Furry/feral checkpoint with a default style influenced by my uwustyle LoRa, emphasizing strong outlines and colors.
Fast versions are optimized for low CFG and 6-12 step generation. I use a blend of DMD2 and SDXL-Flash that I think preserves the style well, but there will still be differences from base versions.
FP8 pruned model stores the UNet and CLIP models in FP8, but the VAE in FP16. This reduces the byte size by almost half in return for a slight quality loss, but most likely won't support hardware-accelerated FP8 inference (so it won't generate faster).
Due to the contrast & saturation enhancement of the VAE, I personally use vaemix v6 (or similar anime/toon VAE) for the hires-fix step to avoid the excessive contrast and color "burn" that can sometimes happen.
Description
This version is uwumerge v2 + DMD2 * 0.65 + SDXL-Flash * 0.55
Optimized for 6 to 12 steps and CFG of 1.0 to 2.0, but quality will be lower than v2.0 in some cases
FAQ
Comments (8)
Any chance we get an FP8 version of this new low step merge? the previous FP8 ones run surprisingly well on very low vram systems
I didn't do one because I thought that the FP8 version really only saved disk space and the quality would be too bad to be worth it, but I'll go ahead and upload one.
@DarkFawkes Thanks! It's kind of perfect for running on my craptop, with 4gb of vram and 8gb ram, fp16 just can't fit entirely without spilling into swap space on it but your fp8 ones fit just fine even if they take a bit to run. It's nice to have while my desktop is busy with other things
@FloatsYourStoat It's uploaded now but the site hasn't virus scanned it yet.
@DarkFawkes Oh wow, that was fast, thanks a bunch!
@DarkFawkes Initial testing seems to be pretty acceptable honestly, with 6 steps and basic guider i get a 1024x1024 image in ~25 seconds on a woefully under powered gpu
@FloatsYourStoat I'm glad it works for you. If you would like to prune & convert other people's models to FP8 yourself the way I do, so that they run on lower VRAM, I use the "Model Pruner (MTB)" node from the ComfyUI extension "comfy_mtb". I don't know if it might require a higher end GPU to do the conversion process though.
@DarkFawkes Thanks for the tip! i'd also need to merge those low step loras too for it to be more reasonable gen times, wish they didn't give that strange texture to the images but such is the cost of using shortcuts i guess.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.