VFXBench is a collection of commercially relevant video editing and generation tasks for professional production pipelines.
| Rank | Model | Creator | Score | Cost / sec |
|---|---|---|---|---|
| 1 | Aleph 2.0 |  Runway | 67.0% | $0.415 |
| 2 | FLUX 3 Video |  Black Forest Labs | 59.6% | $0.030 |
| 3 | Seedance 2.5 |  ByteDance | 56.4% | $0.955 |
| 4 | HappyHorse 1.0 |  Alibaba | 50.1% | $0.140 |
| 5 | Gemini Omni Flash |  Google | 48.6% | $0.111 |
| 6 | Wan 2.7 | Alibaba | 43.5% | $0.100 |
| 7 | Kling O3 Pro |  Kling AI | 41.3% | $0.143 |
| 8 | Hailuo 03 |  MiniMax | 18.9% | $0.204 |
| Rank | Model | Creator | Score |
|---|---|---|---|
| 1 | Kling O3 Pro | Kling AI | 27.2% |
| 2 | Aleph 2.0 | Runway | 18.0% |
| 3 | Wan 2.7 | Alibaba | 13.6% |
| 4 | Seedance 2.5 | ByteDance | 13.4% |
| 5 | FLUX 3 Video | Black Forest Labs | 12.4% |
| 6 | Hailuo 03 | MiniMax | 11.1% |
| 7 | HappyHorse 1.0 | Alibaba | 10.5% |
| 8 | Gemini Omni Flash | Google | 8.1% |
| Rank | Model | Creator | Score |
|---|---|---|---|
| 1 | Kling O3 Pro | Kling AI | 41.8% |
| 2 | FLUX 3 Video | Black Forest Labs | 34.1% |
| 3 | Wan 2.7 | Alibaba | 33.7% |
| 4 | Seedance 2.5 | ByteDance | 32.1% |
| 5 | Aleph 2.0 | Runway | 30.1% |
| 6 | HappyHorse 1.0 | Alibaba | 28.2% |
| 7 | Gemini Omni Flash | Google | 24.3% |
| 8 | Hailuo 03 | MiniMax | 17.6% |
| Rank | Model | Creator | Score |
|---|---|---|---|
| 1 | Kling O3 Pro | Kling AI | 13.5% |
| 2 | Aleph 2.0 | Runway | 6.4% |
| 3 | Seedance 2.5 | ByteDance | 5.3% |
| 4 | HappyHorse 1.0 | Alibaba | 4.3% |
| 5 | FLUX 3 Video | Black Forest Labs | 4.1% |
| 6 | Wan 2.7 | Alibaba | 3.1% |
| 7 | Hailuo 03 | MiniMax | 2.9% |
| 8 | Gemini Omni Flash | Google | 2.2% |
| Rank | Model | Creator | Score |
|---|---|---|---|
| 1 | Seedance 2.5 | ByteDance | 37.5% |
| 2 | Wan 2.7 | Alibaba | 29.9% |
| 3 | FLUX 3 Video | Black Forest Labs | 24.1% |
| 4 | Kling O3 Pro | Kling AI | 22.0% |
| 5 | Aleph 2.0 | Runway | 21.8% |
| 6 | HappyHorse 1.0 | Alibaba | 21.3% |
| 7 | Gemini Omni Flash | Google | 19.6% |
| 8 | Hailuo 03 | MiniMax | 14.1% |
Each spoke is one leaderboard. Each shape is one model, drawn at its rank on that board, with first place on the outer edge. A point at the centre means the model is not ranked on that board.
Hover to preview, click to pin. A dashed edge marks a second model from the same creator.
The pipeline has three tiers, from a social clip to a feature archival master. The highest tier a model can round-trip is the ceiling on what a studio can hire it for. A model can accept a high-bit-depth file and still crush it to 8-bit internally. Only an ACES round-trip shows that.
Scored in ACES2065-1: ΔE-ITP, gamut drift, highlight preservation, banding. A model’s Fidelity Level is the highest tier it survives within a professional tolerance.
Remove a person or object from a live-action plate. Ground truth is a real second take of the same shot without them.
read the spec ↗We send a Rec.709 clip and ask for it back unchanged, so any shift is the model regrading footage it was told to leave alone.
read the spec ↗We send a clip and ask for it back unchanged, so any shift in where the blacks and whites sit is the model regrading footage it was told to leave alone.
read the spec ↗Each model is offered L3, then L2, then L1, and scored on wide-gamut survival at the highest door it opens.
read the spec ↗Each model is offered L3, then L2, then L1, and scored on dynamic-range survival (highlights, roll-off, scene-linearity) at the highest door it opens.
read the spec ↗Removal in reverse: whether the inserted element belongs. Locked-off counterfactual shoot, plus an influence mask for its shadow and reflections.
read the spec ↗Comp-ready alpha: keying, hair, transparency, despill. Stratified by difficulty and never averaged into one number.
read the spec ↗Scene balancing across a sequence of shots. Perturb-and-restore objective core, plus the benchmark's first expert gate.
read the spec ↗Banding, effective bit depth, chroma subsampling. Postponed from v0: film grain masks banding, so real-footage sourcing is unsolved.
read the spec ↗
VFXBench