V2

anima-base-1 + some of my loras + official anima-turbo-lora-v0.2

test in forge-neo

ER SDE BETA

step=4-16 prefer 6

CFG=1-4 prefer 1.25

offset=3-12 prefer 8

The advantage of this version which is using the official Turbo Lora is that it's very fast; a 5070TI can be generated in less than 5 seconds in 6 steps, which is quite sensitive to prompts. The problem is that because it's generated so quickly, it's not sensitive to seed changes, and the output is almost the same with same prompt.

V1

anima-base-1 + some of my loras + Cosmos-Predict2.5-2B-base-distilled-LoRA

test in forge-neo

ER SDE BETA

step=8-16 prefer 12

CFG=1-6 prefer 4

offset=3-6 prefer 4