Changelog
Every shipped improvement, newest first.
Seedance now tells you before submitting when a prompt is too long
When you write long prompts for Seedance, you're now told the moment you submit if you've gone over the limit.
What's different
- You're told at submit time. If the prompt is over Seedance's limit, hitting Generate shows a message right away — nothing is charged, and you don't wait in the queue for a failed generation.
- Roughly how long is the limit. About 20,000 Chinese characters (about 60,000 English characters) per request. Reference images, audio and video take up a small part of that, so the more media you attach, the slightly less room there is for text.
- When submitting from the CLI, MCP or the agent, the error tells you roughly how many characters to remove.
This applies to all Seedance models.
Privacy lock: set your own shortcut for “Lock Now”
Set your own shortcut for “Lock Now”
Go to Settings → Privacy lock, click the key shown on the Shortcut row, then press the combination you want (Esc to cancel):
- Mac: hold ⌘ (or ⌃) plus ⇧ or ⌥, then press a letter or number;
- Windows: hold Ctrl plus Shift or Alt, then press a letter or number;
- If a combination is already used by the system or XinYu, you'll be asked to pick another. To go back to ⌘⇧L / Ctrl+Shift+L, click Reset.
The setting applies to this device only. Once the desktop app updates to 0.3.5, “Lock Now” in its menu uses your shortcut too (apps on 0.3.1 or later update automatically).
Desktop app privacy lock: lock when the app opens
Lock when the app opens
In the desktop app, go to Settings → Privacy lock. Under "Desktop app only" there's a new option, Lock when the app opens (off by default; set an unlock PIN first):
- Close the app and open it again, and you'll need to unlock before anything shows;
- On Windows, after you close the window to the tray or minimize it, restoring it from the tray or taskbar also asks you to unlock;
- Refreshing the page won't lock it.
Older desktop apps will ask you to update
If you're on an older version of the desktop app, you'll see a prompt in the bottom-right corner to download the new one. Install it once from the download page, and the app keeps itself up to date from then on.
Voice input: tap stop and it stops
Stopping voice input is now much more decisive.
What's different
- Tap Stop and it always stops within 1.5 seconds. When you finish speaking and tap the red Stop button, voice input first tries to write your last sentence into the box. Even on a weak connection, it stops and releases the microphone within 1.5 seconds at most.
- In a hurry? Tap again. Tapping Stop once more during that short wait ends it immediately.
- Works the same everywhere there's a microphone: the agent chat box, the prompt boxes on canvas nodes, and every other input with a voice button.
Saving big canvases is more reliable on a shaky connection
Saving is now more reliable on canvases with lots of nodes, even when your connection isn't great.
What's different
- When your save collides with a change made elsewhere, only your edits are sent. If the same canvas was edited in another tab, or a generation result was just written back to it, your save can bump into that change. Now it goes through by uploading only the nodes you changed — no need to download the whole canvas first. The bigger the canvas and the slower the connection, the bigger the difference.
- New things added elsewhere show up for you. When that happens, nodes and connections added on the other side are brought into the canvas you have open. Things you deleted yourself are not brought back.
- The last save before you leave is smaller. Before you head to top-up or close the tab, the final save only sends what changed, so it is far more likely to arrive before the page goes away.
Tip
The save status is shown under the canvas name — "Saved to cloud" means everything is stored.
This applies to personal canvases; team canvases already sync in real time.
The XinYu agent now lets you keep talking while it works — and rewind when you change your mind
Keep talking while it works
- While the agent is working, keep typing: your message waits under "Queued" above the input box and goes out automatically when the agent is done.
- To change course right away, tap Steer next to that message: the agent finishes the step it's on, then follows your new instruction.
- If a task stops before it's done, waiting messages stay put and are marked "Tap Send to send".
Rewind when you change your mind
Hover over one of your messages and tap Rewind to here:
- The agent's canvas changes after that point are undone — nodes, links and groups it created, and names and settings it changed, go back to how they were;
- Images and videos that are already generated or still generating are kept, and anything you edited yourself afterwards is left alone;
- The conversation after that point is removed, and your original message goes back into the input box so you can edit and resend it.
Steadier
- Long replies are no longer cut off midway just because the agent is writing for a while.
- Before deleting anything, one tap on "Confirm" is all it takes.
- Something you stopped stays stopped — the agent won't pick it back up when you send your next message.
Easier start on a new canvas: example nodes in one click
- Starter buttons: a new, empty canvas now shows a "Double-click the canvas to add a node" hint and a row of starter buttons. Click Text to video or Generate an image to drop in a node with an example prompt already filled in. Your sign-up Xin Points cover one run.
- First node at the right size: on desktop, the first node you add to a new canvas appears at a comfortable size, with the generation panel and button fully on screen, so you can generate as soon as your prompt is ready.
- You can also click Upload to add your own media, or More nodes to pick any other node.
Xin TV Publishing Agreement: your work stays yours, and 18+ works are never promoted
- Your work stays yours: we only display and promote your work on a non-exclusive basis — on Xin TV, the XinYu AI homepage and our official social accounts, credited to your nickname. We never sell your work or license it to any third party for commercial use.
- 18+ works are never promoted: works marked 18+ never appear in any promotion.
- Clear cloning rules: once you turn on "Share Canvas", others can clone your prompts, workflow and reference materials (plus finished images and videos if you turn on "Allow cloning of finished assets") to make their own new works, which they may use commercially. They may not resell or repost your original work as-is, or pass it off as their own.
- Tick to agree every time you publish or edit: the publish dialog now has a confirmation box at the bottom — tick it to publish. Editing a published work works the same way. For works marked 18+, you also confirm that everyone depicted is an adult.
- Unlist any time: once you unlist a work in "My Shares", it is no longer shown to anyone or used in any promotion.
- Read the full Xin TV Publishing Agreement.
The XinYu AI desktop app is here for Mac and Windows
- Download: go to xinyuai.app/download, or click "Desktop" in the site's top menu. The Mac app runs on both Apple silicon and Intel (macOS 13 or later). The Windows app supports Windows 10 or later (64-bit), and you can choose which drive to install it on.
- Local files panel: click "Local Files" in the canvas sidebar to keep a project folder from your computer next to the canvas. Browse it as a list, thumbnails or masonry and drag files straight onto the canvas. File names and paths stay on your computer; only the file you drag in is uploaded.
- Save downloads where you choose: downloading opens a save window that starts in the folder you last used for this project. You get a notification when it's saved, and one click shows the file.
- Notified when it's done: working in another app? You'll get a system notification when an image or video finishes or fails.
- A better privacy lock: lock right from the menu bar, lock automatically when your computer sleeps or locks, and on a Mac unlock with Touch ID and turn on screenshot blocking (XinYu shows as black in screenshots and screen sharing).
- Automatic updates: new versions download in the background and install the next time you quit. No need to come back to the website.
- On the web, clicking "Local Files" on the canvas also takes you to the desktop app download.
- For the full tour and a one-shot film, see The XinYu AI desktop app: your folders, right next to the canvas.
New privacy lock: cover your screen in one click, unlock with a 4-digit PIN
- Where to set it up: Settings → Privacy → Privacy lock. Set a 4-digit PIN and you're ready.
- Lock in one click: choose "Lock now" in the avatar menu or the canvas menu (top left), or press ⌘⇧L (Windows: Ctrl+Shift+L). The whole page is covered right away, and the browser tab title changes to "Locked".
- Unlock: tap the on-screen digits or type your PIN on the keyboard. After 5 wrong tries in a row, wait 30 seconds.
- Refreshing won't unlock it: every XinYu tab open in the same browser locks and unlocks together.
- Auto-lock when idle: optionally lock after 1, 5 or 15 minutes without activity.
- Locking only covers the screen. Generations already running keep going.
- Your PIN stays in this browser on this device and is never uploaded. You'll need to set it again after signing out.
XinYu AI is now available in Traditional Chinese and Japanese
- The site and the canvas are now available in Traditional Chinese and Japanese — switch from the language menu in the top-right corner.
- Browsers set to Traditional Chinese (Taiwan, Hong Kong, Macau) or Japanese open in that language automatically on the first visit.
- Playlist controls on the canvas (split / export / zoom), the “Save to Material Library” dialog and the @ reference menu now follow your interface language.
Canvas motion refresh: generating, failure notices and adding nodes
- While generating: a new sparkle-and-sweep animation with elapsed time in the corner. When you regenerate, a blurred outline of the previous result stays visible, and the overlay fades out softly once the new one lands.
