AI

Iris-3B: Going Beyond the Latent with Pixel-Space Diffusion Training, Conversion and Fine-Tuning

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 18:18:24 UTC.

Briefings

Iris-3B, a 3B-parameter pixel-space transformer, was pretrained from scratch via a 256-to-512-to-1024 curriculum. It competes with latent models on text-to-image quality at 1024^2 but failed to improve monocular depth estimation or image restoration/super-resolution compared to latent FLUX.2 Klein. The study documents these negative results and the specific failure modes.

Key points

• Iris-3B is a 3B-parameter pixel-space transformer trained via a 256-to-512-to-1024 curriculum.
• Iris-3B matches latent FLUX.2 Klein on text-to-image quality at 1024^2 but lags on depth estimation and restoration.

Context

This research investigates pixel-space diffusion models to determine if avoiding VAEs offers advantages for downstream tasks like depth estimation and super-resolution.

Limitations

The study explicitly documents negative results and failure modes for depth estimation and restoration. No significant improvements were found over latent models on these specific tasks, and the study does not claim superiority on other metrics.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 11/11/2026, 09:40:03 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiaxxarlrwc2tkcbes6s65e2u52uxoh46kmk7qya4ujol7qlky6hcu