Introduction
A new revolutionary AI image generation Model has been released, which is as good if not better than Midjourney! The new model, dubbed Flux, has been created by Black Forest, a company formed by former members of Stability AI.
Black Forest, have released three versions of their model:
- Dev Model: Non-commercial but highly impressive. Users can request a license for commercial use.
- Schnell Model: A lightning-based model ready for commercial projects with clear terms of use.
- Closed Source Model: Available via API, similar to Stable Diffusion 3, but reserved for their private use.
Both the Dev and Schnell models are significant improvements over Stable Diffusion 3, particularly in handling complex visual elements like hands. However the models do have NSFW filters, they are nowhere near as aggressive as SD3’s
Installation Guide for Flux 1.0 on ComfyUI
To get started with Flux 1.0 on ComfyUI, follow these steps:
- Download the Models: Head to Black Forest Labs’ Hugging Face page and download either the Flux 1.0 Schnell or Flux 1.0 Dev model, along with the AE SFT files. You also need to get the T5XXL CLIP encoder (download it here). Choose either the FP16(16-24+ VRAM) or FP8 (12GB+ VRAM) version based on your system’s RAM and VRAM capacity.
- Organize Your Files: Place the model files in the appropriate ComfyUI directories— the .sft models go in the models/UNet folder, AE files in the models/VAE folder, and the T5XXL models/CLIP in the clip folder
- Set Up the Workflow: Drag and drop the workflows (included here) to start using Flux right away!
Breaking Down the Workflow
Flux’s workflow differs slightly from traditional SDXL workflows but is straightforward once broken down:
DualCLIPLoader: Loads the T5XXL and secondary CLIP_L encoders.
Sampler Custom Advanced: Handles model, conditioning, latent image, and generation parameters.
Noise Node: Acts as the seed.
Guider Nodes: Allow for multiple conditioning inputs.
Performance Comparison: Flux vs. Other Models
Flux 1.0 delivers impressive results across various prompts and styles. Here are some notable observations:
Anime Style Maid: Flux consistently produces better hands and overall proportions compared to AA Flow 0.1 and 0.2.
Realistic Victorian Mansion Scene: Flux outperforms in detail and realism, rivaling SDXL quality.
Text Encoding: Flux handles complex text prompts, like “flux” burning in fire, with remarkable accuracy, although it still faces minor issues with text clarity.
Future of Open-Source Image Generation
Flux 1.0 signifies a significant leap forward in open-source image generation. It addresses many of the shortcomings of previous models and sets a high standard for future developments. The ongoing competition among various open-source model creators, similar to the large language model space, promises to accelerate advancements in this field.
For content creators and developers, keeping up with these rapid changes can be challenging, but the benefits of improved models like Flux 1.0 are well worth the effort. Stay tuned for more updates and experiments with Flux and other emerging models.
Flux Recommended Dimensions & Aspect Ratio
2.0 MP (Flux maximum)
1:1 exact 1448 x 1448, rounded 1408 x 1408
3:2 exact 1773 x 1182, rounded 1728 x 1152
4:3 exact 1672 x 1254, rounded 1664 x 1216
16:9 exact 1936 x 1089, rounded 1920 x 1088
21:9 exact 2212 x 948, rounded 2176 x 960
1.0 MP (SDXL recommended)
I ended up with familiar numbers I’ve used with SDXL, which gives me confidence in the calculations.
1:1 exact 1024 x 1024
3:2 exact 1254 x 836, rounded 1216 x 832
4:3 exact 1182 x 887, rounded 1152 x 896
16:9 exact 1365 x 768, rounded 1344 x 768
21:9 exact 1564 x 670, rounded 1536 x 640
0.1 MP (Flux minimum)
Here the rounding gets tricky when trying to not go too much below or over the supported minimum pixel count while still staying close to correct aspect ratio. I tried to find good compromises.
1:1 exact 323 x 323, rounded 320 x 320
3:2 exact 397 x 264, rounded 384 x 256
4:3 exact 374 x 280, rounded 448 x 320
16:9 exact 432 x 243, rounded 448 x 256
21:9 exact 495 x 212, rounded 576 x 256
Support EndangeredAI &Join the Community
Patreon: https://www.patreon.com/endangeredai
Blackforest Huggingface: https://huggingface.co/black-forest-labs
ComfyUI Example: https://github.com/comfyanonymous/ComfyUI_examples/tree/master/flux
Text Encoders: https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main
Vae: https://huggingface.co/black-forest-labs/FLUX.1-schnell/blob/main/ae.sft