My personal workflow for Flux 1 Dev. I will update it as I make changes.

Model files:

https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/flux1-dev.safetensors

https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors

https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/t5xxl_fp16.safetensors

https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/clip_l.safetensors

Recommended checkpoints:

https://civitai.com/models/1018060/project0

This uses SplitSigmas node, so that it can fit in 8GB or more or vram. I use an RTX 3080 10GB. Image generation is slow, but the images are near zero-shot and text is near perfect every time.

This supports two guiders, the basic one that is faster, and the Perp-Neg Guidance that is slower and allows negative prompts, even with a CFG of 1 (the recommended setting).

Nodes to install that may not appear in the node search:

https://github.com/chflame163/ComfyUI_LayerStyle

https://github.com/ltdrdata/ComfyUI-Impact-Pack

https://github.com/WASasquatch/was-node-suite-comfyui/

https://github.com/pythongosssss/ComfyUI-Custom-Scripts

https://github.com/giriss/comfy-image-saver

https://github.com/Suzie1/ComfyUI_Comfyroll_CustomNodes

https://github.com/SLAPaper/ComfyUI-Image-Selector

https://github.com/ssitu/ComfyUI_UltimateSDUpscale

https://github.com/CosmicLaca/ComfyUI_Primere_Nodes

4-29-2025

• Updated some nodes to work with the newest version of ComfyUI.
• Added links to the required nodes that were not in the newest ComfyUI node search.

1-21-2025

• Some small settings tweaks that I've made over the months.

10-16-2024

• ComfyUI had some updates that significantly increased the speed of the Flux 1 dev. I get more coherent images using a negative prompt with PerpNegGuider, so I only generate images that way now.
• Optimized settings for the best quality during upscaling.
• Added some settings to tweak face brightness and contrast face fix. I also changed from masking the face to the entire head, because it gave much better results.
• I added an option to use CFG 2.0 for upscaling and face correction. I found this to be better when using a negative prompt.
• Added 4x upscale to the end, and a new path for those images.

8-21-2024

• Added a LoRA stack.
• Cleaned up the flow so it's easier to read and change the guidance.

8-20-2024

• Added a face detailer, same as my SD3 node.
• Updated the settings to use ddim/ddim_uniform, and added some notes for the settings that work best in my testing.
• Added FluxGuidance node to the negative prompt conditioning.

8-13-2024

• Removed the PAG and Kohya High Res. Fix, since they don't seem to work. Same issue as SD3. This is because the models switched from convolutional networks (U-Net) to transformer based networks (DiT).
• Perg-Neg does allow negative prompts to work, even though negative prompts do not work at CFG 1. I think it's because it's manipulating the image at the layer level, rather than at the diffusion level.
• Cleaned up the workflow. I moved all the settings, because I found changing the sampler and CFG setting is pointless and does nothing good.
• I did some testing with t5xxl_fp8_e4m3fn.safetensors CLIP text encoder, but I gained no major benefit in speed but a noticeable drop in quality when compared to t5xxl_fp16.safetensors.