V1. Int4 Uncensored Krea 2 Turbo for Faster Generation (Supports NSFW)
V2. Int4 Uncensored Krea2Turbo for Faster Generation (Supports NSFW) (OUT NOW)
Updated the Checkpoint for Better Skin/Texture/Prompt Adherence with Lora
Optimized for RTX 2000 and 3000 Cards
Use Int4 W4a4 Checkpoint Loader
6-8 Steps
Euler with Simple
Baked in Realism
Description
Int4 Version Convrot - Converted from BF16 Version for Faster Generation on 3000 cards
Little Quality Drop
:D Enjoy!
FAQ
Comments (6)
wow! for 6GB model giving pretty much good results, using RTX3060 laptop, loras also good
How long does it take for you to gen a 1mp image? I didn't think it would work on 6gb vram that well
Well, I'm using a Stone Age graphics card, a 1660 Super with 6 GB, and it's slower than a snail; the 6-step image takes 2 minutes at 1 MP.
@Tarendax Try using wan2.1upscale VAE along with a gaussian blur node, I think my sample image workflow would help. Do check it out.
Sageattention does work a bit but wouldn't recommend it as it isnt that much of an improvement imo. Still tinkering about.
@mzimzisa I'll try it.
Just some free advice for Int4 users, the Wan2.1 VAE is compatible with Krea2, and makes hair and skin look much nicer. Use the FP32 version for superior quality. It really takes the edge off of the quality difference with Int4 quants, so you get both speed and sharpness boosts.