Generative Vision Atlas

Research line · contested

Representation-space generation

Freeze a vision foundation model and generate directly inside its feature space; train only a decoder back to pixels.

What defines membership

The generative latent should be a pretrained semantic representation, not a code learned for reconstruction.

The assumption nobody examined

For most of the latent-diffusion era the autoencoder was treated as settled infrastructure. Stable Diffusion shipped one, everything downstream inherited it, and research attention moved to the denoiser. The diffusion transformer replaced the UNet backbone but kept the same latent. Rectified flow replaced the objective but kept the same latent. Text encoders migrated from CLIP to T5 to full vision-language models, and still the latent stayed put.

The assumption underneath was that a compression code is a neutral substrate: squeeze the image, run the generative model in the smaller space, decode. What that assumption misses is that an autoencoder trained purely to reconstruct pixels has no reason to organize its latent semantically. Two images differing only in compression artifacts sit far apart in such a space. Two images of the same object under different lighting may sit far apart too. The generative model is then asked to learn a distribution over a space that encodes appearance without encoding meaning.

The measurement that made it a research question

REPA turned this from a philosophical objection into a number. It added an auxiliary loss aligning a diffusion transformer’s intermediate hidden states to a frozen DINOv2’s features, changed nothing else, and cut the training needed to match a strong baseline by more than an order of magnitude.

That result admits two readings. The conservative one is that the generator benefits from a semantic teaching signal, so add the loss and keep everything else. VA-VAE and REPA-E take this route, pushing the alignment pressure into the autoencoder itself and then into joint end-to-end training. REG sharpens it further by entangling a semantic token into the denoising target rather than attaching a side loss.

The radical reading is that if aligning toward a semantic space helps this much, the semantic space was the better place to work all along.

Two groups arrive at the same answer within days

RAE and SVG appeared within days of each other in October 2025, independently, with the same core move: stop training an encoder at all. Freeze a pretrained vision foundation model, treat its feature space as the generative latent, and train only a decoder back to pixels.

This is a smaller change than it sounds and a larger one than it looks. Smaller, because the architecture is still a diffusion transformer and the objective is still flow matching. Larger, because it inverts the relationship between reconstruction and generation. In the VAE picture, the latent exists to be decodable and generation is a downstream consumer. Here the latent exists because it is meaningful, and decodability becomes the decoder’s problem.

Making it train required more than swapping the encoder. RAE reports that the decoder width has to meet or exceed the token dimension, a constraint no VAE-scale decoder ever had to satisfy, because these latents are high-dimensional where VAE latents are deliberately narrow. It needs a dimension-dependent shift in the noise schedule, and noise-augmented decoding, and a diffusion head rather than a linear projection to decode well. The recipe is specific, and the specificity is the evidence that high-dimensional semantic latents are a genuinely different regime rather than a drop-in substitution.

What the evidence supports, and what it does not

Scale-RAE carries the recipe to open-vocabulary text-to-image and reports something more interesting than a better score: under fine-tuning on high-quality data, VAE-based models overfit catastrophically by epoch 64 while RAE-based models remain stable through 256. Stability under long training is a different kind of claim from a better FID, and harder to explain away as tuning.

The FID numbers themselves need care. RAEv2 reaches 1.06 on ImageNet 256, which reads as a decisive win over a diffusion transformer’s 2.27 until you notice the guidance column. That same 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. The sampler accounts for a 0.59 swing, larger than most of the architectural gaps being argued about across papers. The honest summary is that representation latents train faster and stay stable longer, and that the headline quality numbers are too entangled with guidance choices to settle the question.

Where it breaks

A frozen semantic encoder discards high-frequency detail, because discarding it is what made the representation semantic in the first place. The whole burden of pixel fidelity moves to the decoder. This is not a minor engineering wrinkle: it is the reason no instruction-editing method operates in this space, since editing requires leaving untouched regions untouched, and detail the latent never carried cannot be preserved.

The second problem is geometric. These latents are severely anisotropic on a per-token basis, so transporting an isotropic Gaussian to them forces curved trajectories. Straight-path objectives do not produce straight paths in a space shaped like this, which obstructs the distillation that fast sampling depends on.

What the field did next

Two responses, pulling in opposite directions.

