Which approach is actually better?
The honest answer is that the published numbers do not support a clean ranking, and the reason is worth understanding before reading any table below.
What you can conclude
- Pushing semantic structure into the latent reliably speeds up training. Within a single paper, holding the backbone fixed, representation alignment and frozen foundation encoders both cut the steps needed to reach a given quality by large factors. These are the most trustworthy claims in the field because they are measured against the paper's own baseline.
- At their best-tuned settings, pixel-space, VAE-latent, and representation-latent models all land in roughly the same 1.0 to 2.0 gFID band on ImageNet 256. No family is running away with it.
- Reconstruction quality does not predict generation quality. The clearest single example is inside one paper: the encoder giving the best reconstruction is not the one giving the best generation.
What you cannot conclude
- That any one family beats another. Guidance method, model size, and training budget vary simultaneously across the papers making these claims, and each alone can move the number by more than the gap being argued about.
- That the guidance column is a detail. One paper reports the identical model at 1.65 with no guidance, 1.49 with classifier-free guidance, 1.14 with autoguidance, and 1.06 with representation guidance. That 0.59 swing comes from the sampler alone, and it is larger than most of the architectural gaps in the table below.
- That text-to-image scores are precise. Three papers report the same public checkpoint at 0.66, 0.67, and 0.82 on the same benchmark, purely from differences in prompt rewriting and sampler settings. Treat gaps under roughly 0.05 as noise.
ImageNet 256, with guidance
Lower is better. The guidance column is why this table is not a leaderboard.
| gFID | Paper | Line | Guidance | Budget | Params |
|---|---|---|---|---|---|
| 1.03 | PAE | Semantic-plus-detail hybrids | not stated in abstract | 13x faster convergence claimed | not stated |
| 1.06 | RAEv2 | Representation-space generation | REPA-Guidance | 80 epochs | 839M |
| 1.12 | REPA-E | Representation-aligned latents | CFG (scale not stated) | 800 epochs | 675M |
| 1.13 | RAE | Representation-space generation | AutoGuidance | 800 epochs | 839M |
| 1.35 | VA-VAE / LightningDiT | Representation-aligned latents | cfg-interval + timestep shift (scale not stated) | 800 epochs | 675M |
| 1.42 | REPA | Representation-aligned latents | guidance-interval | 800 epochs (~4M iters) | 675M |
| 1.55 | MAR | Continuous-token autoregression | CFG (scale swept, value not stated) | 800 epochs | 943M |
| 1.67 | MAETok | Representation-aligned latents | cfg=2.0, 250 steps | 4M steps | 675M generator + 176M tokenizer |
| 1.73 | VAR | Discrete-token autoregression | guidance ratio 2.0 + rejection sampling | 350 epochs | 2.0B |
| 1.8 | FlatDINO | Semantic-plus-detail hybrids | not stated in abstract | not stated | 675M (DiT-XL) |
| 1.82 | JiT | Single-stage pixel transformers | CFG-interval | 600 epochs | ~2B |
| 1.92 | SVG | Representation-space generation | CFG (scale unresolved in source) | 1400 epochs | 675M |
| 2.06 | SiT | Flow matching and rectified flow | cfg=1.5 | 7M steps | 675M |
| 2.27 | DiT | VAE-latent diffusion | cfg=1.5 | 7M steps | 675M |
PAE: Taken from the abstract; the guidance method and budget behind the headline are not stated there. Needs a PDF-level check before being treated as directly comparable.
RAEv2: The paper's own Table 4 shows the identical model at 1.65 (no guidance), 1.49 (CFG), 1.14 (AutoGuidance) and 1.06 (REPA-Guidance). Guidance method alone moves this number by more than 0.5 FID, which is larger than most claimed architectural wins in this table.
RAE: AutoGuidance is a stronger guidance method than the plain CFG used by the VAE baselines in most cross-paper comparisons, and contributes a large share of the gap.
VA-VAE / LightningDiT: The widely quoted '21x faster than DiT' refers to the 64-epoch checkpoint (gFID 2.11), not to this 800-epoch headline number. The speed claim and the quality claim come from different rows.
REPA: Headline uses guidance-interval, not the plain CFG its DiT/SiT baselines use. With plain CFG (w=1.35) the same model reports 1.80.
VAR: arXiv v1 reported 1.80 for this configuration and the current version reports 1.73; cite which version you mean. Uses rejection sampling, which the diffusion baselines do not.
FlatDINO: Headline claim is efficiency (8x fewer FLOPs per forward pass), not raw FID.
JiT: Uses roughly 2B parameters against the 675-839M of the latent-space rows, and a guidance-interval technique. No unguided number is reported, so this cannot be compared against the unguided column. Any claim that pixel-space now beats latent-space is confounded by scale and guidance simultaneously.
SVG: Trained far longer than most rows here (1400 epochs). At a comparable 500 epochs it reports 2.10. The source gives conflicting CFG scales (4.0 in a figure caption, 1.5 in an appendix table).
