ariaotp/INT4-ConvRot-W4A4-ComfyUI

These models are quantized INT4 Convrot W4A4 for latest ComfyUI. It has somewhat inferior image quality than BF16/GGUF Q8/INT8 Convrot but fast in Ampere(RTX30x0)/Ada Lovelace(RTX40x0)

Original models that I used:

• Krea2 Turbo: https://huggingface.co/LAXMAYDAY/Krea-2-Turbo-int4-tensorwise-mixed
• Krea2 Raw: https://huggingface.co/LAXMAYDAY/Krea-2-Raw-int4-tensorwise-mixed
• Qwen-Image-2512: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI
• Qwen-Image-Edit-2511: https://huggingface.co/LAXMAYDAY/Qwen-Image-Edit-2511-int4-tensorwise-mixed
• FireRed-Image-Edit 1.1: https://huggingface.co/cocorang/FireRed-Image-Edit-1.1-FP8_And_BF16
• FLUX.2 Klein 9B: https://huggingface.co/black-forest-labs/FLUX.2-klein-9B
• Qwen2.5-VL 7B: (Qwen2.5-VL-7B-Instruct-abliterated): https://huggingface.co/huihui-ai/Qwen2.5-VL-7B-Instruct-abliterated (Needed for Qwen-Image and FireRed-Image-Edit)
• Qwen3 4B (Huihui-Qwen3-4B-abliterated-v2): https://huggingface.co/chinmankokumin/Qwen3-4B-abliterated-v2 (Needed for Z-Image and FLUX.2 Klein 4B)
• Qwen3 8B (Huihui-Qwen3-8B-abliterated-v2): https://huggingface.co/huihui-ai/Huihui-Qwen3-8B-abliterated-v2 (Needed for FLUX.2 Klein 9B)
• Qwen3-VL 4B (Huihui-Qwen3-VL-4B-Instruct-abliterated): https://huggingface.co/ahmed22xa/Huihui-Qwen3-VL-4B-Instruct-abliterated-comfy (Needed for Krea2)
• Qwen3-VL 8B (Huihui-Qwen3-VL-8B-Instruct-abliterated): https://huggingface.co/huihui-ai/Huihui-Qwen3-VL-8B-Instruct-abliterated (Needed for Boogu-Image and Ideogram4)

For Krea2 models and Qwen-Image-Edit 2511, Since they are already quantized INT4 ConvRot W4A4 by LAXMAYDAY, I fixed keys of those models to convrot_w4a4 for working on official ComfyUI.

That python script I created used ChatGPT 5.6 Sol is convert_int4_tensorwise_to_convrot_w4a4.py.

The rest of them are used Starnodes Model Converter, made by Starnodes2024, quantized from BF16 models https://github.com/Starnodes2024/comfyui-starnodes-modelconverter

convert_int4_tensorwise_to_convrot_w4a4.py Script usage (not so useful because I already converted and uploaded)

Not actual convert, but test run

• python convert_int4_tensorwise_to_convrot_w4a4.py model.safetensors --dry-run

Save to new model(convrot_w4a4 model for ComfyUI)

• python convert_int4_tensorwise_to_convrot_w4a4.py model.safetensors model_convrot_w4a4.safetensors *default output name: 'model.convrot_w4a4.safetensors'

Overwrite original model

• python convert_int4_tensorwise_to_convrot_w4a4.py model.safetensors --in-place

Overwrite existing file

• python convert_int4_tensorwise_to_convrot_w4a4.py model.safetensors model_convrot_w4a4.safetensors --force

Keep original scale scape [N,1]

• python convert_int4_tensorwise_to_convrot_w4a4.py model.safetensors --keep-scale-shape

If that model doesn't have "convrot: true", but it's convrot_w4a4 model:

• python convert_int4_tensorwise_to_convrot_w4a4.py model.safetensors --assume-convrot

*WARNING: If you convert non-ConvRot int4_tensorwise model, it should be MALFUCTIONS!

Can convert these models for official ComfyUI:

• https://huggingface.co/LAXMAYDAY/Krea-2-Turbo-int4-tensorwise-mixed
• https://huggingface.co/LAXMAYDAY/Krea-2-Raw-int4-tensorwise-mixed
• https://huggingface.co/LAXMAYDAY/Qwen-Image-Edit-2511-int4-tensorwise-mixed

License

Qwen-Image, Qwen-Image-Edit, FireRed-Image-Edit, Qwen2.5-VL, Qwen3, Qwen3-VL: Apache 2.0

• https://www.apache.org/licenses/LICENSE-2.0.txt

Krea2: Krea 2 Community License

• https://www.krea.ai/krea-2-licensing

FLUX.2 Klein 9B: FLUX Non-Commercial License

• https://github.com/black-forest-labs/flux2/blob/main/model_licenses/LICENSE-FLUX-NON-COMMERICAL

convert_int4_tensorwise_to_convrot_w4a4.py: MIT

• https://opensource.org/license/mit