- Failure notices: the full reason is shown — longer explanations scroll inside the box. Messages have been rewritten to say what happened and what to do next.
- Adding nodes: when you add a node yourself (add menu, paste, dropping files and more), the dot grid lights up around it and ripples outward. It stays off when the canvas is zoomed far out or your system has Reduce Motion turned on.
- Zoomed far out: generating nodes get a green outline and failed ones a red outline, so they are easy to spot.
- Menus and dialogs: open and close with one consistent, lighter animation.
- ▶ Watch the 32-second demo (Chinese captions)
Agent updates for longer tasks and retries
-
New Agent tasks start with a spending budget of 500 Xin Points and 120 minutes of active execution time. You will be asked before extending either limit, and progress is preserved when continuing.
-
Retries retain the images, audio and quoted text attached to your request, including when you switch models.
-
Failed Agent tasks receive a full refund of their Agent charge. Separately confirmed media generation remains billed per generation task.
-
Fixed cases where clicking Send stayed on Processing without submitting a task, and improved status updates after continuing a task.
-
More accurate completion status for storyboard batches. Continue creating in the same conversation by replying “continue” or requesting changes.
-
Seedance 2.5 Draft mode: pick the shot at 480P, then render it in 1080P
How to use it
- Pick Seedance 2.5 (Draft) in the model menu. The resolution locks to "Draft 480P".
- Write your prompt and attach references as usual, then generate — you get a 480P draft.
- Not quite right? Keep editing on the same node (the switch under it stays on "Edit draft"); every try costs the 480P price.
- Happy with it? Hit Render final under the node and confirm 1080P. The final appears as a new node next to it, marked "Final" and linked back to its draft by a line. The draft stays as it is.
What the final looks like
The final reuses everything from the draft — prompt, duration, aspect ratio, sound and references — so the framing, motion and timing match the draft, now at 1080P. Fine details such as faces and eyes are re-rendered at 1080P and can differ slightly from the draft; in our tests, with a character reference image attached, the face matched the draft.
Pricing
A draft costs the same as a regular 480P render; a final costs the same as rendering that clip in 1080P directly. What you save is the cost of iterating: the trial and error happens at 480P, and you only pay for 1080P on the one you keep.
For a 5-second text-to-video clip with sound: a draft is 65 Xins, the final 356 Xins. Three drafts plus one final come to 551 Xins; trying three versions directly in 1080P costs 1068 Xins.
Good to know
- Every Seedance 2.5 mode supports drafts: text-to-video, first frame, first & last frame, all-in-one reference, video editing and video extension.
- A draft can be finalized for 7 days. After that, "Render final" is greyed out — use "Edit draft" to make a fresh one.
- You can render the same draft's final again, but it will look almost the same and is charged again; for a different shot, go back and edit the draft.
- The 1080P final is 10-bit HEVC: better image quality, but some players can't open it (VLC and QuickTime can).
- The Canvas Agent can make drafts for you too; rendering the final is always your own click.
CLI and MCP
- CLI:
xinyu gen video --model seedance-2-5 --draft ...for the draft, thenxinyu gen final <draft jobId> --project <canvas id>for the final. - MCP: call
xinyu_generate_videowithdraft: true, thenxinyu_generate_video_finalfor the final.
Generations with reference media now ride out network blips
When you generate with reference images, videos or audio, we first fetch those files from storage, then hand them to the model.
What's new
If that fetch runs into a brief network hiccup — a timeout, a dropped connection, storage being momentarily busy — it now retries automatically, up to 3 times. It all happens in the background; you don't need to press generate again.
This applies whether the generation comes from the canvas, the Agent, the CLI or MCP, and to all three kinds of reference media.
If it still can't get the file
After three failed attempts the job fails and your Xins are refunded in full.
Only network problems are retried. If the file itself is the problem — say it has been deleted, or it's too large — we won't keep retrying; you'll get told right away.
SeeDream 5.0 Pro: limited-time 20% off
SeeDream 5.0 Pro is 20% off for a limited time — a 2K image drops from 14 to 11 Xin Points, and 1K from 10 to 8. The reference-image surcharge (+1 per image from the 6th) is discounted too. Nothing else about the model changed: same resolutions, same reference limits.
The panel shows the original price struck through next to the discounted one; if you already had a canvas open, refresh to see the new price.
GPT-6 Sol / Luna, Claude Fable 5.1 and DeepSeek V4.1 Flash are now in XinYu chat
Four new brains
Open the model picker in XinYu chat and you'll find four new options:
- GPT-6 Luna — the lowest-priced GPT. Plenty for everyday questions and prompt tweaks.
- GPT-6 Sol — GPT-6's balanced tier, at about a fifth of GPT-6 Astra's price. It keeps reasoning while it uses canvas tools, which makes it a good default for most creative work.
- Claude Fable 5.1 — the newer Claude Fable, with flagship-level reasoning and tool use. On hard problems it decides for itself when to think deeper.
- DeepSeek V4.1 Flash — DeepSeek's newer lite brain; it thinks before it acts, and lets you watch it think.
Which one to pick
- Cheapest: GPT-6 Luna
- Want to see how it reasons: DeepSeek V4.1 Flash
- Most work: GPT-6 Sol
- Hard problems worth paying a little more for: Claude Fable 5.1 or GPT-6 Astra
Every model's price is shown in the picker — glance at it before you switch. You can switch back any time.
Video editing timecodes now cover a span, not a moment
In Seedance 2.5 video editing, "Insert timecode" now gives you a start and an end slider, and what lands in your prompt is a span like
00:05-00:08rather than a single moment.Why a span works better
Video editing is a full re-render: the model regenerates the clip, changing what you asked for and holding the rest. It needs to know where the boundaries are. Give it a moment and it has to guess; give it a start and an end and the change lands inside the window you drew.
The official prompting examples are all continuous spans too, so the panel now just gives you spans directly.
The panel tells you which stretches you left out
Say you wrote
00:05-00:08to swap a character, but the clip runs 18 seconds — the other 13 seconds are unspecified. The panel lists them:00:00-00:05, 00:08-00:18 not described — the model will improvise
Next to it there's Fill gaps: one click turns them into "keep the original footage unchanged".
You don't have to click it. Letting the model improvise in the blank stretches is a perfectly normal way to work, so we don't decide that for you. It stays a line of text and a button, and nothing is added to your prompt when you send.
Two small notes
- The sliders move in whole seconds. The platform wants whole seconds, and previously you could stop on a half second, so what you saw and what you sent could differ.
- The video editing dialog from the node toolbar still inserts single moments — that path is for stepping through frames and marking a spot, where "this frame" is exactly what you mean. Both kinds of chip can live in the same prompt.
XinTV now shows how many people watched your work
Every work on XinTV now has a view count that goes up with real views — visible on list cards, on the work page, and in your own "My shares".
How the number is counted
- Your own visits don't count. Opening your own work repeatedly won't move it.
- One person counts once per 24 hours. Refreshing, switching tabs, clicking in and out of the list — all count once.
- Only published public works are counted.
Why spell out the rules
Because the number is only worth anything if you can act on it. If your own visits counted, and every refresh counted, it would quickly become decoration nobody looks at. With the rules above, it reflects how many people actually opened your work — solid enough to decide what to make next.
Counting never gets in the way of the work
The view count is an observability number. It never affects whether anyone can watch your work, and never affects billing.
XinTV now has an 18+ switch — you decide what you see
There's a new 18+ switch at the end of the filter row in XinTV (All / Animation / Short film …).
- Off by default, so the list shows all-ages works only.
- To include 18+ works, tap 18+ and confirm you are 18 or older.
- That confirmation lasts for the current browsing session only — close the tab and you'll be asked again.
- Tap it again to hide them.
Nothing changes for creators
Whether a work is marked 18+ is still whatever you chose when you published it, and no work was unlisted. Works marked 18+ are still on XinTV; viewers just turn the switch on to see them. Your own works are always visible to you under My Shares, regardless of this switch.
The gate on the work itself is unchanged
Turning the switch on only affects what appears in the list. Opening a work still goes through the same age confirmation and unlock flow as before.
Switching teams takes you straight to the project list
Switching teams
Pick any team in the workspace sidebar, or leave a team canvas via Back to workspace — you now land straight in that team's project list.
The address bar stays scoped to that team, so refreshing the page or bookmarking the link brings you back to the same team.
