This is my own personal workflow that I will update over time as I make improvements.

This workflow is MUCH faster than my fp16 Flux workflow. This uses the nunchaku to handle SVDQuant 4-bit quantized models that can fit in 4GB of VRAM. A RTX 3080 10GB generates images in under 10 seconds that are higher quality than Flux.1 Dev fp16 models!

https://github.com/mit-han-lab/ComfyUI-nunchaku

For this you need a Flux model with SVDQuant or int4 in the title. These are around 5GB. Extract the folder to your "\ComfyUI_windows_portable\ComfyUI\models\diffusion_models/" folder.

If needed, you can save more memory by using: https://huggingface.co/mit-han-lab/nunchaku-t5/resolve/main/awq-int4-flux.1-t5xxl.safetensors

Recommended models tested during development:

CreArt Ultimate: https://civitai.com/models/1545303?modelVersionId=1748507

Project0 v40Art: https://civitai.com/models/1018060?modelVersionId=1839533

Jib Mix Flux: https://civitai.com/models/686814?modelVersionId=1595633

7-14-2025

• Changed to use flash-attention2, since it gave better quality.
• Sampler and scheduler tweaks to give better and more consistent results.

6-9-2025

• Improved the face detailer by adding logic to first look at head + hair, then face, and then eyes. It will only upscale based on the largest masked area that is under 25% of the total image (to account for 2x upscale). This way it will only upscale the eyes for extreme close-up images, but will upscale the head and hair for full-body images. This is to avoid banding that happens with Flux on very large images. If you have an eye larger than 25% of the image it will not detail anything to avoid image banding.

6-8-2025

• Release of new workflow