If you like this model, remember to give a like to / follow the original!
Mol_Keun released an Anime 3.8B version of their Mol_Keun Mix. But it seems it isn't working for some people who don't use ComfyUI? So here's a conversion of it I tested and confirmed works on Forge Neo. It's also quantized to INT8_convrot to run on my potato and run faster.
Update: Added the original bf16-precision denoiser without the embedded TE and VAE, people with more VRAM than I can test if it works on Neo for them.
Update: Also added the non-3.8B version converted to int8_convrot since it can go on the same model page. If using a Turbo LoRA with it, you'll need to turn the strength down, 0.7 worked for me.
Update: Added int8_convrot conversion of Mol Keun COMIX, OGMustard's merge of Mol Keun Anima and JetlistiC's MaturaComix.
Get Anima's Qwen3 0.6B text encoder from their HuggingFace page. Optionally, get the VAE there too, or use any Anima VAE. I believe Mol_Keun originally bundled Qwen2D, which also has an Anime version.
All credit to the original merge author Mol_Keun. If they don't want quantizations like this posted just dm me and I'll take it down. Wow they featured it on their reccomended resources, guess they approve! Glad I could help get their checkpoint into the hands of Forge Neo users! =M
Apologies for the low quality showcase images, I'm on a GTX 1060.
Mol_Keun's recommended settings:
Sampler: Euler a
CFG Scale: 4~6
Steps: 25~40
Clip Skip : 2
Tested settings for Turbo LoRAs: (with Anima-2.9B-Turbo-v2.9-remap, probably works with Anima 2.9B Turbo Lora Remapped too)
Samplers: ER SDE, Res Multistep, Euler a
Schedulers: Beta, Simple, Linear Quadratic, Bong Tangent
CFG Scale: 1-2
Steps: 8-16
Clip Skip: Whatever the default is