Where team settings live
To rename a team, invite members, or change member permissions:
- Desktop: avatar menu (top right) → My Teams → pick the team
- Mobile: drawer menu → the team name with the gear icon
Videos added with “+” now really do edit and extend
Both ways work the same now
There are two ways to attach a reference video for Seedance 2.5: draw a connecting line, or use “+” to pick one from your canvas assets.
Until now only the first one actually ran as the task you picked. Videos added with “+” ran as general reference instead — the label was right, but what came back wasn't the kind of result you asked for. Both ways behave identically now.
The panel stops showing values that don't apply
On Video edit or Extend, the output aspect (and for edit, the length) follows the source video. So the panel now says exactly that, instead of showing a number you can change but that wouldn't take effect.
How to tell
Attach the video, pick Seedance 2.5, then choose Video edit or Extend in the prompt panel — the footer should read “follows the source”. When you see it, this run is using the task you picked.
Video editing, now right in the prompt panel
Edit right where you write
Connect a video into a generation node, pick Seedance 2.5, and a Video edit tab appears in the prompt panel's tab row.
Once you're on it:
- Just say what you want changed — no need to phrase it with words like "edit" or "replace", we handle that
- The price sits right next to it, visible before you hit generate
- Output length and aspect follow the source video, so those two pickers step aside (they wouldn't apply anyway)
Point at a moment with a timecode
Hit Insert timecode, drag the slider to the moment you mean (the range is this video's own length), and hit Insert — a timecode marker lands in your prompt. Write what should change right after it.
The full editor is still there
When you need to step through frames and circle a spot on the picture, the editor on the video node's toolbar works exactly as before. Two routes side by side: type a sentence in the panel, or open the editor when you want to draw.
Wan 3.0 Prime is live, and 21:9 is open
Wan 3.0 Prime — for when you're waiting on the result
Aliyun describes Prime as the high-speed edition: capabilities aligned with the standard edition, with end-to-end speed significantly improved. So Prime sells speed, not quality — parameters, ratios, durations, resolutions and reference-media limits are identical to the standard edition. The only difference is how fast the clip comes back.
Pick Prime when you're in a hurry, stay on the standard edition when cost matters more.
Tier Standard (50% off) Prime (40% off) 480P 3 Xins/s 4.9 Xins/s 720P 6 Xins/s 10.08 Xins/s 1080P 12 Xins/s 20.16 Xins/s A 30-second 720P clip: 180 Xins on the standard edition, 302 Xins on Prime.
Pick "Wan 3.0 Prime" straight from the model dropdown; the CLI and MCP server have it too (
--model wan-3.0-video-prime).21:9 widescreen is open — on both Wan 3.0 models
The upstream opened 21:9 up, so we did too. We measured the actual output dimensions at all three resolutions rather than just checking that the job succeeded:
Resolution Actual output 480P 980×420 720P 1470×630 1080P 2206×946 21:9 now appears in the ratio picker for both the standard edition and Prime.
Also
Wan 3.0 accepts any whole number of seconds from 2 to 30. When the Canvas Agent planned one of these for you, it used to budget as if the clip were 20 seconds; it now matches what you are actually charged.
GPT Image 2.5 limited-time price cut — up to 22% off
High-quality tiers are cheaper across the board
The discount our upstream gives us, passed straight through to you. Already live — nothing to set.
Size High Ultra Max 1K 7 → 6 12 → 10 27 → 22 2K 14 → 11 24 → 20 54 → 43 4K 23 → 18 40 → 32 89 → 72 4K Medium also drops from 6 to 5.
A 4K Max shot: 89 Xins → 72
Cuts run 14–22%; weighted by real usage that works out to roughly 20% off.
Low tiers, and Medium at 1K and 2K, are unchanged — those already cost just 1–4 Xins.
The panel shows the exact cost for the shot before you run it.
Why "limited-time"
This cut tracks the discount our upstream gives us. If that discount ends, our prices follow — and we will say so here first, not quietly.
You can take a generation back now
Those seconds are yours
After you hit generate on the canvas, the node shows a countdown with two buttons: Cancel and Submit now.
- Clicked the wrong thing → hit Cancel. The request never left your browser, so nothing is charged.
- Just a typo in the prompt → don't cancel, fix it right there in those seconds. What runs is your corrected version.
- Don't want to wait → hit Submit now, or click generate again to start immediately.
You decide how long
Open Canvas settings → Submit delay and set anything from 0 to 10 seconds. The default is 3 seconds — long enough to catch a misclick, short enough to stay out of your way.
Set it to 0 to turn it off entirely and go straight to generating, exactly like before.
Where it applies
Every generation on the canvas that spends credits: image, video, audio, text, inpaint, outpaint, background removal, subtitle removal, relight, trim, export and storyboard parsing.
Images on your phone: pinch to zoom, swipe to browse
Photo-album gestures for images on your phone
Open an image in Assets:
- Pinch to zoom in and out, then drag with one finger to look around
- Double-tap for 2.5x — it zooms into the spot you tapped; double-tap again to go back
- Swipe left or right to move to the next or previous image
- Swipe up or down on the picture to scroll straight to the file details and prompt below
Portrait images and videos are no longer squeezed into a thin strip either — you get a proper view as soon as you open one.
For video: the fullscreen, play, mute and restart buttons are bigger on phones, so you don't have to aim. (Video and audio still use the player's own controls; the gestures above are for images.)
On a mouse, nothing about the desktop changes — scroll-wheel zoom, the zoom slider and the arrow keys are all exactly where they were.
One look for the mobile menu
Whichever page you're on, the menu now shares the same design: icon rows, section headings and the language switch at the bottom. Once you're signed in you get the account and settings entries as well, in the same style. On the English interface the section headings are in English too.
The prompt box now shows each model's character limit
Know the limit before you write
Models differ in how long a prompt they accept, and until now you only found out after sending. The prompt box now shows a live count of characters written / that model's limit — amber as you approach it, red when it is full.
Pasting a long prompt no longer goes nowhere
Paste a long prompt and, if it exceeds the limit, we keep what fits and tell you exactly how many characters did not make it. No counting by hand, and no "I pasted and nothing happened".
The limits
Model Limit Seedance (all) No limit Wan 3.0 20,000 MiniMax H3 7,000 Wan 2.7 5,000 Kling 3.0 2,500 HappyHorse 1.1 2,500 Kling O3 No limit These follow each model's own upstream specification. The same rules apply when you submit via the CLI or MCP.
Free local subtitle cleanup: select, preview and export
- Select a generated video on the canvas and open Local subtitle cleanup · Free. Mark the subtitle area and time range, preview the result, then export.
- Supports videos up to 30 seconds and 1080p at 24/30 fps. Available for verifiable XinYu-generated assets; uploaded videos are not supported.
- Designed for white subtitles in a fixed position. Add or erase strokes to refine the repair area, and check the preview on complex backgrounds.
- Export a new video for 0 Xin, preserving the original and its audio track. The question-mark button opens an illustrated guide; Esc closes the guide first, then the editor.
GPT Image 2.5 can output transparent backgrounds; reference images are now priced per image
Transparent backgrounds
GPT Image 2.5 (Flare and Sunburst) now has a Background control in the parameter panel, with three options:
Auto · Opaque · Transparent
Pick Transparent and you get a PNG with a real alpha channel — everything outside the subject is genuinely transparent, not white. That saves a cutout step for product shots, stickers and asset pieces.
The control only appears on GPT Image 2.5, because transparent output is specific to it.
Aspect ratios now follow the size tier
1K / 2K / 4K offer almost the same set of ratios: 1:1, 4:3, 3:4, 3:2, 2:3, 16:9, 9:16 and 21:9 are available at every tier, and 2K additionally offers 9:21. The ratio you choose is the ratio you get.
The panel now lists only the ratios a tier can actually produce. So 1:1 disappears when you switch to 4K — that isn't an omission, it's that the tier can't produce it.
Reference images are now priced per image
Starting today, attaching reference images to GPT Image 2 and GPT Image 2.5 counts toward the price:
Model Pricing GPT Image 2 First one free, +1 Xin for each after that GPT Image 2.5 Flare / Sunburst +1 Xin per reference image For example, a 1K low-refinement GPT Image 2.5 render is 1 Xin; attach one reference and it's 2. The price in the panel tracks the reference count live, so you see the final amount before you run it.
Why: the upstream for these two models bills for reference images per image, and larger references cost more — a 4K reference costs over three times a 1K one. That part wasn't reflected in the price before.
Not affected: reference images on every other image model remain free of extra charge; nothing about their usage or pricing changes.
