The previously released Stable Diffusion 3 model has been criticized for being over-audited, resulting in a "regression" in the ability to generate human structures.
Now an excellent pinto has appeared, the Black Forest Laboratory'sFLUXFLUX was released to rave reviews, and was named the "strongest open source literate graph model".
Compared to header closed-source tools such as Midjourney and DALLE, FLUX's performance is still respectable.
FLUX currently supportsComfyUIGraphical User Interface, described nextLocal deploymentThe methodology:
Step 1: Install the latest version of ComfyUI
1. Go to the following URL: https://github.com/comfyanonymous/ComfyUI?tab=readme-ov-file#installing
2. Click Direct link to download to download a zip file of about 1.5G.
3. Extract the zip package to your designated hard disk.
4. Enter the installed directory and run run_nvidia_gpu.bat, then the system will automatically download the necessary files.
5, after which will automatically pop up a web page similar to this, indicating that the installation was successful.
Note: This background window should be kept open, otherwise ComfyUI will not work properly.
Step 2: Download the FLUX model
1. Go to FLUX's hug face page: https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main and click the Files button:
2. In the file list, download the master model file and the VAE self-encoder:
flux1-dev.sft: 23.8 GB, put this file in the ComfyUI/models/unet/ directory
ae.safetensors: 335 MB, put this file in the ComfyUI/models/vae/ directory
3. Enter the ComfyUI Hugging Face page: https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main
4. In the file list, download the CLIP pre-training model:
t5xxl_fp16.safetensors: 9.79 GB
t5xxl_fp8_e4m3fn.safetensors: 4.89 GB
clip_l.safetensors: 246 MB
All of the above files are placed in the ComfyUI/models/clip/ directory, with at least one of the 9GB and 4GB files downloaded.
5. Go to the following URL: https://openart.ai/workflows/maitruclam/comfyui-workflow-for-flux-simple/iuRdGnfzmTbOOzONIiVV and click the Download button on the right to download the workflow.
6. Drag the workflow file onto the ComfyUI page to get the result shown below:
7. Make sure you have selected the correct model and encoder (refer to the picture below) and then you can start painting!
It should be noted that FLUX has high hardware requirements. During the test, it took up to nearly 30G of RAM:
FLUX is also slow in generating images, even with a 4090 graphics card with default settings, it takes more than 20 seconds to generate a 1024x1024 image:
But considering the increased odds of a one-time success, slow seems acceptable?
The article covers the URL:
FLUX's Github page:
https://github.com/black-forest-labs/flux
ComfyUI's Github page:
https://github.com/comfyanonymous/ComfyUI
