DataVoid

SeFi-Image shipped with inference code but no training path. Two days later, first working LoRA finetune, built from scratch. Had to implement their core novelty from the paper: Semantic-First Diffusion, where the semantic latent denoises ahead of texture by At=0.1, giving the model a structural anchor before generating fine details. 500 steps, 40 images, 4 prompts stay diverse throughout:

All the Varients thus far for hand editing Krea2 Turbo weights

And finally, all.

It blends realism and digital drawing into the image when not requested. Further investigation is needed for this and the following results, as it's not as clean-cut.

b3 attn %10 pushed more detail into the water. Lightens the tone of the entire scene. reduces details in the subject but increases artistic flair

Krea Turbo 2 works with hand editing weights! Took the 12B DiT, inverted 10% of block 12-14 weights (w*0.8 blend), no retraining, no LoRA, just raw weight math. Model stays fully functional and gets notably more artsy. Block-selective weight manipulation on DiTs is relatively unexplored territory. Left: original. Right: 10% inverted blocks 12-14. (same infer settings & seed) It's shocking to me that nobody is hand-editing weights in these models even now. Next: attention-only targeting, magnitude-selective (TIES-style), and value-projection inversion across all blocks. 6 variants generated, this is the first functional result. #ComfyUI #Krea2 #Krea #DiffusionModels

@elder_plinius @max_paperclips @amasad

exactly! What I am doing right as of this moment. its ticking me off because i need some of those resources for a project i am working on.