Anima 2.9B Preview v1 (SD.Next Diffusers Conversion)

Diffusers-format conversion of Anima 2.9B preview v1 by Gazingstars for use with SD.Next. Please find the original repository here.

Anima 2.9B is a depth-expanded fine-tune of Anima 1.0 Base: the transformer is grown from 28 to 40 layers by interleaved block insertion and trained on 1.7M additional anime and illustration samples. In preview v1 only the inserted layers are trained.

Original model: Gazingstars123/Anima-2.9B (Anima-2.9B-preview-v1.safetensors)

Conversion

Transformer: converted from the ComfyUI single file with the diffusers Cosmos 2.0 key mapping; transformer/config.json carries num_layers: 40; weights stay bf16 as in the original.
LLM adapter: taken from the same file, where it is bundled; unchanged from Anima 1.0 Base.
Text encoder (Qwen3-0.6B), VAE (Wan), tokenizers, scheduler and pipeline code: copied verbatim from Anima 1.0 Base, identical across Anima variants.
Verification against Anima 1.0 Base: all 28 original blocks are byte-identical at their source index and every inserted block has trained output projections.
Size: 6.9 GB total (5.2 GB transformer, 1.2 GB text encoder, 257 MB adapter, 243 MB VAE) against the 5.4 GB single file.

Architecture

Transformer: CosmosTransformer3DModel (2.9B params, 40 layers)
Text Encoder: Qwen3-0.6B (replacing Cosmos T5-11B)
LLM Adapter: Custom cross-attention adapter bridging Qwen3 to the transformer
VAE: AutoencoderKLWan

License

CircleStone Labs Non-Commercial License v1.2 (see LICENSE.md); this model is a derivative under that license. As a derivative of Cosmos-Predict2-2B-Text2Image, the model is also subject to the NVIDIA Open Model License. The non-commercial restriction applies to the model weights, not to generated images.

Anima-2.9B is officially supported in ComfyUI and Forge-Neo!

Join my discord server!: https://discord.gg/em4PjwZvWx

If you'd like to support me or to support the training progress:

Vast.ai: thangquay347@gmail.com

Every bit of support helps expand the model's scope and capability even further!

You may still need to install ComfyUI-Anima-2.9B to the custom node folder if your comfyui are not up-to-date (version 0.33.1). Plug and play, there is no custom node needed. Sometimes may not work with other custom nodes

Overview

Anima-2.9B is a fine-tune and layer-expansion of circlestone-labs/Anima. The base Anima model targets anime, illustration, and non-photorealistic art; this release continues training on that foundation with the expanded architecture. The model is trained on an additional 1.7M anime/illustration samples, with knowledge cutoff in July 2026, making Anima-2.9B one of the most up-to-date anime/illustration model at release.

Versions

Anima-2.9B-preview-v1: initial release

Training/Dataset

Trained using Muon optimizer on a 8x 5080s cluster, with earlier steps trained locally on my PC
As of preview v1, only the new layers have been trained, with roughly 70% of the compute spent on 1024px
Knowledge cutoff is July 2026, training data included both new and old samples prior to September 2025
Mixed captioning, including both tags and natural languages, using a mix of Gemini 3.1 Flash-Lite, Gemini 3.5 Flash-Lite, and Claude Sonnet 5
NO score tags

Architecture

Transformer depth expansion: expanded from 28 transformers layers to 40, growing the model to ~2.9B parameters. Each new layer is added by deep-copying its neighboring layer's weights, using interleaved insertion with zeroed-out output projections, making the new model functionally identical to Anima-base at initialization.

Prompting tips :

Follow Anima prompting tips: quality tags, year/period tags, @artist tags, character count (1girl, 1boy), character tags (follow Danbooru and Gelbooru tags), series/copyrights, base appearance.

Character name/tags should be follow with series/copyrights tags or else the model might confuse.

For multi-character images, attribute the character and names with their respective tags/appearance.

The model does improve the base art style slightly, but I'd still recommend using artist tags.

The dataset does not include scores in its captions, however, you can still use them.

(IMPORTANT) THE MORE DETAILED THE PROMPT, THE BETTER, short prompt will often generate a bland simple background, and may not able to produce the desire results

Generation (Recommendation)

Sampler: Euler/Res-multistep/Er-sde
Scheduler: sgm-uniform/beta/beta57/linear-quadratic
Resolution: 812x1216, 1152x1536, 1536x1536 (iffy)
Steps: 28-50
CFG: 3.5-5

My personal usage is euler + sgm-uniform, which has a good balance between composition and fine details. Additionally res-multistep + linear-quadratic spend more time at high noise steps, which does lead to visibly better composition. My recommendation for the highest quality is 50 steps, there are some images where 3.5 CFG do better than 5 CFG and vice versa. Experiment yourself!