I used flux dev de-drilled for this training(url:https://civarchive.com/models/843551/fluxdev-dedistilled)
At the same time, I remade the training set
The improvement this time mainly lies in the fitting of the art style and the diversity of characters (now diversity is much better)
By the way, there were no trigger words this version
The training effect of de distortion has indeed improved, but not significantly
The main issues are the 'strange' image quality of flux (it likes to generate smooth textured skin) and the lack of rigidity of flux itself (poor learning of atmosphere creating)
But I think this may be the limit of flux's ability to fit thick painted styles (I have tried my best)
Flux has particularly good magnification ability, so it can still generate visually appealing images under specific prompts
But this may be my last training session
On the one hand, Flux no longer has a successor
On the other hand, the training of flux has been explored enough, and I feel that there is not much room for improvement anymore
Thank you to those who have used my model from SD1.5 to Flux. Although we have never met before, we appreciate your companionship over the past three years
Please use the flux dev de-retail model to use Lora, but the original flux model can also be used
This is the workflow, t2img:https://drive.google.com/file/d/1xmQOCsO0d3s7CNztA_3LQS4YrncY2Cel/view?usp=sharing
upscale:https://drive.google.com/file/d/1-42TV4-iFEGfXlTMZshmBLk3LrtAV3pG/view?usp=sharing
All prompt words come from previous model prompt words and are directly generated using Comfyui
have fun
Description
For learning and communication purposes only, not for commercial use.
This version currently has some issues with diversity, and the next version is under training.
The image quality and semantics of this version are quite good.
The upscaling effect of webui-forge is not ideal. To achieve detailed and high-quality results, it needs to be used with comfyui workflows.
I will share my own comfyui workflow settings :
It must be noted that achieving the desired image quality requires parameter adjustments, proper prompt organization, and the use of upscaling to fully realize the quality. Avoid directly using reverse-engineered prompts, especially those containing terms like "realistic," as they may degrade image quality. Using GPT to help organize prompts can yield good results.
The prompts shown in the sample images are the final upscaled prompts and may not match the initial text-to-image prompts.
Setting the LoRA weight to around 1.2 is recommended. Adding a trigger word like "wgz style" (a newly coined term for AI) may also help.
If you have any questions, leave a comment below. All sample images are directly generated by the model without any post-processing or Photoshop editing.
FAQ
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.