GitHub radar
Anima 2.9B: Anime Art on Cosmos Architecture
Anime illustration text to image model fine-tuned on 1.7 million anime images with July 2026 cutoff.
Anime illustration text to image model fine-tuned on circlestone-labs/Anima on NVIDIA’s Cosmos-Predict2-2B, has been trained on 1.7 million anime images with a July 2026 cutoff. It has 2.9 billion parameters and was created by increasing the 28 layers of the original architecture to 40 layers through interleaved insertion and setting the output projections to zero. It was trained on a mix of captions from both Gemini and Claude (including both tag and natural language descriptions but no score-based tags), and can be loaded into ComfyUI as a single file diffusion checkpoint (.safetensors). License: Non-commercial.
Why a vibe-coder should care
Anima-2.9B is a fine-tuned text-to-image diffusion model, based on NVIDIA's Cosmos-Predict2-2B, with 2.9B params. It was fine-tuned on circlestone-labs/Anima and gazingstars123 increased the 28 layers to 40 with interleaved insertion and zeroed out output projections. It was trained on 1.7M anime and illustration images, with a July 2026 cutoff, using a mix of captions from Gemini and Claude (both tag-based and natural language), sans score-based tags. It is a single file diffusion checkpoint (.safetensors) for ComfyUI. License: Non-commercial.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Help me set up Anima-2.9B in ComfyUI: https://huggingface.co/Gazingstars123/Anima-2.9B — download the .safetensors checkpoint, put it in models/checkpoints, load it as a single-file diffusion model and show me how to generate the first anime illustration from text.
Requires a GPU with 8 GB of VRAM or more, or an Apple Silicon Mac — the model is about 6 GB in bf16 format for ComfyUI.
Open on Hugging Face▌ More finds