LCM Inference without blurring? WHAT KIND OF WITCHERY IS THAT!?!?
Note the change in the Clip has made the use of Hiresfix obligatory.
Please, see the "About this Version" section for info..
✦•··· Recommended Settings for DMD2 ···•✦
LoRa weight: 0.8
Steps: 8 ~ 10
CFG scale: 1.5 ~ 1.8
Sampler: LCM - Normal / SGM_Uniform / Karras / Beta
Hiresfix; 15+ hires steps, 0.3 denoising strength, any upscaler..
Description
IL 1.1 , LunaXL baked VAE, Fixed Clip for LCM (It's for Drawthings), use DMD2 LoRa.
The model creates distorted images in A1111 and Forge, but this does not happen in ComfyUI. You've been warned..
Model Precision: FP16
Apple Devices: Run in Drawthings, tested on MacBook Pro M5, Mac Mini M4, iPad Pro M2
FAQ
Comments (15)
I absolutely loved this model — to me, it's simply the best. I can’t understand why it doesn’t get more recognition. Anyway, I’d love to see an FP16 version, and I really need to get in touch with the creator. It’s urgent. So if you’re the creator and you’re reading this, please reach out.
need fp16 invoke wont run it
It feels that the fp16 version leans much more heavily into wanting to just create nudity. Which is fine if you want that sort of thing, but fp8 was a bit more varied.
There are two Loras that were removed from FP16 that caused certain artifacts, while I adjusted the weights of others to make the style more prominent.. But yep, now the model tends towards NSFW..
I don't know if I should return them, or retrain them and add them back in.
Karely_ai Up to you! This site leans more towards nsfw, but personally I prefer the other version.
Great model. Loving the soft skin renders.
I couldn't for the life of me get the LCM DMD2 LoRA working properly with it. I got close, but the results always had small artifacts and unpolished fingers. As soon as I switched to 24 steps with Euler A and the DMD2 LoRA, it was beautiful.
Was I just doing something wrong? Or is there a different DMD2 LoRA you're using? These are the settings I was using:
https://civitai.com/models/1608870?modelVersionId=1820946
DMD2 LoRA 0.7 model and clip weight.
LCM, Normal scheduler
1.7 CFG
8 steps
I use my own merge of DMD2, HyperSD and PSO, but that lora should work similarly bcuz that's the base I've used..
NO WAYYYYYYYYYYYYYYYYY!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
It's an amazing model, also for img2img - has good prompt adherence, when tasked with 're-drawing' the source, and yet add something from the description. :)
The generations are beautiful and the model seems awesome
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.