Photo background removers get judged one image at a time. That's the wrong test for a GIF or video — a tool can get every single frame almost right and still look broken once they're played back, because a flicker is far more visible than one imperfect pixel. Here's how we actually test RemoveGifBG's model, category by category, including where it still fails.
Does the alpha match the character's actual silhouette on each individual frame
Does the result hold steady across the clip instead of flickering on isolated frames
Flat-color fills, scene backgrounds, photo-real content — scored separately, not averaged together
Before testing model quality, it's worth being clear about why a head-to-head table against the usual names wouldn't be a fair test in the first place:
Once you account for that, an honest animation benchmark mostly ends up testing one thing: the quality of the underlying matting model on sequences. That's what the rest of this page measures.
An internal library of real animation content — sticker packs, VTuber overlays, game sprite sheets, short clips — spanning white/black flat backgrounds, colored scene backgrounds, and photo-realistic content. Not synthetic, not cherry-picked for a good result.
| Category | What it tests | Status |
|---|---|---|
| Clean, simple content | The majority case — no background ambiguity, a clear subject | Solved |
| Flat white/black backgrounds | Character fill painted the same color as its own backdrop | Solved |
| Static-scene flicker | A background element staying removed consistently across every frame | Solved |
| Scene / photographic backgrounds | Real 3D-rendered or photo backdrops behind an animated subject | Mostly solved |
| Adjacent near-background elements | Two separate objects, both close to the background color, sitting near each other | Open |
| Very low-contrast line art | Faint white-on-white artwork below normal color-difference thresholds | Open |
Two categories remain genuinely unsolved and are published here rather than left out. Both are the direct target of the dataset expansion described on the Cel page.
Fill accuracy: every clip is reviewed frame by frame against the source art, by a person, not just an automated diff. We publish the method, not just a score, so "passing" means something concrete you can check for yourself rather than a number with no definition behind it.
Frame-to-frame consistency: each frame's result is compared against its neighbors. A real edit changes smoothly from one frame to the next as the subject moves; a flicker doesn't — a pixel that's confidently background in every neighboring frame but suddenly isn't in one frame is a strong signal something's wrong, independent of whether that single frame "looks fine" in isolation. On one internal test clip — a static desert scene behind a moving character — a sand patch that should have stayed removed the whole time fully reappeared on 2 of every 5 frames before this was fixed; after, 0 of 5.
Two categories, published rather than hidden:
Full detail on both, and what closes them, is on the Cel page.
Try It on Your Own File — Free
Building the model that closes the gaps above: see Cel
Not yet — it's our own internal test methodology, published openly so anyone can see exactly what we test and how we score it, including the categories that still fail. An independently-run comparison is on the roadmap; see the Cel page for what funding it unlocks.
Not head-to-head, because they're not solving the same problem: remove.bg has never supported animated GIFs, ezgif's tool is a color-key filter rather than a model, and Unscreen shut down in December 2025 and was folded into Canva Pro. Once those are accounted for, an honest animation benchmark ends up mostly testing matting-model quality itself — which is what this page measures.
A category passes when every clip in it is reviewed frame by frame against the source art and holds up on two measures: the alpha matches the character's actual silhouette, and the result doesn't flicker.
An internal library of real animation content spanning sticker packs, VTuber overlays, game sprite sheets, and short clips — not a synthetic or cherry-picked benchmark set.
Whenever a category's status actually changes, not on a fixed schedule. Last verified August 2026.