This workflow allows you to harness the power of FLUX.1 [dev] within ComfyUI. Follow these steps to set up and run the model.
The true "simplest" version of a Flux workflow can be found here, but it uses the FP8 version which is slightly worse. For that reason, I do not use the "Load Checkpoint" node version of the workflow.
Note: To train LoRAs, see my video guide here: https://civarchive.com/articles/6533/how-to-train-a-lora-for-flux1-dev-with-simpletuner
Instructions:
Model Placement:
Place the FLUX.1 [dev] checkpoint in
ComfyUI/models/unet(not in checkpoints folder)
Download Required Files:
Original weights: flux1-dev.sft
FP8 version (for <24GB VRAM systems): flux1-dev-fp8.safetensors
Text encoders (place in
ComfyUI/models/clip): flux_text_encodersVAE (place
ae.sftinComfyUI/models/vae): ae.sftFor lower RAM usage, download the FP8 T5XXL encoder (with some quality degradation)
Low VRAM Setup:
Launch ComfyUI with the "--lowvram" argument (add to your .bat file) to offload the text encoder to CPU
Performance Notes:
Confirmed to run on:
RTX 3090 (24GB): 1.29s/it
RTX 4070 (12GB): 85s/it
Both tests used the FP8 quantized version. Note that the 4070 is significantly slower.
Preview Setup:
Update your ComfyUI to the latest version
In the Manager menu, set "Preview Method" to "Auto"
Using the Workflow
Load the provided workflow file into ComfyUI
Adjust your prompts and parameters as desired
Run the workflow to generate images
Note: This workflow includes a custom node for metadata. It can run in vanilla ComfyUI, but you may need to adjust the workflow if you don't have this custom node installed.
Introduction to FLUX.1 [dev]
FLUX.1 [dev] is a groundbreaking 12 billion parameter rectified flow transformer for text-to-image generation, developed by Black Forest Labs. As part of the innovative FLUX.1 suite of models, it represents a significant leap forward in image synthesis technology, offering:
State-of-the-art image quality and prompt adherence
Competitive performance with closed-source alternatives
Enhanced efficiency through guidance distillation training
Support for diverse aspect ratios and resolutions (0.1 to 2.0 megapixels)
Black Forest Labs, founded by distinguished AI researchers and engineers, brings a wealth of experience from developing foundational generative AI models like VQGAN, Latent Diffusion, and Stable Diffusion. FLUX.1 [dev] showcases their commitment to pushing the boundaries of AI-generated imagery.
License and Usage
FLUX.1 [dev] is released under a non-commercial license. Please refer to the official FLUX.1 [dev] license for full terms and conditions. This model is intended for research, personal projects, and non-commercial use only.
Enjoy exploring the capabilities of FLUX.1 [dev] with this ComfyUI workflow!
Description
This version has all of the advanced parameters as the previous versions, but I pestered Comfy into adding LoRA support natively.
FAQ
Comments (11)
3090 here, should i use fp16?
you will find this out in your own testing. short of the long, a 3090 can run both
Nice setup. Now waiting for the first nice lora's
So this is not ready to be used yet? Because of lacking Lora?
@dillion1920 It works really nice, but will improve with Loras as always.
This means that the flux1 model has its own lora, it cannot use lora from SD1.5? 😅😅
Thanks for the workflow!
Can you please also post the link to the Lora you used? Or can it be any Lora? (SD1.5/XL)
Sadly you won't be able to use 1.5/XL LoRAs with Flux but you can train your own if you'd like: https://civitai.com/articles/6533/how-to-train-a-lora-for-flux1-dev-with-simpletuner
The LoRA I used in the workflow was just a test to make the cat look more like my cat, but it didn't fully work as expected. It's on my HuggingFace though if you wanted to try it just for fun- https://huggingface.co/markury/AlbertFLUX
I admit, I'm a bit confused. I thought no LORAs exist for FLUX yet?
No worries, you can find a guide on training them here: https://civitai.com/articles/6533/how-to-train-a-lora-for-flux1-dev-with-simpletuner
I will also admit- my LoRAs suck. Hopefully the community will make some better ones. But yes, they do exist!
They do. As of today one here at Civitai and 4 on Huggingface, growing...