Spaces:
Running on Zero
Running on Zero
| title: PixelModel v6 | |
| emoji: 🖼️ | |
| colorFrom: yellow | |
| colorTo: indigo | |
| sdk: gradio | |
| sdk_version: 6.22.0 | |
| app_file: app.py | |
| short_description: 155M-parameter MMDiT text-to-image model | |
| python_version: "3.12" | |
| startup_duration_timeout: 30m | |
| # PixelModel v6 | |
| A 155M-parameter text-to-image diffusion transformer (MMDiT) that generates 256×256 images using rectified flow sampling. Features joint attention over T5 text tokens and CLIP pooled conditioning. | |
| ## Model details | |
| - **Architecture:** MMDiT with joint attention, 2D RoPE, QK-norm, SwiGLU | |
| - **Parameters:** 155M trainable (411M total incl. frozen VAE + text encoders) | |
| - **Text encoders:** T5-base (joint stream) + CLIP ViT-B/32 (global adaLN) | |
| - **VAE:** `madebyollin/sdxl-vae-fp16-fix` | |
| - **Resolution:** 256×256 | |
| - **Sampling:** 50 steps, CFG 3.0 | |
| Based on [bench-labs/PixelModel-v6](https://huggingface.co/bench-labs/PixelModel-v6). |