sherwin217/krea-2-turbo-int8-convrot

Krea 2 Turbo · INT8 ConvRot

Text-to-image Gradio demo using the exact conservative checkpoint AX1Y2JP/Krea-2-Turbo-INT8-ConvRot/krea-2-turbo-int8-convrot.safetensors.

The model author warns against loading this checkpoint with ComfyUI's native INT8 loader. This app explicitly uses BobJohnson24/ComfyUI-INT8-Fast, with ConvRot enabled and on-the-fly quantization disabled. It never substitutes the base BF16 model or another quantization.

Features: three full-size aspect ratios, a 512px preview, 4–12 sampling steps, repeatable seeds, downloadable PNGs with settings metadata, examples, and an API/MCP endpoint. Default sampling: Euler, beta schedule, 8 steps, CFG 1, AuraFlow shift 4, following the existing ComfyUI-based Krea 2 demo recipe.

Deployment

Create a Gradio ZeroGPU Space (zero-a10g) and upload the files in this folder. Hardware must be selected in Space settings or the Hub API; README metadata alone cannot activate ZeroGPU. No paid dedicated GPU is requested.

Pinned ComfyUI and INT8-Fast sources are downloaded on startup. Model assets are downloaded from public repositories at pinned revisions. All three models are loaded at module scope before generation. First startup downloads roughly 20 GB of model files and can take several minutes.

License and attribution

Model weights are covered by the Krea 2 Community License. Use is subject to that license. Model credits: Krea and AX1Y2JP. Inference uses ComfyUI and INT8-Fast, whose source licenses remain applicable. Pipeline parameters were checked against misukisu's Krea 2 demo.

Validation status

Python syntax checked locally. Live ZeroGPU startup and image generation must be verified after repository write authentication is available.