GPT Image 2.5 is here: Flare and Sunburst, five refinement levels
Two new models are in the image list: GPT Image 2.5 Flare and GPT Image 2.5 Sunburst.
How they differ
- Flare — faster generation. The everyday pick.
- Sunburst — same image quality, about twice as slow, tuned for edit precision. Reach for it when you are retouching or chasing fine detail.
They cost exactly the same; price follows the refinement level and size you pick.
Five refinement levels instead of three
GPT Image 2 had Low / Medium / High. 2.5 adds two more on top:
Low → Medium → High → Ultra → Max
Higher levels spend more compute on the image and cost more. At 1K that ranges from 1 to 27 Xins, and the panel shows the exact price for the shot before you run it.
Note: "High" on 2.5 and "High" on GPT Image 2 are not the same thing — each generation splits its levels by its own compute. For the best 2.5 output, pick Max.
16 reference images at once
Double GPT Image 2's eight. More room for character consistency, multi-image blends, and style references.
Both models also edit with references — just connect images to the generation node on the canvas.
Sizes and aspect ratios
1K / 2K / 4K. Each tier supports a different set of ratios (2K is the widest — it adds 21:9 and 9:21), and the panel only lists the ratios that tier can actually produce, so you can't pick one that won't come out.
Prompts up to 32,000 characters
GPT Image 2.5 accepts prompts up to 32,000 characters. Go over and you'll be told, instead of getting an image made from half your prompt.
GPT-6 Astra is now available in XinYu chat
A new top tier
GPT-6 Astra — OpenAI's current flagship — is now in the model list in XinYu chat.
Where it shines is work that needs step-by-step reasoning: turning a dense brief into a shootable shot list, finding a workable option inside a pile of constraints, restructuring long-form content. It holds up noticeably better than the other tiers on that kind of task.
Its deep thinking stays on while it uses canvas tools — letting it read your canvas and reason at the same time is where it's at its best.
Two things to know first
It's slow. Measured at roughly 3x the time of GPT-5.6. For everyday questions or tweaking a prompt, a faster tier is a much better experience.
It's the priciest tier. The exact rate is shown in the model picker — glance at it before you switch.
In short: give it the hard problems that are worth waiting for, and use something else for the rest.
How to switch
Open XinYu chat, click the model picker, and choose GPT-6 Astra. You can switch back any time.
You can now speak your prompts
Tap, and just say it
Image, video, audio and text nodes, storyboard scripts, prompt nodes, and the XinYu chat on the right — all of them now have a microphone next to the send button.
Tap once to start, tap again to stop. What you say lands straight in the box, appended to whatever you had already written.
Nothing is sent automatically. Look it over, fix a word, add your @ references, and send when you're ready. For long prompts, say a chunk, pause, then keep going.
Chinese and English
To the left of the microphone there's a 中 / EN toggle. Pick one and it recognizes in that language; your choice is remembered for next time.
Speaking a Chinese prompt with English model names or craft terms mixed in? Stay on 中 and say it naturally — common creative terminology is boosted.
No credits
Recognition happens in your browser. Your audio never reaches our servers, and it costs no credits.
Chrome or Safari work best. The first time you tap the microphone your browser will ask for permission — allow it. Safari needs Dictation or Siri enabled in system settings. If your browser can't do it, the button simply won't appear, so you'll never tap something that does nothing.
One-tap Seedance 2.0 Mini, and your welcome bonus is now claimable
Your first Mini, covered by your welcome bonus
Your welcome bonus covers a free account's first 4-second Seedance 2.0 Mini. Whatever the panel shows is what you're actually charged.
Free accounts get one Mini clip, up to 4 seconds. Upgrading removes the limit entirely and raises the cap to 15 seconds.
Your welcome bonus: the email now arrives on its own
Signing up now sends you a verification email automatically. Click the link and your welcome bonus lands.
If you registered earlier and never verified, there's a notice at the top of your workspace with a one-tap resend. The bonus has been waiting for you.
One click to start
The Mini write-up and the Seedance page now each have a direct entry point. Click it and we open a canvas for you with a Mini node already set up — model, duration, resolution and a sample prompt all filled in. Change what you want and hit generate.
Faster cropping, and upscales now keep their full resolution
Cropping is faster
Cropping used to download the entire original image into your browser before it could do anything. The bigger your image, the longer the wait — 4K renders and upscaled images routinely run to tens or hundreds of megabytes, which is painful on a phone or a long-haul connection.
The cut now happens on our servers; your browser only sends the position of the crop box. The image already lives next to the server, so it reads once, cuts once, and writes once.
While it runs you get a "generating" node on the canvas that fills itself in when it's done — the same experience as generating an image, so you can carry on with something else while you wait.
Upscales keep their full resolution
Upscaling, background removal and annotation now all work from the original image.
Until now they read the compressed copy the canvas uses for fast previews. Upscaling was hit hardest: a 5504×3072 image doubled should give you 11008×6144, but came back at 3200×1786 — smaller than your own original. These now start from the original, so an upscale gives you the full resolution.
Grid split and combine
Images produced by grid split and combine are now registered in your asset library too, so you can reference and manage them like any other material.
Crop tool: 7 more ratios, full-resolution output, and it works on touch
19 ratios, up from 6
The crop ratio list used to carry only six presets, far fewer than what generation itself can produce. The two are now aligned:
21:9, 3:2, 2:3, 5:4, 4:5, 2:1 and 1:2 are new, plus 9:21 for tall ultra-wide crops. With Free and Original, that is 19 in total.
Original doubles as a real reset: after cycling through a few ratios, tapping it brings the box back to the whole image instead of shrinking a little more each time.
High-resolution images keep their resolution
Cropping now reads the original file. It used to read the compressed copy the canvas uses for fast previews, so 2K and 4K renders — and anything you had upscaled — came out around 1600 pixels wide, with nothing on screen to tell you.
Those images now keep their full resolution. In our testing, a 7078-pixel-wide image that previously came back at 1600 now comes back at 7078.
The size readout in the corner switched to original-image pixels too, so the number you see is the number you get.
It works on phones and tablets
The crop box used to respond to a mouse only. On touch you could open the tool, pick a ratio and hit confirm — but the box itself would not move. Finger and Apple Pencil dragging now work, and the grab areas around the corners and edges are larger, so you no longer have to hit the small white dot exactly.
Esc closes the crop tool.
Type an exact size
Above the ratio grid there is now a size field. Enter the size in original-image pixels; it applies on Enter or when you click away. The Lock ratio switch next to it pins the current aspect ratio — including an arbitrary one you dragged out yourself — so dragging scales without distorting.
Cropping also shows progress now, instead of leaving you guessing how long it will take.
Reference links on your canvas stay put — and the ones you lost are back
Reference links stay put
A reference lives on the link between two nodes: the link is what lets @Image 1 in your prompt resolve to the right asset. Previously, when the same canvas was being written from more than one place at once — two tabs open, or the agent placing something on the canvas while you worked — links could go missing. What you saw was a whole batch of reference tags greying out, without you having deleted anything.
Those cases are now reconciled item by item before saving: links created elsewhere are kept, and links you cut yourself are never reconnected behind your back.
The links you lost are back
We ran a recovery using the most conservative test we could: the link was genuinely created, you never deleted it, both nodes still exist, and your prompt still carries the reference tag pointing at it. Only links meeting every condition were restored. 111 links came back — just open the canvas, nothing for you to do.
Anything that failed those conditions was left alone, so nothing you meant to remove gets pushed back at you.
No more English error page after an update
When a new version went live while your page was still open, opening certain panels could turn the whole page into a single line of English error text. Now the page saves your unsaved work first, then refreshes itself onto the new version — at most you see a brief load.
Reference tags in your prompt no longer disappear on their own
Tags grey out; they are never deleted
Reference tags such as @Image 1 in your prompt used to be stripped out of your text whenever a reference could not be read for a moment. Once removed, a refresh would not bring them back, and you had not deleted anything.
Your text is now left alone. When a reference cannot be read the tag simply greys out and explains itself on hover. Whether to clear it is your decision, via the “Remove stale tags” button.
Swapped references say so
When you replace the reference in a slot, the tag used to turn into an unexplained grey question mark, even though generation was in fact using the new image.
That case is now labelled explicitly: hovering says the slot holds a different reference and that generation will use the current one. The tag stays usable; you just know it is not the original.
Tags follow the asset itself
Reference tags now bind to the asset's own identity rather than only to the image address it had at the time. Re-generating, changing the cover, or moving storage no longer breaks them, so you don't have to go back and re-@ everything.
