News & Updates
Product updates, new models, and platform announcements.
Nano Banana 2.1 is here: four characters, any scene, same faces
Google's newest image model, with a big jump in multi-character consistency. Attach your character references and they keep their looks across scenes and poses. Now in the Featured section, 50% off.
Read →
Seedance 2.5 is live: 30-second clips, change one part, extend from the end
Seedance 2.5 is live on the XinYu AI canvas: up to 30 seconds per clip, change one part of a shot while the rest stays, extend a video you already have from its end, attach 30 images / 10 videos / 10 audio tracks, and let a music track drive the cutting. Output at 480p, 720p or 1080p. The film above is Seedance 2.5's own work: 1080p, one 30-second generation, made on the canvas.
Read →
Membership Plans Are Live: Three Tiers, Credits That Never Expire
XinYu AI now offers BASIC / PRO / ULTIMATE plans — monthly or annual, with credits that never expire and bigger top-up bonuses.
Read →
The XinYu AI desktop app: your folders, right next to the canvas
The desktop app for Mac and Windows is here: a local files panel, downloads saved where you choose, system notifications when generations finish, a privacy lock with screenshot protection, and automatic updates. With a one-shot intro film.
Read →
Why a trimmed reference video gets rejected: it plays fine, but the model says Bad Request
A clip trimmed without re-exporting still carries the cut data at its start. Players skip it; the model's decoder doesn't. Two ffprobe commands to check, one ffmpeg command to fix, and why the popular no-re-encode fix puts the trimmed part back.
Read →
Writing Seedance 2.5 video edit prompts: removing is easy, replacing is hard
Removal works first time; replacement kept failing (the new subject is added, the old one stays). Change one spot and the whole frame is redrawn (23.7 dB vs 42.6 dB for no change). Boxes and arrows don't help. Timecodes held on an 18-second clip and did nothing on a 4-second one. Five rules for prompts that work.
Read →
How Seedance 2.5 pricing works: what resolution, length and a reference video each add
1080p costs 2.46× of 720p, not 2.25×. Audio and reference images are free. A short reference video barely changes the price, but past about 2/3 of the output length each extra second costs about 0.6 of an output second. An edit is about 1.2× a text-to-video clip of the same length. Every ratio comes from our production billing functions.
Read →
Free subtitle cleanup for XinYu videos: preview, then export a copy
Clean up fixed white subtitles for 0 Xin. Select an area and time range, preview, then export a copy while preserving the original and audio. For verified XinYu-generated videos only; uploads excluded. Up to 30 seconds and 1080p, at 24 or 30 fps.
Read →
Seedance's cheaper input-video tier saves you nothing. We measured it.
The rate card really does price input-video jobs about 40% lower. Controlled test — same prompt, 4 seconds, 480p, the only variable being a 2-second input video — returned 38,830 versus 67,652 billed tokens: about 4% more, not less. Work through the formula and the discount turns out to be uncapturable at any duration.
Read →
When a generation is rejected, do you get your credits back?
Rejections happen in two different places: one stops at the door in about 6.5 seconds and costs nothing; the other generates the clip first, fails at 17–18 seconds, and bills anyway. Our rule keys on whether we were actually charged — never on the model name or the error text. When in doubt, you get refunded.
Read →
Why Seedance rejected your reference image: the 0.40–2.50 aspect ratio limit
"expected the aspect ratio to be between 0.40 and 2.50" — the error never tells you which image caused it. We measured all three reference images from a real failed job, mapped the common aspect ratios, and explain why the same image is perfectly legal on Wan 3.0.
Read →
GPT Image 2.5 is live: Flare and Sunburst, five refinement tiers, transparent output
OpenAI's GPT Image 2.5 is now on the XinYu AI canvas: two variants, five refinement tiers, up to 16 reference images per run, and PNG output with a real alpha channel.
Read →
MiniMax H3 at Half Price
Our upstream cut its price and we are passing all of it through. For two weeks, MiniMax H3 is 50% off at every resolution and in every mode.
Read →
MiniMax H3 is live: reference with images, video, or audio
A new video model: MiniMax H3. Beyond text- and image-to-video, its reference mode accepts images, video and audio at the same time — hand it a clip for motion, a track for rhythm. 2K output, any length from 4 to 15 seconds, with native audio.
Read →
Canvas Agent: describe it once, let it finish the job
Tell the Agent what you want, right on the canvas. It reads the canvas, plans, generates, checks its own output, and fixes what's off — in the background, so you can close the tab. In closed beta, rolling out gradually.
Read →
Now Live: Seedream 5.0 Pro — Precise Editing · Layer Control · Multilingual Text
ByteDance's Seedream 5.0 Pro is now live on the XinYu AI canvas. It moves from "generation" to "controllable creation" — precise local edits, layer separation, and multilingual typography. Available with standard credits, no extra subscription.
Read →
Full-Power Seedance 2.0 in Native 4K
Full-power Seedance 2.0 now renders native 4K with 10-bit HDR — idea to final cut on one canvas. Lock faces with reference images, control pacing with reference video.
Read →
Seedance 2.0 Mini — Try One on Your Signup Credits
The faster, cheaper Seedance 2.0 Mini is here at half the per-second cost. Your signup credits cover a 4-second Mini clip — see it before you commit.
Read →
Xin TV Is Here — Get Seen, Get Paid
Share your canvas cuts to Xin TV in one click — 10G uploads, auto quality ladder, and viewers can tip or pay to unlock, straight to your wallet.
Read →