Revised and updated 10/08/2024
(save this article if you want to receive updates)
We are at a new beginning, a paradigm shift for many of us. After the failure of the SD3 launch due to the problems of that model and the commercial licensing issue, it seemed that our path was going to run for a good season between SDXL, PONY and 1.5 models until the arrival of something better.
But the arrival of this new model created by Black Forest Labs has shown that the journey in the desert may not be as long as we thought.
After doing some first tests, with the two new models available, (FLUX 1 schnell), which gives optimal results in only 4 steps, and (FLUX 1 dev), optimized to work with 20 steps I can say that these models give a very high performance, although it is true that the technical requirements of your PC must be quite high to make it work (Optimal 24 gb of Vram, but works reasonably well with 12 GB of Vram).
Things to keep in mind:
-The resolution and detail of the images are above almost anything known so far.
-The texts are almost always perfect.
-Hands no longer have fused, crooked or too many fingers.
-Negative prompts are no longer necessary, so there is not even a node for them.
-Perfect compatibility with Comfyui, (here is a sample workflow)
Installation:
Download Model - flux1-dev.sft (Standard) or flux1-schnell.sft (Need less steps). put it into \models\unet // I used dev version
Download Vae - ae.sft that goes into \models\vae
Download clip_l.safetensors and one of T5 Encoders: t5xxl_fp16.safetensors or t5xxl_fp8_e4m3fn.safetensors. Both are going into \models\clip // in my case it is fp8 version
Add --lowvram as additional argument in "run_nvidia_gpu.bat" file
Update ComfyUI and use workflow according to model version, be patient ;)
Model + vae: black-forest-labs (Black Forest Labs) (huggingface.co)
Text Encoders: comfyanonymous/flux_text_encoders at main (huggingface.co)
Flux.1 workflow: Flux Examples | ComfyUI_examples (comfyanonymous.github.io)
UPDATE 03/08/2024:
-A Flux1-FP8 model is now also available, with half the weight of the original model, at the following link https://huggingface.co/Kijai/flux-fp8/tree/main
-Img2Img workflow added in this article.
UPDATE 10/08/2024:
-Added Workflow for the first LORA FLUX that are coming out. (Thanks to https://civitai.com/user/markury )
You can see his original post here: https://civitai.com/models/618997?modelVersionId=702900
-First LORAS Available:
https://huggingface.co/XLabs-AI/flux-RealismLora/tree/main (the first one)
https://civitai.com/models/633553/aesthetic-lora-for-flux (for anime)
Personally, they have not given me the expected results, but try them for yourselves.
Want More Info About FLUX? You can see my other Articles Here:
https://civitai.com/articles/6558/updated-flux-lora-flux-controlnet-here-we-goooo
https://civitai.com/articles/6527/flux-likes-soccer-and-much-much-more-what-does-the-future-hold