The “+” picker now lists only what's actually on this canvas — with real names
“This canvas” means this canvas
When you open the “+” to pick a reference, the “This canvas” tab used to include things you couldn't actually see on the canvas: copies left behind when you picked from another canvas, outputs replaced by a re-generation, leftovers from nodes you had deleted.
It now lists only what is really on this canvas: if the node is there, so is the asset; delete the node and it disappears immediately. “Other canvases” follows the same rule.
“Generation history” is unchanged — that tab is the generation record for this canvas, including output from nodes you later deleted. That's where to look for older images.
Cards have names now
Asset cards used to be titled “Untitled” almost every time (generated images have no filename). They now show the name of the node they belong to — the title you gave that node, and renaming it updates the list. Especially handy when picking from another canvas: you can tell at a glance which node an image came from.
Rejected references explain themselves
When a reference can't be used, the message now says which kind of problem it is and what to change — for example that the model does not take reference audio, that start/end-frame mode doesn't use reference video or audio, or that a referenced asset no longer exists.
Image nodes start with a “+”, and Seed Audio picks reference voices without edges
Image nodes start with a “+”
The “+” used to appear only once the tray already had a reference — a brand-new image node with nothing attached had no “+”, so the first reference could only come from an edge. New nodes now show the “+” right away; pick from your library and it registers on that node.
Seed Audio: reference voices without edges
With Seed Audio 1.0 selected, the audio node's tray now has a “+” — pick an audio clip from your library (this canvas / generation history / other canvases) as the reference voice, no edge needed.
- Up to 3 clips, wired and picked ones counted together; the “+” hides itself when full
- ElevenLabs shows no “+” (it doesn't take reference audio)
Picking an uploaded audio with “+” now generates
Picking an uploaded audio file with “+” (on a video or audio node) used to fail at generation. It now generates normally — nothing to configure.
The “+” picker now takes video and audio references
The “+” now picks video and audio
Yesterday's “+” only took images. The picker now has Video and Audio tabs; whatever you pick is registered on the current node just like images — no edge to draw, and later changes to the source canvas don't touch this node.
Tabs follow the model: you only see the tabs for the reference types the current model actually takes, so you can't pick something that would never be used.
Model Reference video Reference audio Seedance 2.0 / Fast / Mini up to 3 clips up to 3 clips Seedance 2.5 up to 10 clips up to 10 clips Wan 3.0 up to 5 clips up to 5 clips Wan 2.7 up to 3 clips — MiniMax H3 up to 3 clips up to 3 clips Wired references and “+” picks share one limit; the panel price and the final charge are both computed on the merged count.
Start/end frames switch only from a clean state
Start/end-frame mode uses two images and no reference video or audio. So with video/audio references attached, or more than two images, the “Start/End” tab is still there — but tapping it tells you exactly what to remove first, instead of switching and dropping the extras. Once both frame slots are filled, the “+” hides itself.
The panel says it up front
- Pick a model that doesn't take reference audio and the audio slot is greyed out with a note that it won't be sent — you never generate with an audio clip nobody uses.
- In video edit, choosing a tier that has no matching edit model (the 4K tier, for example) shows plainly that Kling O3 Standard will generate and bill at the Standard rate.
Canvas references: attach without wiring, and drag to reorder
Attaching a reference image to a canvas node no longer requires wiring an edge.
How — there's a new "+" at the end of the reference tray. Pick from your asset library; both your own canvases and team canvases are available. Whatever you pick is registered onto the current node, so later changes to the source canvas won't affect it.
One row, mixed freely
- Wired references and "+" references now live in the same row and can be interleaved in any order
- Drag to reorder — the
@Image1/@Image2numbers in your prompt follow along - The badge on each thumbnail is its number in the prompt; click a thumbnail to insert it
Works for first/last frame too
- A "+" reference can be the first frame or the last frame — whichever comes first is the first frame
- Badges now read "First Frame / Last Frame", and the one-click swap is still there
- For models that only accept a first frame, the second slot is greyed out with an explanation — your image is never silently dropped
The limit follows the model, and the dialog shows how many you can still add. Available in both image and video generation.
New: Zhipu GLM-5.3 and GLM-5.3 Flash, with a 1M-token context
Two Zhipu models have been added, selectable in both the canvas agent and the text node.
GLM-5.3 Flash — available on the free tier
- 11 credits per 1M input tokens, 35 credits per 1M output tokens
GLM-5.3 — for members
- 97 credits per 1M input tokens, 303 credits per 1M output tokens
What both share
- A 1M-token context, and up to 131k tokens of output in one go
- Automatic caching: when you reuse the same context, the cached part is billed at roughly a fifth of the normal input price — nothing to switch on
- Tool calling, so in the canvas agent they can create nodes and change parameters directly
When Flash is the cheapest choice
Its input price is the lowest we offer (11, against DeepSeek V4 Flash's 12), but its output costs more than DeepSeek V4 Flash (35 against 23). So Flash wins when you feed in a lot and want a little back — long-document summaries, passing over a large body of material, bulk rewrites. When you need to generate a lot of text, DeepSeek V4 Flash is the better deal.
What you need to do
Nothing — just pick them from the model list.
MiniMax H3 at half price (through Sept 13)
MiniMax H3 is 50% off at every resolution and in every mode until September 13 — a 15-second 2K clip drops from 263 to 132 Xin Points. The upstream price cut passed straight through; nothing about the model was reduced.
No more page zoom when you tap an input on mobile
Tapping an input no longer zooms the page
On iPhone, tapping any input — a prompt field on the canvas, a search box, a filter — used to zoom the page in, and it would never zoom back out. You had to pinch out manually, and even then the page stayed scrolled somewhere odd and could be dragged sideways.
That's fixed iOS Safari behaviour: it zooms whenever an input's text is below a certain size. Inputs on narrow screens now clear that threshold, so tapping one leaves the page exactly where it was.
Nothing to turn on. Desktop is unaffected.
Enter no longer fires destructive confirmations
Delete-style actions ask for confirmation. Until now, once that dialog was open, a stray Enter counted as "confirm". Enter no longer triggers the destructive action — focus starts on Cancel, and running it takes a deliberate click.
Keyboard handling got tidied up while we were there: these dialogs close with Esc, and Tab no longer wanders behind the dialog.
Your membership tier, right in the canvas
The canvas top bar now shows a membership tier badge next to your credits.
- See your plan at a glance — FREE / BASIC / PRO / ULTIMATE, no digging through settings.
- Click the badge for membership plans — compare pricing and monthly credits for all three tiers without leaving the canvas.
- Clicking the credits number still opens top-up, exactly as before.
Note: on team canvases that button switches which wallet pays (canvas pool / personal) — a different thing, so no tier badge there.
Wan 3.0 is here: up to 30 seconds, with text, image, audio and video references
There is a new Wan 3.0 option in the video node. Two things set it apart:
One: up to 30 seconds per clip. Most models stop at 15. Wan 3.0 goes to 30, and any whole number of seconds between 2 and 30 works.
Two: text, image, video and audio all work as references. Up to 10 reference images, 5 reference videos and 5 reference audio clips in a single request. Locking a character, a location, or driving the picture from an audio track all happen in the same model.
Also:
- 480P / 720P / 1080P
- First and last frame: give it an opening and a closing image, it works out the middle
- Audio toggle: leave it on for a finished clip with sound, turn it off to score it yourself
- Fixed 30fps
Wan 3.0 is 50% off through 16 October. 6 Xins per second at 720P — 180 Xins for a 30-second clip.
Update (2026-09-16): our upstream cut its price again and we passed it straight through — previously 30% off (8.4 Xins/s at 720P).
⏳ One heads-up: 1080P and longer clips take a while, and the time varies a lot. The panel will tell you; submit it and go do something else, the node updates on its own.
The model picker has been rebuilt
The image and video model lists used to be one flat grid — a dozen-plus models dumped at once. Now:
- A recommended section up top, so you don't have to read the whole list
- Video is grouped by maker (Seedance / Kling / Wan); images are ordered newest first
- A search box — if you know the name, just type it
- Every model now says when to pick it, instead of listing specs
Switching models now tells you which settings it changed
Switching models now announces the settings it adjusted.
Every model offers a different set of resolution and duration tiers. When you switch, your current settings land on the closest tier the new model supports — and that step now raises an instant notice spelling out exactly what changed:
This model doesn't support 5s — switched to 4s
Resolution works the same way. The notice only appears once you've actually picked a tier — a freshly created node is just taking its defaults, so it stays quiet.
