Iris-3B generates megapixel images without a VAE

Most modern image generators compress pictures into a small latent grid, run diffusion or flow matching on that grid, and use a variational autoencoder to reconstruct the pixels. Speridlabs’ Iris-3B removes that compression stage and generates raw RGB pixels through patch tokens at resolutions up to roughly one megapixel.

Released under Apache 2.0, Iris-3B includes weights, code, a demo, and documentation through its project page and model files. The 3-billion-parameter backbone supports text-to-image generation, monocular depth estimation, and 4x image restoration. Its most useful research result is a negative one: under the reported training budget, the pixel-space model did not outperform a latent baseline on the downstream vision tasks.

Why pixels change the compute bill

A VAE reduces an image to a compact latent representation before generation begins. That compression lowers the number of values the generative model must process, while the decoder handles the final conversion to RGB. Compression can also discard fine spatial information or introduce artifacts in textures, edges, and skin.

Iris tests whether direct RGB prediction can produce a richer visual prior for detail-sensitive tasks such as depth estimation and restoration. The model still groups pixels into 16 by 16 patches, so it does not process every pixel as an independent transformer token. At 1024 by 1024 resolution, the image becomes a 64 by 64 grid containing 4,096 patch tokens.

Pixel-space training carries a substantial compute cost because those tokens must preserve enough information to reconstruct every RGB value. Latent models operate on smaller spatial grids, which has made them the practical default for high-resolution generation. Iris-3B tests how far a direct-pixel design can scale with a 3B-parameter transformer.

Inside the patch pipeline