Source: https://discord.com/channels/851231576494702613/1268987858602168474

Img2Img workflow for Black Forest Labs' new model, Flux. FLUX.1 [pro]: Top-tier performance (API only) FLUX.1 [dev]: Open-weight, efficient, non-commercial use FLUX.1 [schnell]: Fastest for local development and personal use  Download the Models: CLIP: https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main VAE: https://huggingface.co/black-forest-labs/FLUX.1-schnell/blob/main/ae.sft Dev Model Weights: https://huggingface.co/black-forest-labs/FLUX.1-dev Schnell Model Weights: https://huggingface.co/black-forest-labs/FLUX.1-schnell  How to Use in ComfyUI: https://comfyanonymous.github.io/ComfyUI_examples/flux/ Kijai 8fp version: https://huggingface.co/Kijai/flux-fp8/tree/main

• This workflow is set up for the [dev] variant of the FLUX.1 model.
• This workflow is HIGH VRAM only, I've only tested it with my 24gb card. Chances are this won't work with less, but you MIGHT be able to do it with 16gb, let me know!
• We look at an input image, determine a prompt for it, then mix the image in with the prompt and control the level of noise injected into the result to control the 'wildness' of the dream.
• You can convert this to prompt only by plugging the Empty Latent Image into the Ksampler, instead of the VAE Encode.
• Same workflow works with Schnell, set your steps to 4. Schnell decreases inference time with the penalty of a less great image, however it uses the same amount of VRAM.