Early-Signal Detector

method · first seen 5 Oct 2026

pixel-space diffusion

Also: pixel-space flow models, pixel-space diffusion model, pixel-space video diffusion

Diffusion image generation operating directly on pixels rather than in a latent autoencoder space.

Pixel-space diffusion trains image and video generators directly on raw pixels instead of inside a pretrained VAE latent space. That avoids lossy compression and two-stage training. What is new in these sources is that the approach is being made to work at scale. Iris-3B is a 3B open text-to-image model with weights and an Apple Silicon port. A two-person lab released a 2B Pyramid-JiT model trained for about $21K and posted it on Show HN. Several papers study why v-prediction fails, add adaptive and multi-scale tokenization, or convert pretrained latent video models (PixVid, JVT, U-JiT) into pixel space.

If pixel-space models match latent models in quality and cost, the VAE could drop out of the generation pipeline. That would allow end-to-end training, finer detail for restoration and depth tasks, and per-frame control in video. Several independent groups are publishing open weights and papers, which suggests this is a real research direction and not one vendor's push.

Analyst note (emerging), 17 Nov 2026.

Mentions per week

0419 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 6 on 16 Nov 2026 · 13 mentions in 28 days · 13 independent groups · code → papers → discussion · first seen 6 weeks ago · acceleration +0.88

Earliest sources

Latest sources

Mentioned together with

diffusion transformerflow matchingvelocity predictionimage super-resolutionadaptive tokenizationquadtreeMetal Performance Shadersmixture of expertslow-rank adaptationWiener filter

In the ledger