Flux_Sevenof9_12Step_Distilled版本Sevenof9_shut_dist_Q5KM (ID: 1256818)

RE-UPLOAD v2 - 11.1.

at least 12 Steps max 25

Its a merge of:

DEV_Shuttle https://civitai.com/models/943001?modelVersionId=1058817
and
FLUXDEV_FP8 https://civitai.com/models/622579/flux1-dev-fp8?modelVersionId=695998
with my lora
distilled https://civitai.com/models/977247/flux-distilled-lora

so you can use CFG ~2 (but you don't have to)
Q3_K_M running fully on 12GB VRAM
dont go above1200pix

for me it is a start to merge models and loras because i cant train a model, iam only a single one
v2(FP8,Q3) - nudity, poses on board, BUT dont expect to much fine details, (add freckles, body freckles), all girls look like next door
v1(FP8,Q5) - its a well balance between steps (not to low) and CFG 2 for complex prompts

Euler with Beta or SGMUniform
or
(swarmui)DPM++2Sa with Beta (better skin tone)
or
dpm++2M with Beta

Tell me please how it works ... Loras should be work at least the main direction (maybe some details gone)

some pictures with my Loras to show how it works, my impression some times bit lower weight than usual

Here are some comparisons between Steps and CFG

https://civitai.com/images/46982169
https://civitai.com/images/46984474

描述:

Q5KM version of first FP8 model

训练词语: