Anima-Expanded 2.9B — BASE & Aesthetic
Layer-expanded Anima. 40 DiT blocks, 2.9B params, and it still runs on a completely normal Anima setup - native Qwen3-0.6B encoder, same VAE, no custom nodes. If you already have Anima working, this is a checkpoint swap.
Trained on roughly 4.1M images, about 50/50 anime and furry.
Two versions in this drop:
BASE - the plain expanded model. Neutral, does what you prompt and nothing you didn't ask for. This is the one for stacking your own LoRAs, or for comparing against your usual Anima 2B.
Aesthetic — tuned hard on lighting, color and backgrounds. Much stronger look straight out of the box, less neutral. Take this one if you want the model making some decisions for you.
Same prompt, same seed, both checkpoints. Fastest way to see what changed.
What I pushed on
Furry — species traits that actually read, fur/ears/tails, faces with expression
NSFW — anatomy and multi-character posing between adults
Backgrounds — full environments instead of blurred filler behind a character
Anime — color and expression, compositions that aren't just centered bust shots
Backgrounds and furry moved the most in my own testing. Multi-character NSFW is still inconsistent — if you find prompts that land it, post them.
ComfyUI
Needs a recent ComfyUI with Anima support. No custom nodes for either version.
ComfyUI/models/diffusion_models/ BASE or Aesthetic (~5.8 GB, bf16)
ComfyUI/models/text_encoders/ qwen_3_06b_base.safetensors
ComfyUI/models/vae/ qwen_image_vae.safetensorsLoad Diffusion Model → weight dtype
defaultLoad CLIP → type
stable_diffusionThen the normal CLIP Text Encode → KSampler → VAE Decode
Switching versions is just the diffusion checkpoint. Encoder and VAE stay put.
Recommended settings
Sampler: Euler a (
euler_ancestral)Scheduler: simple
Steps: 24
CFG: 4.0
Resolution: your usual Anima resolutions
Notes
Post your gens in the gallery - prompts and seeds if you can, Comments and bug reports welcome, I read them.
If you get something good out of it, a review or a like helps more than you'd think.
Description
Aesthetic version was trained at multi res 1024x1024 1536x1536 on a custom high quality dataset.
FAQ
Comments (6)
50 50 split is good news. did you do anything for tag aliases between the boorus? or any specific processing?
I also don't really understand why they are divided into 3.8B and 2.9B, or what the difference is between the Expanded 3.8B version and the non-Expanded version. As for Expanded, if you use the default Anima workflow, the Expanded version and the previous version will generate exactly the same results.
I think it's divided because this is a reworking of two separate projects, the 2.9b anima variant and the 3.8b variant, modified to require less plumbing and function like a normal anima model.
Is this version have more new characters? I can't generate Sparxie when using this model.
Base version feels like a common finetune of Anima, styles are altered, but still recognized.
But Aesthetic is worse - though overall composition is improved, most artist names just slightly change the default bland AI style
If you recommend 24 steps and cfg 4, why all your post are with 20 steps and cfg 6?