Anima 3.8B · Qwen 0.6B

Hey, this is Anima-Expanded 3.8B 👋

52 diffusion transformer layers. Roughly 4.1 million images seen during training. The native Qwen 0.6B text encoder.

This model comes out of our Anima 3.8B + Qwen 3.5 4B experiment: we expanded Anima 2.9B from 40 to 52 layers and trained the expanded model on a roughly 50% anime / 50% furry data mix.

For this release, we kept the trained diffusion model and brought conditioning back to Anima's original Qwen3-0.6B text encoder. The additional Qwen 3.5 connector has been removed, while all 52 diffusion layers and the original native text adapter remain.

The expanded model, ready for your regular Anima workflow. ✨

More room for the details you care about

The goal is simple: stronger furry generation, better explicit-content handling, and richer backgrounds, alongside the anime illustrations that brought us to Anima in the first place.

• 🐾 Furry characters: more convincing species traits, expressive faces, fur, ears, tails, and character detail.

• 🔥 Adult / explicit imagery: better anatomy, posing, and interactions between adult characters.

• 🌄 Backgrounds: more complete environments, clearer scene structure, and details that make the setting feel like part of the image.

• 🎨 Anime illustrations: expressive characters, appealing color, and compositions worth revisiting.

These are the areas we want the expanded model to improve. Bring your comparison prompts—we want to see which gains hold up across different styles, subjects, and seeds.

What's inside?

• 52 diffusion transformer layers, expanded from the original 40.

• Approximately 3.8B parameters in the diffusion checkpoint, including its native text adapter. The text encoder is loaded separately.

• Roughly 4.1M images seen during training, with an approximately even anime / furry mix.

• Native Qwen3-0.6B conditioning, using the original Anima text adapter.

• A standalone BF16 diffusion checkpoint that loads through standard ComfyUI nodes.

This is a native-text-encoder variant of the trained 3.8B experiment. It has not been separately retrained or distilled for this release.

Load it like a normal Anima model

Use an up-to-date ComfyUI with Anima support. No custom node is required, including for the expanded 52-layer architecture.

ComponentFileFolderDiffusion modelAnima-3.8B-v1.1-Qwen06B.safetensorsComfyUI/models/diffusion_models/Text encoderqwen_3_06b_base.safetensorsComfyUI/models/text_encoders/VAEqwen_image_vae.safetensorsComfyUI/models/vae/

1. Select the checkpoint in Load Diffusion Model, with weight dtype set to default.

2. Select the Qwen 0.6B encoder in Load CLIP, with type set to stable_diffusion.

3. Load the VAE, then use the usual CLIP Text Encode → KSampler → VAE Decode workflow.

ComfyUI detects the 52 layers automatically. You can use your existing native-Anima workflow and swap the diffusion checkpoint.

A starting point

The following settings were used for a successful generation check:

Steps 24

CFG 4.0

Samplereuler a

Schedulersimple

Resolution 832 × 1216 or other 1MP

Quality tags not needed but supported

Treat these as a starting point and experiment with your usual Anima settings. The diffusion checkpoint is approximately 7.5 GB; runtime memory also depends on resolution, the text encoder, and the VAE.

Take it for a spin

This is an experimental release, and we're curious to see what you make with it. Try your favorite prompts, revisit an old seed, and show us where the extra layers take you.

Anime, furry, familiar characters, original designs—bring your prompts and share your results. Images with prompts, seeds, and settings are especially useful for comparing the expanded model with the original Anima. 💛