JustAnotherFurryMerge

JustAnotherFurryMerge (JAFM) is my personal merge I made using ComfyUI. It first started as experiment with merging random furry models I have. Turns out it gens very well so I decided to share it with anybody out here. It is intended to generate furry anthros. It can do males and possibly anime as well, however I'm not into that. Using mostly e6 tags for my gens.

This merge is eps prediction. It has a piece of v-pred in it, but since basemodel is eps the output model is eps as well.

Recommended settings:

Steps: 10 - 30 (16+ for zoom-in better quality)
CFG: 2 - 6 (6+ if you like oversaturation)
Sampler: Euler a, dpm++ 2m sde, ddpm, seeds 2
Scheduler: beta > karras > normal
Other uncommon works too, just keep experimenting

Prompting: Using quality tags seems to improve quality in some way. Probably does not matter if you put them on beggining or end of prompt. Use whatever you prefer.
Resolutions: Anything between 0.9 and 2.4 MPx is okay. Unless too vide or tall. Or try this :)

I personally gen with Euler a@beta on 30 steps@3 cfg at 1080p + 2-pass img2img upscaling to 4k with dpm++ 3m sde gpu@karras on 16 steps only and 0.16 denoise, same cfg (v4.0 showcase workflow). Takes a while but deffo worth the wait. I don't guarantee it to work under 32GB ram/12 vram, anyway nothing stops you from trying it :P

Running showcase workflows:

First make sure you have latest version of ComfyUI with Manager with and all custom nodes installed. Copy pasting again from Civitai should not throw any missing nodes.
Don't forget to check suggested resources for detectors I used. They need to be placed in "models/ultralytics/segm/" on the same drive your ComfyUI is saved. Having "ultralytics: models/ultralytics/" in "extra_model_paths.yaml" file does not work and probably never will.

If workflows from older versions doesn't run with correct model/vae selected, try using from never ones, copy prompt, seed and choose model. This might happen as ComfyUI with customnodes getting frequent updates. I also genned v2.0 showcases on high resolutions. Anyway I changed workflow on some showcases in v3.0 and v4.0 are all 4k upscaled.

Version history:

• 1.0 & 1.0unstyled - First release

• 2.0 & 2.0unstyled - Hires rawgens quality improvement

• 3.0 & 3.0unstyled - Small knowledge enchancement

• 4.0 & 4.0unstyled - Enhancing the style

As for v4.0, unstyled is on my HF repo since I don't think it's worth posting here. You can also find there my vae, beta versions, premixes and my experiment as int8 convrot quants of v4.0 & 4.0unstyled. Try gooning with 2GB less memory!

Credits

Thank you all for creating such nice models! Without you, my own would not be possible:

Checkpoints
Agri_Yob, AI_nstein, Anomalous_Anonymous, BlessedBooba, Chenkin, Crody, DarkFawkes, ecaj, EJtech, GreyFerret, iDeNoh, jhonny09, kftiger, LatteLeopard, Mol_Keun, multilazor, Ocean3, Old_Arctenox, ONOMAAI, ReedVaughn, RodAndRog, SXCD2719, thebiologist92137

Loras
Anzatiridonia, EauDeNoire, iamddtla, Meower2024, ntc, UnholyDesiresStudio, Volnovik, w4r10ck

Vae
First 2 releases stabilityai, ongoing myself

Post your gens, share if you like it! And don't forget, no reprints please.

And as always
>> Just keep generating
>> Don't ever try to stop
>> Share your best works
>> Seek new motivation