The bet
This line agrees with single-stage pixel transformers that no learned tokenizer is necessary, and disagrees about why pixel generation was hard. The diagnosis here is not the prediction target but the resolution: model a small image well, then refine upward, and pixel diffusion comes within a fraction of a FID point of latent diffusion with no VAE anywhere in the pipeline.
The arc
Cascaded Diffusion Models sets the pattern in 2021 — a chain of pixel-space models, each super-resolving the last, with conditioning augmentation to stop errors compounding along the chain. simple diffusion then asks whether the chain is avoidable: with a shifted noise schedule and capacity concentrated at low resolution, can one end-to-end model be comparable to a cascade and to latent diffusion? It argues comparability. It does not claim cascades are unnecessary, and this atlas previously overstated it as doing so.
HDiT supplies the architecture that makes single-stage pixel training practical, a hierarchical transformer whose cost scales as O(n) rather than the O(n²) of a standard DiT, by restricting global attention to the coarsest level and using neighbourhood attention elsewhere. It trains natively at 1024×1024 on FFHQ; its ImageNet results are at 256×256 and are competitive rather than state of the art.
SiD2 is the result that should change how the current wave is read, though it is usually reported less carefully than it deserves. In October 2024 it reached FID 1.48 on ImageNet 512 in pixel space, at 512 steps with guidance scale 1.0 over the interval (-3, 5). It is often cited as matching latent diffusion. It does not: in the paper’s own matched-setting table EDM2-XXL with a guidance interval reaches 1.40, and the authors concede that latent diffusion still has slightly better scaling properties. What is true, and sufficient, is that pixel-space generation was within 0.08 FID of the best latent model a year before the tokenizer-free papers argued the case — and that no pixel-space paper in this atlas has beaten 1.48 at 512 since.
Edify Image shipped industrial-scale pixel-space generation the following month using Laplacian frequency bands, though it reports no quantitative benchmark, so it cannot be placed against these numbers. PixelFlow carried the multiscale bet across to flow matching in 2025.
Why this line is marked superseded
Not because anything in it was refuted. Nothing here was. It is superseded in attention: the field moved to single-stage transformers, and the cascaded literature stopped accumulating. The honest reading is that a cascade is more moving parts than the tokenizer it was meant to replace, that errors compound between stages and need explicit machinery to control, and that most of this work predates or sits outside the transformer scaling story the field now expects a generative method to have.
Why it still matters
It is the strongest available evidence that the 2025–26 tokenizer-free wave rediscovered something rather than discovered it. The argument that pixels were never the real obstacle, and that a well-designed pixel model comes within a fraction of a FID point of the best latent model, was made here with numbers years earlier. Note the careful wording: this line got close and said so honestly. It is the current wave, not this one, that reaches for the word parity.
That is worth stating plainly, because it is the kind of thing an atlas organised around papers rather than ideas would miss entirely. These papers do not cite the current wave and the current wave does not always cite them, so nothing in the citation graph connects the two. The connection is in the claim, not the references.