The FLUX version is coming!
This time, a brand new FLUX large model is introduced, adding many scene elements, and the generalization of materials has been improved. Whether it is details or the understanding and restoration of prompts, there has been great progress. Welcome everyone to try it out.
Tip: Because the miniature figures are relatively small, it is recommended to turn on 2 times high-definition restoration: R-ESRGAN 4x + Anime6B to correct the detail problems of the small figures.
V2 version raid update, optimize many problems appeared in V1 version, add your more whimsical ideas!
Optimization Notes:
1. Fixed the problem of lower quality of the image when the high resolution fix (Hires. fix) is not turned on.
2. Improve the accuracy of the rendering and add more texture details.
3. Double the number of materials, add more common items and food, such as: fabric, wool ball, chocolate, ice cream, beehive, etc.
4. Reduced the number of mandatory words, now you don't need to enter: low saturation colors, miniature characters, soft light, etc., you just need to enter the trigger word weisuo to create a picture.
5. Remove the limit of repainting range, you can choose the repainting range according to the actual situation.
6. Welcome to try to return to the picture ah.
Recommended Parameters
Positive quality word reference:
weisuo.
Negative tips word reference:
NSFW, (can be filled in or not, basically does not affect the quality of the picture)
Big model: majicMIX_realistic_mai orange realistic_v7, https://www.liblib.art/modelinfo/bced6d7ec1460ac7b923fc5bc95c4540
Lora weight: 0.8
Sampling method (sampler): DPM + + 3M SED Karras (according to the actual situation can choose sampler)
Iteration steps (steps): 24
CFG Scale: 7
Resolution reference: 512*768 recommended (vertical map)
High resolution fix (Hires. fix): 1. zoom algorithm: ESRGAN 4x, 2. redraw amplitude: (unlimited), 3. zoom speed: 2x
Description
The FLUX version is coming! This time, a brand new FLUX large model is introduced, adding many scene elements, and the generalization of materials has been improved. Whether it is details or the understanding and restoration of prompts, there has been great progress. Welcome everyone to try it out.
FAQ
Comments (7)
Hello, I encountered the following problem when using it:
calculate_weight() got an unexpected keyword argument 'intermediate_dtype'
My checkpoint uses FLUX_dev_fp8, and the controlNet used is also FLUX. The problem should not be my controlnet.
After I disabled your lora, the K sampler can work normally again
How should I solve this problem?
Have you tried changing to other samplers?
You can also try to update Comfy UI to the latest version.
Thank you. Can I see your workflow? I also want to make it cleaner and better.
Just eliminate extraneous cues when labeling and keep the training images consistent in style!
@1jiliguala1 Without the need to add nodes, is it a matter of prompts?
I'll give it a try. Thank you!
@schampion456159 Just load it like a flux lora.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.