| --- |
| tags: |
| - text-to-image |
| - comfyui |
| - prxpixel |
| - fp8 |
| - safetensors |
| - custom-inference |
| pipeline_tag: text-to-image |
| base_model: |
| - Photoroom/prxpixel-t2i |
| --- |
| |
| # PRX Pixel FP8 Mixed Transformer |
|
|
| This repository contains a **single-file PRX Pixel transformer checkpoint** converted from `Photoroom/prxpixel-t2i` into a **mixed fp8 safetensors** format for the `lumina_prx_pixel` custom ComfyUI runtime. |
|
|
| ## What This File Is |
|
|
| - This is **not** the full original model repo. |
| - This file contains the **transformer only**. |
| - The **tokenizer** and **text encoder / clip side** should still be loaded from the original base model source: |
| `Photoroom/prxpixel-t2i` |
|
|
| ## Mixed FP8 Means |
|
|
| This checkpoint uses a **mixed fp8 storage recipe**, not full end-to-end fp8 for every tensor. |
|
|
| - Large transformer projection weights are stored in `float8_e4m3fn` |
| - More sensitive tensors stay in higher precision |
| - Runtime is **mixed precision** |
| fp8 weights are used for selected large linear layers, while activations / outputs remain in bf16 or the loader compute dtype |
|
|
| In other words: |
|
|
| - **weight storage** is reduced for selected heavy layers |
| - **compute** is still mixed, not pure fp8 everywhere |
|
|
| ## Intended Runtime |
|
|
| This file is intended for: |
|
|
| - `lumina_prx_pixel` |
| - `Load PRX Pixel model` |
| - `Load prx clip model only` |
|
|
| from the matching custom ComfyUI node pack. |
|
|
| It is **not meant for stock diffusers loading**. |
|
|
| ## How To Use In ComfyUI |
|
|
| 1. Place the `.safetensors` file inside `ComfyUI/models/diffusion_models` or `ComfyUI/models/unet` |
| 2. Load it with `Load PRX Pixel model` |
| 3. Load the text side with `Load prx clip model only` |
| 4. Set that clip source to `Photoroom/prxpixel-t2i` |
| 5. Connect both into `lumina_prx_pixel` |
|
|
| ## Runtime Notes |
|
|
| - For the real fp8 matmul path, you want a CUDA setup with fp8 support in PyTorch |
| - On unsupported backends, the runtime may fall back to dequantized math for execution |
| - This is still useful for packaging and selective storage reduction, but the best VRAM win is on supported CUDA fp8 paths |
|
|
| ## Conversion Notes |
|
|
| This file was produced from the original PRX Pixel transformer shards and packed into a custom single-file format used by the local runtime. |
|
|
| The embedded metadata includes: |
|
|
| - transformer config |
| - scheduler config |
| - PRX Pixel generation settings |
| - fp8 scale metadata for converted layers |
|
|
| ## Limitations |
|
|
| - This is a converted checkpoint, not an original upstream release artifact |
| - Output quality and numerical behavior can differ slightly from the source bf16 weights |
| - Compatibility is tied to the custom `lumina_prx_pixel` runtime that knows how to read this format |
|
|
| ## Credit |
|
|
| - Base model: `Photoroom/prxpixel-t2i` |
| - FP8 mixed single-file packaging and runtime support: `Lumnatrix` |
|
|
