modelos/dmd2 4step color fix - v1.0

dmd2 4step color fix - v1.0

11/3/2025

5:05:21 PM

Palavras-chave e Tags Relacionadas

4 step,accelerator,ays sd1,dmd2,dmd2 4step color fix,fix,illustrious,kirai,lcm,lora,power merger lora,red bias fix,saturation/color lora,sdxl vpred model,ferramenta,v1.0

Silhouetted couple standing on a misty hilltop against a dreamy teal and pink sky, with a red-orange moon and birds flying in the background.

Close-up portrait of a woman with dark hair featuring a double exposure effect blending misty mountains and sea in a cinematic and surreal style.

Anime girl with long purple hair sitting on a wooden chair, wearing an oversized shirt with a creeper print, holding a steaming coffee mug with pastel background

Parâmetros Recomendados

steps

cfg

Dicas

Use POWER MERGER LORA with 0.1 strength saturation slider for best results.

File size is reduced from 1GB to 64 rank using resizing utilities.

LCM and AYS SD1 with cfg 1 and 12 steps are the recommended settings.

Workflow images can be obtained using tinyterra on comfyui.

this is a niche case fix where i noticed that when using LCM and AYS SD1 on a sdxl vpred model brings red-bias, this lora aims to try to fix that by merging with a saturation/color lora

recipe uses

https://huggingface.co/tianweiy/DMD2

1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,0 of dmd2

concat method POWER MERGER LORA with 0.1 strenght saturation slider

https://civitai.com/models/1093089?modelVersionId=1227749

file ends up being 1gb but resized with some utilities down to 64 rank

i knew the last layer of the lora was the cause of the red bias but didnt know how to fix it without removing the block so a mix of strenght was my way of solving it, can have perfect white, perfect red and perfect colors with sometimes 2-off, so its fine for me at this point XD

thanks and credits to a certain supernova

for pointing out the methods and the lora

recommended settings that i use

cfg 1

LCM

AYS SD1

12 steps

https://civitai.com/models/510261?modelVersionId=567119

as an extra step lora

PD: you can get the workflow of the images if you have tinyterra on comfyui