In short: after switching models you no longer have to re-check the panel line by line. If something moved, you'll know right away.
Full-screen voice panel, and prompts that fit your tablet
The voice panel expands to full screen. Click the expand icon in the panel's top-right and the input area grows, so a long script fits on one view — image, video and text panels already worked this way, and voice now matches.
Tablets and narrow windows: the prompt panel fits itself to the screen. It sizes to the available width and always stays fully in view — even when the node sits at the very left or right edge of your canvas, at any zoom level.
Nothing changes on a wide screen; it works exactly as it did.
Canvas videos: mute, download the master, enlarge
Select a video node — three buttons sit in the top-right of the frame.
- Download the master — you get the master file, ready to edit, publish or hand off. The version-history dialog downloads it too, so you can grab an older take without restoring it to the node first.
- Mute — sound on or off in one click.
- Enlarge — it opens right on the canvas, no fullscreen jump, no leaving the layout you're working on.
Double-click to enlarge: images and videos alike — double-click the frame, no need to find a button.
Node toolbars are back to a single always-visible row — select a node and every action is right there.
Batch video generation now runs in parallel
Submit several videos at once and they now all start together — no more waiting for the ones ahead.
Batches
Generate a batch of shots together, or run a few variants of the same prompt: they all start at once. Up to 12 run in parallel.
Everything keeps its own pace
The last step of a video render converts the result into a version that plays smoothly on the canvas. That step no longer holds up images or audio running at the same time — each goes at its own pace.
Unchanged
How long a single video takes still depends on the model itself. What changed is how many can run at once.
Director Desk is live: block the shot before you shoot it
A clapperboard button now sits in the canvas toolbar on the left. Click it and you get a director-desk node — open it and you're standing in a 3D stage.
What you can do in there
Block your actors — add figures, drag them around, turn them. Height, build and body type are adjustable. Props (tables, crates, railings) can be dropped in as stand-ins.
Place the camera — free-fly the main camera, switch focal lengths from 18mm to 135mm, and watch the live viewfinder in the corner: what you see is the shot. Happy with it? Save it as a recorded shot.
Pose them — 93 pose presets, from stand/sit/walk/run to sword, aiming, spellcasting, zombies and farm labour, grouped and labelled. Picked one? You can then tweak it joint by joint — 17 in all (head, neck, chest, spine, pelvis, plus upper arm / forearm / hand and thigh / calf / foot on both sides), each rotatable on its own, with a dot marking the ones you've touched.
Save the poses you like — up to 6, stored on your account (they follow you to another machine). One click brings a pose back, on this actor or any other. Double-click a slot to rename it.
Draw the walk — set a start and an end for any actor, add waypoints in between to route around obstacles. Waypoints are draggable right in the 3D stage. Path shape can be polyline, smooth or step, and the easing is adjustable. Hit play and the actor walks it, facing where they're going.
What you get out of it
Three things drop straight back onto the canvas:
- Clean reference still — the bare frame, to feed an image node as composition reference
- Annotated reference still — with actor labels, facing arrows and the axis line, for you or a collaborator to read
- Motion reference video — the walk rendered frame by frame to MP4, to feed a video model as a motion cue
Each of them becomes a canvas node with one click. Wire it up and keep going.
In one line
Until now "how is this shot framed" lived in your head or on a napkin. Now you can block it in 3D first, then have the model generate against that exact framing and that exact movement.
The agent brings a storyboarding playbook to shot breakdowns
Shot breakdowns
When you ask the canvas agent to break a scene into shots, it now brings a storyboarding playbook:
- Camera values — what focal length, height and angle this shot wants
- Composition hazards — which framings fall apart most often in the render
- Cross-shot continuity — what has to carry over from one shot to the next
Transformation and suit-up shots
For armour plates flying into place, suit-ups and form changes, five new rules:
- Two similar-looking pieces of gear: give each a semantic name, and state explicitly that B's features do not belong on A
- Don't write the whole transformation as one shot — split it into fast cuts, each showing one local area
- A time code on the block isn't enough; inside the shot, spell out what happens in each second
- Naming a movie doesn't give you motion — write how each plate flies, lands and locks
- Don't write a shot shorter than 1 second; it gets stretched and eats the shots after it
What you need to do
Nothing. It's already live.
The canvas agent now writes video prompts as flowing prose
When you ask the canvas agent to write a Seedance prompt, its output looks different now.
- Bracketed blocks:
[GLOBAL][SCENE][CAMERA], filled in slot by slot - Flowing prose: one continuous passage that still carries everything those slots used to
Why prose holds up better
Filling slots slides very easily into adjectives — drop one word per slot and it looks finished. But adjectives don't produce pixels:
- "tense atmosphere" doesn't make the shot tense
- "a magenta light" can come back orange-red — the model slides toward whatever prior sits closest
Prose forces you to finish the sentence. Same lighting example, written as three things: where the light comes from, what it lands on, and what the skin looks like once it does. Write all three and the colour holds.
What you need to do
Nothing. It's already live.
The old format is still there
The bracketed template hasn't been deleted — it's kept as a compatibility format.
- Bracketed blocks:
Wan 2.7 now supports end frames: give it a first and a last image
When generating with Wan 2.7, the start/end frame mode now takes two images:
- The first sets how the video opens
- The second sets how it ends
The model works out the transition in between. When you need a shot to travel from one definite state to another, this is far more precise than giving a first frame and describing the rest in words — the same character going from seated to standing, a camera pulling from close-up to wide, daylight turning to night.
How to use it
In a video node pick Wan 2.7 → switch to start/end frames → drop an image into each slot. Using only the first one still works; that's ordinary image-to-video.
Two notes
- Keep the two images at similar dimensions — a big mismatch tends to warp the transition
- An end frame alone won't work; the first frame is required
HappyHorse 1.1 is here — 9 aspect ratios, up to 15s, audio included
A new model has landed in the video node's model list: HappyHorse 1.1, Alibaba's in-house video model.
What it does
- Text to video — write a prompt, get a clip
- Image to video — hand it a first frame and let it move
- Multi-image reference — attach up to 9 reference images to lock character and object consistency
Things worth knowing
- 9 aspect ratios: 16:9, 9:16, 1:1, 4:3, 3:4, 4:5, 5:4, 9:21, 21:9 — including 21:9 ultra-wide and 9:21 ultra-tall, which most models don't offer
- Any length from 3 to 15 seconds, not just a few fixed steps
- Audio comes built in — every clip ships with synced sound, no toggle, no surcharge
- 720p / 1080p
Pricing
16.8 credits/sec at 720p, 21.6 credits/sec at 1080p. A 5-second 720p clip costs 84 credits.
Two notes
- In image-to-video mode the output ratio follows your input image, so the aspect selector doesn't apply — that's the model's own rule
- This version doesn't do video editing; it takes image references only, not video
Batch download: review the list, then name every file
Select several assets on the canvas and hit Download — instead of packing immediately, you now get a list.
Decide the names before anything is packed
Rename any single file, or apply a rule to the whole batch:
- Index + name (default), name + index, name only, index only, type + index, prompt excerpt
- Prefix, separator (
_-space, none), start number, digits, append timestamp - Hit Apply to all when you want the rule to overwrite the ones you edited by hand too
While you type a name, the actual filename is shown underneath — the suffix added for duplicates, the characters your filesystem won't take — so you see it before downloading, not after unzipping.
Order, exclusions, zip name
- Drag the handle on the left to reorder; numbering follows
- Click × to exclude a file; the footer has Restore if you didn't mean to
- Name the zip yourself
Hover a thumbnail to see it large, so you can check which one you're renaming.
Filenames follow your node titles
Single downloads and batch downloads both give you the title you see on the canvas: your own title if you renamed the node, otherwise the node's own name; uploaded files keep their original filename.
Prefer naming by prompt? It's still there — pick Prompt excerpt as the naming rule.
Seedance 2.5 now outputs 1080p
Seedance 2.5 now has three resolution tiers: 480p · 720p · 1080p.
- All six aspect ratios: 16:9, 9:16, 1:1, 4:3, 3:4, 21:9
- Reference images, reference video and audio all work at 1080p too
- Available from the CLI and from Claude / Cursor via MCP
Pricing
1080p costs 2.46× what 720p does. For example, 5 seconds at 16:9 is about 356 Xins (the same clip at 720p is about 145). The price updates live as you switch tiers, so you see it before you run.
