Benchmark Leaderboards
Loading leaderboardsβ¦
Loading leaderboardsβ¦
Track which models hold the current SOTA on the benchmarks the field actually uses. Click a column header to sort; the leading row is highlighted.
Class-conditional image generation quality β Frechet Inception Distance (FID) on ImageNet at 256x256, the canonical benchmark for image generators. Lower is better. Values use classifier-free guidance, traced to each method's paper.
Metric: FID β lower is better Β· Source β
| # β² | Model / Paper | FID | Ξ vs SOTA | Links |
|---|---|---|---|---|
| 1 | REPA (SiT-XL/2 + REPA)SOTA | 1.42 | β | github.com β |
| 2 | TexTok (DiT + text-conditioned tokenizer) | 1.46 | +0.04 | arxiv.org β |
| 3 | MAR-H (Diffusion Loss) | 1.55 | +0.13 | github.com β |
| 4 | VAR-d30 (Visual AutoRegressive) | 1.73 | +0.31 | github.com β |
| 5 | M-VAR-d32 | 1.78 | +0.36 | arxiv.org β |
| 6 | SiT-XL/2 | 2.06 | +0.64 | github.com β |
| 7 | DiT-XL/2 | 2.27 | +0.85 | github.com β |
| 8 | Generative Modeling via Drifting | 3.36 | +1.94 | arxiv.org β |
| 9 | LDM-4 (Latent Diffusion) | 3.6 | +2.18 | github.com β |
| 10 | Diffusion Image Generation with Explicit Modeling of Data Manifold Geometry | 3.62 | +2.20 | arxiv.org β |
| 11 | ADM-G (Guided Diffusion) | 4.59 | +3.17 | github.com β |
| 12 | FrequencyBooster: Full-Frequency Modeling for High-Fidelity Pixel Diffusion | 4.64 | +3.22 | arxiv.org β |
| 13 | Contrastive Flow Matching | 11.14 | +9.72 | arxiv.org β |
| 14 | One-step Latent-free Image Generation with Pixel Mean Flows | 34.61 | +33.19 | arxiv.org β |
| 15 | Mean Flows for One-step Generative Modeling | 328.91 | +327.49 | arxiv.org β |