PRX Pixel: Photoroom's Open Source Image Model Skips VAE Compression Entirely

Photoroom released PRX Pixel on July 6, 2026, a 7 billion parameter text to image model available under Apache 2.0. The release includes full model weights on HuggingFace, training code on GitHub, and a live demo on HuggingFace Spaces. Every component is available for commercial use without restriction.

The model follows from the original PRX architecture Photoroom published in November 2025. PRX Pixel scales the parameter count from 1.3 billion to 7 billion and introduces a fundamental architectural change: it generates images directly in pixel space, with no Variational Autoencoder involved at any stage.

Why Pixel Space Matters

The dominant architecture in text to image generation today, used by Stable Diffusion, FLUX, and most major commercial models, is latent diffusion. An image is compressed by a VAE into a smaller representation, the diffusion process runs in that compressed space, and the result is decoded back to pixels at the end.

That round trip introduces artifacts. VAE encoding and decoding shifts colors subtly, particularly in skin tones and solid color surfaces. Textures can drift across the compression boundary. Fine detail, most visibly in text embedded in images including signage, labels, and typography in generated scenes, is frequently lost or distorted during reconstruction. These are acceptable tradeoffs when the goal is fast generation at scale.

They become a problem when the output needs to meet a production specification: a specific Pantone code for a costume, a readable product label on a prop, an exact color match for a set design reference. PRX Pixel eliminates the compression step. The model denoises directly from noise to pixel values without ever encoding the image into a compressed representation. The output is what the model computes, without a reconstruction pass that can alter it.

Architecture

The model uses FlowMatchEulerDiscreteScheduler and a Qwen3-VL vision language model as its text encoder. The Qwen3-VL encoder is multilingual, which means PRX Pixel accepts prompts in languages other than English without requiring translation preprocessing.

Photoroom trained the model on 32 H200 GPUs in under 10 days. The training incorporates three techniques not standard in published diffusion model recipes:

• TREAD (Token Routing Efficient Attention Diffusion): a routing mechanism that selects which token pairs participate in attention at each step, reducing compute cost in pixel space where sequence lengths are significantly longer than in latent space
• REPA alignment with DINOv3: a representation alignment approach that trains the model's internal representations to match a strong visual feature extractor, improving structural coherence in generated images
• LPIPS and DINOv2 perceptual losses: training losses that evaluate image quality in feature space rather than pixel space, pushing the model toward outputs that score well on learned perceptual metrics beyond accuracy at the pixel level

Try the Live Demo

The model is available as a Gradio interface on HuggingFace Spaces. It accepts a text prompt, negative prompt, step count, guidance scale, scheduler shift, resolution, and seed as inputs. The Space loads without requiring a login.

Production Applications

For filmmakers and production designers, the absence of VAE compression has three practical implications.

Color accuracy is the most direct. Concept art used as a production reference often needs to communicate specific color information to a team: a costume color that matches a character's established palette, a set paint color coordinated with a cinematographer's lighting plan, or a prop color that photographs accurately under specific conditions. VAE reconstruction introduces small but measurable color drift. Pixel space generation preserves the color the model computes without an additional translation step.

Fine detail retention matters for any generated image that contains text or structured visual elements. Prop design references, product labels on set, signage in concept art, or typography in visual development work are all candidates where conventional latent diffusion models distort or lose the detail entirely. PRX Pixel's direct output path preserves them.

Multilingual prompt support via the Qwen3-VL encoder makes the model accessible to production teams working in languages other than English without requiring a separate translation step. This is not a common feature in open source image models, most of which use CLIP or T5 encoders trained primarily on English text.

For image generation in the cloud without local model setup, AI FILMS Studio's image workspace provides access to the latest models in the browser.

Open Source and License

PRX Pixel is available under Apache 2.0, which permits commercial use without restriction, including use in commercial productions, client work, and products built on top of the model. The Apache 2.0 license requires attribution in derivative works but places no limitation on revenue or production scale.

The full training code is published in the Photoroom/PRX repository on GitHub, including the architecture implementation, training loop, evaluation scripts, and ablation experiments. This level of release is less common than releases that provide only model weights, and makes PRX Pixel useful as a research baseline for teams developing custom pixel space generation systems.

For comparison with other open source pixel space approaches released in 2026, HiDream-O1-Image is an 8B MIT-licensed model that takes a similar no-VAE approach using the Pixel-level Unified Transformer architecture and adds instruction-based image editing and subject personalization alongside text to image generation.

---

Sources

Project Page: photoroom.com/blog/prx-pixel GitHub: Photoroom/PRX HuggingFace: Photoroom/prxpixel-t2i HuggingFace Space: Photoroom/PRX-Pixel HuggingFace Blog: PRX Open Source T2I Model License: Apache 2.0