Two things worth knowing first
The format is 10-bit HEVC. Finer detail than 8-bit, but some players and editors can't open it — the built-in Windows player and older editing software may show a black screen or refuse the file outright. VLC and QuickTime play it fine; for editing, transcode to H.264 first.
Generation time varies a lot. The same 1080p job can take a few minutes or twenty-odd, depending mostly on upstream queue depth rather than how many seconds you asked for. Jobs run in the background — close the tab and come back.
Canvas Agent is now open to everyone — no request needed
Canvas Agent is no longer in closed beta. Every signed-in account can use it — including brand-new ones.
Getting started
Open any canvas; the Agent panel is on the right. Just say what you want:
- "Break this script into 6 shots and generate an image for each"
- "Make a character sheet for this character, four angles"
- "Look at these clips on the canvas and flag the ones where the face drifts"
It lays out a plan for you to look over before it starts, and the work runs in the background — close the browser and pick it up later.
What it can drive
Generating images, video and audio; editing images; looking at images and video; reading the canvas; creating nodes and edges; tidying the layout — most of what you can do by hand on a canvas, you can hand to it instead.
You pick its brain model yourself (Claude, GPT, Gemini, DeepSeek, Kimi and others) from the top of the Agent panel.
Spending stays visible
Every step the Agent takes is itemised in your billing details, there to check any time. If your balance runs short it stops and tells you, rather than pushing on regardless.
Node tools: hover to open, laid out as a grid
- Hover to open — no click needed. With a node selected, move your cursor onto the dot at its top-right and the tools open on their own. Move away and it closes by itself, or click the ← at the top left.
- The dot grows into the panel rather than popping a separate window beside it — your cursor never has to cross a gap to reach the menu.
- Tools are laid out as a grid, so the panel is about half its old size and every tool is a shorter trip away.
- Tool names appear at the top of the panel: whichever icon you rest on, its name is spelled out up top.
- Touch devices still open by tapping; Apple Pencil hover works too — on the same iPad, finger taps and pencil hover each do the right thing.
- Clicks are ignored until the open animation finishes, so you can't trigger a tool while your hand is still on its way over.
Zoom out to 5% — see the whole canvas at once
- Minimum zoom goes from 25% down to 5%. On a canvas with many nodes you can now take in the whole layout in one screen instead of panning around.
- Below 14% it switches to a minimal overview: nodes become plain blocks, leaving just structure and position. A line at the bottom of the screen tells you how to get back — zoom in and node content returns exactly as it was.
- Animations now stop as you zoom out. At 20% and below, generating effects, selection glow and edge flow all stop; which nodes are still generating stays visible as a static marker, so nothing is lost. The further out you go, the calmer the canvas — and your machine stops rendering animation you can't see anyway.
- The zoom ruler gained ticks for the low end and now marks where the minimal overview begins, so you can see the boundary before you hit it.
- Also fixed a grey haze the background dot grid produced at very small zoom levels.
Batch download now keeps your node titles
- Batch downloads now use your node titles. Select a batch, download once, and the files come out named like
Xiaoyu_Episode4_Outfit.png— no more guessing which random string is which. - Nodes you never renamed no longer share one name — the start of the prompt becomes the filename, e.g.
A girl in white before the snow mountain.png. - Real duplicates get a counter (
_2,_3), so nothing overwrites anything else in the same batch. - Markdown exports from text nodes follow the same naming.
- Batch downloads now use your node titles. Select a batch, download once, and the files come out named like
Subtitle removal paused · Common actions moved to the top
Subtitle removal is paused
Video subtitle removal relies on an external service that is currently down. Rather than let you start a job, wait, and get a failure back, we have taken the feature offline for now.
It is hidden from the video node's tool list, and the canvas agent will no longer call it either. It will return once the service is fixed — we will post here when it does.
Videos you have already generated are unaffected.
Tool list: the everyday actions moved to the top
Download, Set as cover, Expand and Info now sit at the top of the list. These are the "take the result away" actions you reach for most often, and they used to sit below a stack of editing tools.
Everything else keeps its order.
Node tools now live behind one small button
The toolbar is now a single small dot
Selecting a node used to park a wide toolbar right above it. On image nodes that's up to 9 buttons — about 380px wide, while a portrait node is only 280px. The toolbar was wider than the picture.
Now it collapses into a small dot at the node's top-right corner. Click it and the tools open as a labelled list beside the node, clear of the image. It closes again once you pick something.
No more guessing what an icon means
The old toolbar was icons only — you had to hover to find out what each one did. Every entry now carries its name: Redraw, Erase, Enhance, Outpaint, Remove background, Multi-angle, Panorama, Lighting, Annotate, Crop, Grid split, Info, Download, Set as cover, Expand.
Tools that were buried under "More" are one click away
Outpaint, remove background and multi-angle used to sit behind a second "More" menu. They're in the main list now.
Image, video, audio and document nodes all work this way
Text nodes keep their formatting bar as is — bold, italic, headings and lists need to stay one click away while you are writing.
Zooming never puts it over your artwork
The dot keeps the same size at any zoom level, and its distance from the node scales with the node, so it never ends up on top of your image when you zoom out.
A third desk pet: Xiaoyu, as a figure
There are three canvas pets now.
Xiaoyu
The new one is Xiaoyu, styled as a collectible figure — one plastic material, one lighting setup, oversized white hoodie and black shorts.
She has the same full animation set as the other two: idle, blink, glance around, wave, beckon, cheer, slump, being picked up, landing back down — plus a heads-down typing pose for when work is in progress.
How to switch
Right-click the pet on your canvas → Switch pet, then click the one you want. The swap is instant and remembered per device.
Gugu Gaga and Doro are both still there, and the default is unchanged.
Switching models now keeps your settings
Switching models no longer resets your settings
Changing the model on a video node used to snap resolution straight back to that model's default — your carefully chosen 4K or 1080P, gone. The rule now:
- The new model also has your current tier → kept exactly as is
- You picked 480P, the new model starts at 720P → lands on 720P
- You picked 4K, the new model tops out at 1080P → lands on 1080P
Duration works the same way. Your prompt and the audio toggle never change.
Portrait stays portrait
When the new model does not support your aspect ratio, it used to become 16:9 every time. Now it snaps to the closest shape instead: 3:4 lands on 9:16, still portrait.
MiniMax H3 keeps 768P
The cheaper 768P tier that shipped today used to jump back to 2K (40% more) if you switched away and back. It stays put now.
Grok Imagine 1.5 keeps 1080P
Switching to it used to force 720P with no way back. Text-to-video and image-to-video now both hold 1080P.
The panel only shows options that actually do something
The "Adaptive" aspect ratio is gone for Gemini Omni and Grok Imagine 1.5 — both always render 16:9, so that button never had any effect.
MiniMax H3: a cheaper 768P tier, and reference videos now work
A 768P tier, 29% cheaper than 2K
H3 now has a 768P resolution option. 50 Xins for 4s (2K is 70), 100 for 8s (2K is 140). Draft at 768P, finish at 2K.
Reference videos and audio now work
H3's reference tray takes three kinds of material: up to 9 images, 3 videos, and 3 audio clips, 12 files total.
Editing needs no separate mode
H3 has no separate "edit" entry point — attach a reference video and just say what to change in the prompt. Attach a clip with Chinese subtitles, write "replace the text on screen with English", and the layout and footage carry over.
Clip lengths are checked up front
Reference videos and audio run 2–15 seconds each, 15 seconds combined. If a clip is over, you hear about it before you're charged — not after the render.
Seedance 2.5 is live: 30-second clips you can edit and extend
Seedance 2.5 is now available on the canvas.
Up to 30 seconds per clip
The Seedance 2.0 family caps at 15 seconds; 2.5 doubles that to 30 seconds. It supports 480p and 720p, with up to 30 reference images.
Video editing: change what you don't like, in place
Video nodes now have an Edit video button in the toolbar. It opens a player — step to the exact frame you want to change, draw on it to point at the spot, then describe the change.
Every frame you capture drops a timestamp tag into the prompt automatically — no counting seconds by hand:
00:02 change the outfit to blue 00:03 switch the background to a night bar scene 00:09 make the hair dark blue with highlightsVideo extension: add to either end
Two new entries on the panel:
- Extend backwards — generates what happened before the clip begins
- Extend forwards — continues the story after it ends
You pick the output length, 4 to 30 seconds. The output ratio follows the source video.
How reference videos are priced
Correction (2026-08-10): this section previously said reference videos were often cheaper and showed a "32% off" table. That was our miscalculation. They are not cheaper — the correct explanation is below.
