Flux.1 dev Hires Fix Workflow is a high-definition repair workflow specifically designed for the Flux.1 dev model. I recommend using the Flux.1 fp8 version of the model. This model reverse-engineers prompt words through unsloth–Meta-Llama-3.1-8B-Instruct and enhances details using ultimator, so you don’t need to input any prompt words to effectively run this workflow.

1. Parameters

1. Denoising Level: The denoising amplitude should be around 0.4 to 0.6. If you are changing the style, a higher denoising amplitude is required.

2. ControlNet Upscaler Weight: Set it between 0.8 to 1.0.

3. Magnification Factor: Do not exceed a magnification of 2x, as it may introduce many artifacts (such as grid patterns). If you need to go beyond 2x, perform the HD fix in two steps—use the result from the first HD fix as the input for the second.

2. Models

You will need the following three models:

1. LLM Model: This is the large language model set. It will automatically download on the first run. If the download fails, you can manually download it and place it in the ComfyUI/models/LLM/unsloth--Meta-Llama-3.1-8B-Instruct folder.

Download Link:https://huggingface.co/unsloth/Meta-Llama-3.1-8B-Instruct

2. Clip Vision: This is the clip model for vision tasks. Download all files and place them in the ComfyUI/models/clip_vision/google--siglip-so400m-patch14-384 folder.

Download Link: https://huggingface.co/google/siglip-so400m-patch14-384/tree/main

3. Joy Caption Lora: This is the lora model for captioning. Download all files and place them in the ComfyUI/models/Joy_caption/cgrkzexw-599808 folder.

Download Link: https://huggingface.co/spaces/fancyfeast/joy-caption-alpha-two

4. ControlNet HD Repair Model: Download this model and rename it to Flux.1-dev-Controlnet-Upscaler. Place it in the models/controlnet/flux folder.

Download Link: https://huggingface.co/jasperai/Flux.1-dev-Controlnet-Upscaler/resolve/main/diffusion_pytorch_model.safetensors?download=true