The hybrids accept the diagnosis and patch it. Give the semantic latent a detail channel: a residual branch, queries that pull fine structure from shallower encoder layers, or a VAE compressing the foundation features into something smaller. Several report better reconstruction and better generation than pure RAE, which is the outcome the reconstruction-generation dilemma said was hard.

The pixel-space line rejects the premise entirely. JiT argues that pixel-space diffusion was never unstable because of pixels, only because of epsilon-prediction, and that fixing the prediction target makes a plain transformer on raw patches competitive with no tokenizer at all. If that holds, the entire debate about which latent is best was optimizing something optional. The most pointed evidence comes from an unrelated direction: a one-step method reports 1.54 in latent space and 1.61 in pixel space with everything else fixed. A gap that small is weak support for the latent being essential.

The open question

Nobody has run the experiment that would settle it: one architecture, one budget, one guidance method, four latents — pixel, VAE, foundation, hybrid — measured on the same benchmark. Every comparison currently available varies at least two of those simultaneously. Until someone does that, the field’s confident statements about which space is better are, at best, well-motivated conjecture.

How the line developed

Read top to bottom: what came before, the idea itself, the evidence for it, what improved, and where it breaks.

Before

unCLIP / DALL-E 2 · 2022-04landmark

Generates a CLIP image embedding from text, then decodes it — the first system whose primary latent was a semantic space.

Emu2 · 2023-12core

Autoregressively predicts CLIP visual embeddings at multimodal LLM scale, with a diffusion decoder.

The idea

RAE · 2025-10landmark

Frozen DINO/SigLIP/MAE encoder plus trained ViT decoder; identifies the width-vs-token-dimension requirement, a dimension-dependent noise shift, and noise-augmented decoding as what makes it train.

SVG · 2025-10core

Reaches the same conclusion independently within days, using frozen DINOv3 plus a residual detail branch, reaching the same conclusion by a different route.

Evidence

VFM-VAE · 2025-10strong-followup

Finds distilling foundation features into a trained tokenizer weakens them, supporting the keep-it-frozen choice from the tokenizer side.

Improvement

RAEv2 · 2026-05strong-followup

Fixes classifier-free guidance, adds multilayer feature aggregation, and reaches gFID 1.06 in 80 epochs.

At scale

Scale-RAE · 2026-01core

Carries the recipe to open-vocabulary text-to-image with SigLIP-2; VAE baselines overfit by epoch 64 where RAE stays stable to 256.

SVG-T2I · 2025-12strong-followup

The competing scale-up of the same bet, about a month earlier.

Limitation

Distilling Drifting Transformers · 2026-06strong-followup

Shows these latents are severely anisotropic per token, which forces curved trajectories and obstructs distillation.

What it gets right

  • Semantically rich latent by construction, not by auxiliary loss
  • Strong reconstruction and generation simultaneously, which the dilemma said was hard
  • Inherits every future improvement in vision foundation models for free
  • Notably more stable under long fine-tuning than VAE latents

Where it is weak

  • Semantic features discard high-frequency detail, so the decoder carries the whole fidelity burden
  • High-dimensional and anisotropic, which complicates sampling and distillation
  • No published instruction-editing method operates in this space yet

Reported numbers

As published, with the guidance method, budget, and model size that produced them. Seethe comparison page for why these cannot be ranked naively.

PaperBenchmarkValueGuidanceBudgetParams
Scale-RAEDPG-Bench76.9~30K iterations
SVG-T2IDPG-Bench85.782.6B
Scale-RAEGenEval0.495~30K iterations0.5B-9.8B swept
SVG-T2IGenEval0.752.6B
Distilling Drifting TransformersImageNet 256x256 gFID (NFE=1)1.48extrapolation-based16 epochs distillation839M (inherited from RAE)
RAEImageNet 256x256 gFID1.51none800 epochs839M
RAEImageNet 256x256 gFID1.13AutoGuidance800 epochs839M
RAEv2ImageNet 256x256 gFID1.65none80 epochs839M
RAEv2ImageNet 256x256 gFID1.06REPA-Guidance80 epochs839M
SVGImageNet 256x256 gFID3.36none1400 epochs675M
SVGImageNet 256x256 gFID1.92CFG (scale unresolved in source)1400 epochs675M
RAEImageNet 256x256 rFID0.49
RAEImageNet 256x256 rFID0.16
SVGImageNet 256x256 rFID0.65

Competing answers

Open problems it has not solved

Also in this line

dRAE, PixelREPA