Seedance 2.5 charges a lower rate when your request includes video input. But video-input requests also carry a minimum billing amount, and the two cancel each other out almost exactly.
The result: attaching a reference video costs about the same as text-only generation, and only becomes more expensive once the reference runs longer than two-thirds of your output.
For a 30-second 720p output:
Credits Text-only 867 With a 4s reference 865 (same) With a 10s reference 865 (same) With a 20s reference 865 (same) With a 25s reference 951 (10% more) So attach a reference when your work needs one — there is no need to keep clips short just to save credits. The credit figure on the panel is live, so you can see the exact price before you generate.
Collapsed nodes now show their settings
- MiniMax H3 now shows its resolution ("2K") when collapsed — every other model already did; H3's slot was blank.
- More settings on the collapsed bar: Seed Audio's volume, pitch and sample rate; ElevenLabs' stability; Grok Imagine's "Pro" mode.
- A chip only appears once you change a default — a node you haven't touched stays exactly as wide as before.
- Video-edit resolution is visible too: 720p and 1080p differ by 2× in price, so you can now see which one you picked without expanding.
- When Kling has a reference image, the output ratio follows that image and the panel doesn't let you change it — so the collapsed bar no longer shows a ratio that does nothing.
- The "generate audio" toggle is gone for Gemini Omni and Grok Imagine 1.5: audio for those two is decided upstream, so the switch had no effect.
- Z-Image's "Advanced settings" (negative prompt, steps, seed) have been removed: we measured that the upstream simply doesn't read them, so changing them did nothing — better no knob than a knob that isn't wired to anything.
MiniMax H3: a 4-second option, and room for more references
- A 4-second length — 20% cheaper than 5s, handy when you are still testing a prompt.
- More reference material: up to 9 images, 3 videos and 3 audio clips, 12 files in total (was 9 combined). Mix them freely.
- Go over the limit and you get told right away, with nothing charged — no more quietly dropping the extras.
Both DeepSeek brains are now 50–63% cheaper
The new prices
In the canvas agent's brain picker, both DeepSeek tiers (per million tokens):
Brain Input Output DeepSeek V4 Flash 30 → 12 Xins 60 → 23 Xins DeepSeek V4 Pro 109 → 55 Xins 218 → 109 Xins Already live — nothing to configure. New conversations bill at the new rate.
Who can use them
- DeepSeek V4 Flash is available on the free tier. It is now the cheapest brain in the picker, with output an order of magnitude cheaper than the other free brain. For everyday work — editing the canvas, creating nodes in bulk, setting parameters — it is plenty.
- DeepSeek V4 Pro is on the paid tier, and worth switching to for long context (1M) and harder reasoning.
Flash also moved to the official July 31 release in the same update.
Switching brains
Open the agent panel (click the desk pet at the bottom-right of your canvas, or press ⌘I / Ctrl I). The brain picker is at the top, and each brain shows its input/output price right underneath.
Reference materials, reworked: a tray, numbering, and video refs on Wan 2.7 & MiniMax H3
The video, image and audio panels now share one layout, with reference materials collected in a tray above the prompt.
Thumbnails carry numbers, and the number is the @ in your prompt. The third image shows a 3; writing
@image3in the prompt points at exactly that one. Drag a thumbnail to reorder and the numbers follow.Past twelve, they fold. The last slot becomes a counter button; opening it reveals a panel to the right holding everything, split into image / video / audio sections, with drag-to-reorder inside. Prompt and thumbnails stay on screen together.
Wan 2.7 and MiniMax H3 now take video references. Attach video nodes in reference mode — up to 3 clips. MiniMax H3 additionally accepts images, video and audio at once.
Seed Audio reference voices are now picked from the canvas. Hit the reference button and select an existing audio node instead of uploading the file again.
Each model's reference-image limit is now honoured by the panel, so nothing sits in the tray that the model would never actually use.
New model: MiniMax H3
- MiniMax H3 joins the video models: text-to-video, image-to-video (with optional last frame), and multimodal reference (image + video + audio). 2K, 5–15s, native audio, 17.5 Xins/sec. See the announcement.
Grok Imagine 1.5: up to 7 reference images, and text-to-video
Multi-image reference: up to 7 at once
Put the person, the product and the look into separate reference images, then name each one in your prompt with
<IMAGE_0>,<IMAGE_1>:Handheld UGC-style clip. The person from
<IMAGE_0>holds the skincare jar from<IMAGE_1>and talks to camera, casual phone-camera framing.Images map in wiring order — the first is
<IMAGE_0>, the second<IMAGE_1>, and so on.Name every image you attach. An image you pass but never mention gets ignored, or blends into the shot unpredictably.
Good for: one character across a whole series, talking-head product clips with the real product in frame, or borrowing the palette and texture of one image for a new scene.
Runs from text alone
Write a prompt and go — no starting frame required. Attach an image and it animates from that frame; attach nothing and it generates from the text.
1080p added
Text-to-video and image-to-video now support 1080p. Multi-image reference mode tops out at 720p.
Lower price per second
Quality Before Now 480p 11.2 Xins/s 10 Xins/s 720p 19.6 Xins/s 17.5 Xins/s 1080p — 31.25 Xins/s Already live, nothing to switch on.
It comes with sound
Picture, lip-synced dialogue and ambient effects are generated together in one pass, not dubbed afterwards. Any whole number of seconds from 1 to 15, seven aspect ratios.
Let your AI assistant tidy your canvas directly
- Name your nodes, then actually find them. Your AI assistant can now set node titles directly. Once a canvas is titled, you can search it by name instead of scrolling through hundreds of nodes.
- Wiring, cleanup and parameter changes too. It can link nodes as references, remove leftover draft nodes, and switch a node's model / aspect ratio / resolution — parameter changes are saved as a draft only: nothing is generated and no Xins are spent until you run it.
- Your finished work is protected by default. Nodes that already hold a result, and nodes still generating, cannot be deleted or edited unless you explicitly say so — "tidy up my canvas" will never quietly take your paid output with it.
- Prompt text still lives in the canvas editor. Rich-text prompts and
@图片Nreferences are position-linked, so that stays where it's handled properly. - Applies to personal canvases; team canvases are collaborative and are still edited in the UI.
To get it: MCP users restart the client; CLI users run
npm i -g @xinyuai/cli@latest.The canvas agent is now available to everyone
Now open to everyone
Open any canvas and click the desk pet in the bottom-right corner, or press ⌘I (Ctrl I on Windows), and the agent panel appears.
What it does: read what is already on your canvas, create generation nodes on request, set the model and parameters, and wire up reference images. Once the nodes are prepared, you press “Generate all” to actually start them — whether and when anything is spent stays your call.
Two brains are free
- Gemini 3.5 Flash
- DeepSeek V4 Flash
Both are available without paying. Ten stronger brains (Claude Opus 5 / Sonnet 5, Gemini 3.1 Pro, the GPT-5.6 family, Kimi K3, DeepSeek V4 Pro and more) are on the paid tier and are marked as such in the picker.
It tells you the truth about what it did
- While nodes are still generating, the task card reads “Media still generating · N still running”, and only says complete once they all finish.
- A plan step that could not be ticked automatically is marked ❓ “not auto-confirmed”, with a note that the work may well have been done — we just could not confirm it automatically. No guessing.
- Image models do not all offer the same quality tiers; the agent only picks from the tiers a given model actually supports, so the tier and price on the confirm card are exactly what runs and what gets charged.
Tell us when it is wrong
Every reply has a 👎 and a ⚠️ button. If a run goes sideways, or it does something you did not ask for, press one — that is the only way we can see which specific run went wrong.
Canvas agent: an unticked step now tells you what actually happened
The canvas agent is in whitelisted beta — these changes are visible only on accounts that have it enabled.
An unticked step no longer leaves you guessing
Each step in the plan is ticked automatically as it runs. The tick is matched by the tool name written on that step — so if the agent reaches the same outcome with a different tool, that step never gets ticked.
Previously it just sat there blank, and you could not tell whether the work had been skipped or done without being ticked.
Such a step is now clearly marked ❓ “not auto-confirmed”, with a line under the plan: it may well have been done — we just could not confirm it automatically, so the result and the canvas are the better check.
Why not simply say “not run”
Because that would usually be wrong. What normally happened is that the agent used a different tool to achieve the same thing. Saying “not run” would send you off to redo work that is already finished.
We state only what we actually know: this step could not be auto-confirmed.
Also
Failed or stopped runs are unaffected — there, a step that never ran is already explained by the run itself.