ImageNet 256, without guidance
The cleaner comparison, because it removes the single largest confound. Fewer papers report it, which is itself telling.
| gFID | Paper | Line | Budget | Params |
|---|---|---|---|---|
| 1.51 | RAE | Representation-space generation | 800 epochs | 839M |
| 1.65 | RAEv2 | Representation-space generation | 80 epochs | 839M |
| 1.69 | REPA-E | Representation-aligned latents | 800 epochs | 675M |
| 2.17 | VA-VAE / LightningDiT | Representation-aligned latents | 800 epochs | 675M |
| 2.31 | MAETok | Representation-aligned latents | 4M steps | 675M |
| 2.35 | MAR | Continuous-token autoregression | 800 epochs | 943M |
| 3.36 | SVG | Representation-space generation | 1400 epochs | 675M |
| 5.9 | REPA | Representation-aligned latents | 4M iters | 675M |
| 8.3 | SiT | Flow matching and rectified flow | 7M steps | 675M |
| 9.62 | DiT | VAE-latent diffusion | 7M steps | 675M |
Note what happens here. The gap between a plain VAE-latent transformer and a representation latent is enormous without guidance and much smaller with it. Guidance is doing a large share of the work that headline numbers attribute to the latent.
One-step and few-step
A separate contest. These models sample in one or a few passes rather than dozens, so they are not comparable to the tables above.
| gFID | NFE | Paper | Line | Note |
|---|---|---|---|---|
| 1.48 | 1 | Distilling Drifting Transformers | Natively few-step objectives | DiT^DH-XL backbone in RAE latent, one-step distilled |
| 1.54 | 1 | Drifting Models | Natively few-step objectives | ImageNet 256x256, latent space, one-step |
| 1.61 | 1 | Drifting Models | Natively few-step objectives | ImageNet 256x256, pixel space, one-step |
| 1.99 | 8 | IMM | Natively few-step objectives | ImageNet 256x256, 8 steps |
| 3.43 | 1 | MeanFlow | Natively few-step objectives | ImageNet 256x256, one-step, trained from scratch |
The result worth pausing on: one method reports 1.54 in latent space and 1.61 in pixel space with everything else held fixed. A 0.07 gap is weak support for the latent being essential.
Reconstruction quality of the latent itself
No generator involved. This is the ceiling the tokenizer imposes, and the point is that it does not predict the generation numbers above.
| rFID | Tokenizer | Paper |
|---|---|---|
| 0.16 | RAE with frozen MAE-B encoder | RAE |
| 0.28 | VA-VAE f16d32 | VA-VAE / LightningDiT |
| 0.48 | MAETok, 128 tokens | MAETok |
| 0.49 | RAE with frozen DINOv2-B encoder | RAE |
| 0.62 | SD-VAE f8d4, as re-measured by MAETok | LDM / Stable Diffusion |
| 0.65 | Frozen DINOv3 + residual detail branch | SVG |
| 0.69 | DC-AE f32c32 | DC-AE |
RAE: MAE-B gives RAE's best reconstruction but not its best generation — a compact illustration that rFID does not predict gFID.
VA-VAE / LightningDiT: The strongest cross-paper agreement in this table: three independent papers report the same value.
DC-AE: DC-AE reports no ImageNet 256x256 gFID at all; its generation results are 512x512 only, and its headline 'beats SD-VAE' result there uses 4x the training iterations.
Text-to-image
Higher is better. Read these as coarse bands, not rankings: the same checkpoint scores differently across papers, and the benchmark itself has measured drift.
| GenEval | System | Line | Params |
|---|---|---|---|
| 0.91 | Qwen-Image Qwen-Image after RL tuning | VAE-latent diffusion | 20B |
| 0.87 | Qwen-Image Qwen-Image base model | VAE-latent diffusion | 20B MMDiT + Qwen2.5-VL 7B encoder |
| 0.75 | SVG-T2I SVG-T2I 2.6B DiT, frozen DINOv3 latent | Representation-space generation | 2.6B |
| 0.74 | SD3 / MMDiT SD3 8B (depth 38) at 1024px, with DPO | Flow matching and rectified flow | 8B + 4.7B T5-XXL |
| 0.73 | Lumina-Image 2.0 Lumina-Image 2.0 | VAE-latent diffusion | 2.6B |
| 0.495 | Scale-RAE Best ablation config, deliberately budget-limited (~30K iterations) | Representation-space generation | 0.5B-9.8B swept |
Qwen-Image: Post-RL. Illustrates that RL alignment moves GenEval by ~0.04, comparable to the gap between several distinct systems in this table.
SVG-T2I: Reported with an LLM prompt rewriter, which is not comparable to scores measured on raw prompts. The same paper reports 0.82 for FLUX.1-dev under the same rewriter, where other papers report 0.66-0.67 for that identical checkpoint.
SD3 / MMDiT: SD3's own paper reports neither DPG-Bench nor MJHQ-30K; any such number attributed to SD3 is a third-party re-measurement.
Scale-RAE: Scale-RAE is a controlled scaling study, not a system release. Reading this as a state-of-the-art claim is a category error. The legitimate comparison is its within-paper ablation at matched budget: RAE 49.5 versus a VAE baseline at 39.6.
| DPG-Bench | System | Params |
|---|---|---|
| 88.32 | Qwen-Image | 20B |
| 87.2 | Lumina-Image 2.0 | 2.6B |
| 85.78 | SVG-T2I | 2.6B |
| 76.9 | Scale-RAE | — |
How to read a claim in this field
When a paper says it beats a baseline, check three columns before believing the headline: the guidance method, the training budget, and the parameter count. If any of the three differs from the baseline, the gap is not attributable to the idea being sold. The within-paper ablation, where those are held fixed, is almost always the more informative number, and it is almost always smaller than the abstract implies.