Nearly every mainstream background remover was trained to look at one photograph at a time. Cel has to get a result right, and get it the same way, across dozens or hundreds of consecutive frames — a different, harder problem that almost nobody has built training data for. Cel is RemoveGifBG's answer, and it's being built in the open.
Temporal alpha matting — the exact model already running inside RemoveGifBG
Built on open, MIT-licensed segmentation research — see "How Cel is built" below
The animation dataset, and the pipeline that keeps a result consistent frame to frame
A cel — short for celluloid — is the transparent sheet an animator traditionally painted a character onto, then laid over a separately painted background so the two could be shot together without redrawing the whole scene for every frame. That's literally what this model produces: character art, on transparency, background gone.
A single-image tool only has to be right once. A GIF or video tool has to be right the same way dozens or hundreds of times in a row — and the failure mode that creates is almost invisible in a photo benchmark. A background patch that's 98% correctly removed across a whole clip doesn't look 98% good; if it flickers back in on 2 frames out of 60, it looks broken, because motion makes inconsistency obvious in a way a single wrong pixel in one photo never is.
There's a second problem that's specific to hand-drawn and vector animation: a character is sometimes filled with the same white, black, or flat color as its own background. A photo-trained model has never had to solve that ambiguity — real photographs don't put a white shirt on a pure-white seamless backdrop on purpose. Animation does it constantly. In our own review of the public models and tools in this space, we didn't find one that handles it reliably.
Honestly: early. The current training set is 2,052 hand-reviewed frames across 20 animated clips — enough to prove the approach works, nowhere near enough to call the problem solved. For scale, a comparable public fine-tune (ToonOut, an anime-focused segmentation model) moved its pixel accuracy from 95.3% to 99.5% using roughly 1,200 labelled images of a single, narrower style. An animation-general dataset — GIFs, sprite sheets, stream overlays, sticker loops, every art style — needs to be a full order of magnitude past that, and every frame of it is reviewed by a person, not just auto-labelled.
That gap between "works" and "solved" is exactly what the dataset campaign exists to close.
The goal isn't a general-purpose segmentation model — it's the best background remover for animation specifically, and "best" isn't a slogan here: it's measured category by category on the public benchmark, updated as each one moves from open to solved.
How Cel is built
Cel is built on open segmentation research — the same open, MIT-licensed foundation used across this field — and retrained on our own animation dataset. The temporal matting pipeline around the model — the part that keeps a result consistent across every frame of a clip, and that fills in flat-color content standard tools miss — is our own work.
Solved: flat-color fills (a white shirt on a white background is filled in, not left as a hole), frame-to-frame consistency (a removed background element doesn't flicker back on isolated frames), and real 3D or photographic backdrops. Still open: two separate near-background-colored objects sitting close together, and very faint low-contrast line art. The full category-by-category status, and how each one is scored, is on the benchmark page.
Compute is the cheap part of training a model like this — a full fine-tuning run costs a few dollars of GPU time. The real cost, and the real bottleneck, is people reviewing frames: correcting an auto-generated mask, confirming an edge, flagging what still looks wrong. Funding goes toward:
The dataset campaign runs on a crowdfunding platform — it's not open yet. Register your interest and we'll email you the moment it launches. Every tier gives more credits per dollar than any pack on the site — 25–50% more — and the credits never expire.
Tier preview — final rewards and pricing are set on the campaign page at launch.
Or just use RemoveGifBG today — every file that trips Cel up becomes tomorrow's training data. Try it free.
Cel is the alpha matting model at the core of RemoveGifBG — the model that decides, frame by frame, which pixels belong to the animated subject and which are background. It's built on open segmentation research and retrained on RemoveGifBG's own dataset of animated frames, a kind of training data most public models never see.
A cel — short for celluloid — is the transparent sheet an animator traditionally painted a character onto, then laid over a separately painted background so the two could be shot together. That's literally what this model produces: character art, on transparency, background gone.
Neither — it's the engine, not a SKU. Every job you run on RemoveGifBG (in the app or via the API) already uses Cel; there's no separate model to select, no "Cel plan." This page exists so you can see what's actually under the hood and how it's improving.
The architecture is — Cel is built on open, MIT-licensed segmentation research used widely across this field. What's ours, and not open, is the animation-specific dataset and the temporal pipeline built around the model to keep a result consistent across every frame, not just accurate on one frame in isolation.
Almost every public segmentation model is trained on single photographs. A GIF or video needs the same subject correctly identified dozens or hundreds of times in a row, with the result agreeing with itself frame to frame — a background patch that flickers in and out is far more visible than one slightly wrong pixel in a still photo. Photo-trained models were never asked to solve that.
Two ways. Use RemoveGifBG and tell us about any file it doesn't handle cleanly — those become training data. Or back the dataset campaign (see above) when it opens: your money goes directly toward labelled frames, which is the actual bottleneck.