Photoroom Ships Prxpixel-t2i, Raw Pixel Diffusion Model

Photoroom has released Prxpixel-t2i, a 7-billion-parameter text-to-image model that generates images directly in pixel space. Unlike most diffusion models, it does not use a variational autoencoder (VAE) to compress pixels into a latent space. The model combines a large transformer backbone with a Qwen3-VL text encoder and a flow matching scheduler to produce 1024×1024 outputs.

The model was built by Photoroom’s machine learning team as a pixel-space variant of their PRX architecture. By operating on raw RGB values, the design simplifies the generation pipeline and removes the need for a separate VAE decoder. The code and weights are publicly available under the Apache 2.0 license, encouraging open research and commercial use.

Pixel-space generation without a VAE

Researchers exploring pixel-level image synthesis will find this model useful for studying how diffusion works without latent compression. Developers who need direct access to pixel gradients can also benefit from the simplified architecture. The model opens up possibilities for applications where VAE reconstruction artifacts are a concern.

Installation and current limitations

The model is not yet part of the stable diffusers library, so users must install diffusers from the prx-pixel-pipeline branch on GitHub. Photoroom also requires the transformers package version 4.57 or newer to load the Qwen3-VL text encoder. A future merge into the main diffusers release is anticipated but no date has been given.

“PRXPixel is a pixel-space variant of PRX: it denoises raw RGB directly (no VAE), conditions on a Qwen3-VL text encoder (rather than T5Gemma), and feeds the generation resolution into the timestep modulation.” — Source: Hugging Face