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

Iris-3B. Going Beyond the Latent with Pixel-Space Diffusion Training, Conversion and Fine-Tuning. By Hanqiu Li Cai and one other, from SperidLabs. Published on the seventh of October.

Pixel-space diffusion models skip the lossy VAE that latent models depend on. That suggests an advantage on downstream tasks where fine-grained detail matters, since nothing has been thrown away before the model sees it. The paper tests that claim directly, and its finding is negative. Using a pixel-space generative prior brought no significant improvement.

The test runs along both routes to a pixel-space backbone. The first is training from scratch. Iris-3B is a three billion parameter pixel-space text-to-image transformer, pretrained through a resolution curriculum. It starts at two hundred and fifty-six, moves to five hundred and twelve, and ends at one thousand and twenty-four. Before scaling, the prediction target and representation alignment were ablated at two hundred and fifty-six squared, to decide what to scale. The second route is conversion: a pretrained latent model, FLUX two Klein base four B, was converted to pixel space.

Both families were then fine-tuned for monocular depth estimation, and for image restoration and super-resolution. For depth, using one matched direct-regression recipe, Iris-3B comes out level with the latent FLUX.2 Klein, and the converted pixel FLUX.2 Klein falls behind it. On four times DIV2K restoration, neither pixel model beats a latent FLUX.2 Klein fine-tune, with the converted one trailing it slightly.

The claim has edges the authors draw themselves. It covers these two downstream tasks, these two routes to pixel space, and a fine-tuning recipe matched across models. The paper documents the recipes, the failure modes, and the confounds that remain behind the negative result, so it is offered as a careful null rather than a proof that pixel space cannot help.

Nor is it a verdict against pixel-space generation itself. Iris-3B shows that pixel-space pretraining with the pixel-transformer head of PixelDiT scales to three billion parameters, and reaches text-to-image quality competitive with latent models, matching Qwen-Image on OneIG under the official evaluators at one thousand and twenty-four squared.

So two results sit side by side. Pixel space can generate as well as latent space at this scale, yet the detail it preserves did not translate into better depth or restoration once fine-tuned. The weights and training code are released, in the hope that they help pave the way for further work on pixel-space generation.