anima as an illustrious replacement?
i’ve been playing around with the anima image model these past few days and i’m absolutely thrilled with it — it works 3–4 times faster, while the results are even better
for anyone who wants to try it, here’s a small guide for now, only via comfyui. the upcoming update will also include koboldcpp support •ᴗ•
→ launch comfyui → drag the workflow .json file directly into the comfyui window → download the models and place them into the proper folders:
anima-base-v1.0.safetensors → ComfyUI/models/diffusion_models/
qwen_3_06b_base.safetensors → ComfyUI/models/text_encoders/
qwen_image_vae.safetensors → ComfyUI/models/vae/
anima-turbo-lora-v0.2.safetensors → ComfyUI/models/loras/
example workflow .json for 3 loras. if you don’t need two of them, remove them, but i recommend keeping the first one, anima-turbo-lora, for speed
json
anima_lora_3slots.json3.44 Kb
base prompt: masterpiece, best quality, incredibly absurdres, adult
loras from the workflow used for the images above:
nerd stuff, for anyone curious how it works:
diffusion without turbo works like this: the image is born from noise iteratively. on each step, the model “denoises” the latent a little bit. the base model needs 20–30 of these steps, and every step is a full pass through the entire model.
what turbo-lora does: it’s a distillation lora, trained separately to reproduce the result of a full 20–30 step generation, but using larger jumps along the same trajectory. the weight difference — “student minus base” — is packed into the lora. you plug it in, and anima can reach a clean image in 8–12 steps instead of 25–30
why cfg-scale is 1 with it — and this is the second boost: cfg is the force that pulls the model toward the prompt. to do that pulling, the model needs a comparison. on every step, it draws two versions: one with the prompt and one without it, then shifts the result toward the first one. the higher the cfg, the stronger the shift
at cfg = 1, there’s nothing to compare: the model draws only one version, and the second pass doesn’t happen at all. that’s exactly what the turbo lora needs — it was trained to follow the prompt directly, without the comparison-guidance trick
lora
anima
illustrious
comfyui
turbo lora
images
Jussi Vesa
Okay this is actually quite incredible and so fast! Can it be included in game somehow?
Jul 13 13:18
Jussi VesaReplying to Multiic
Multiic, Nevermind, I found your guide in game. This is truly a game changer simply because it's so fast and I personally think it generates better images with the models you suggested.
Jul 13 20:56 
1
MultiicReplying to Jussi Vesa
Jussi Vesa, no, i mean some of the comfyui settings will be integrated directly into the game, so u won’t need to launch comfyui separately with the anima model. it’ll start automatically in the background together with the game
Jul 13 21:01
