ifsin/Anima-Aesthetic

Anima Aesthetic 1.1: quantized files with the turbo LoRA

A ready-to-use file set for Anima Aesthetic 1.1 (2B, anime) with the turbo 4-step LoRA: the diffusion model in two sizes, the text encoder, the VAE and the LoRA. The model, text encoder and VAE are from Anima by CircleStone Labs. The turbo LoRA is from sorryhyun/anima-turbo-4step.

These are conversions of other people's work. They are not official releases and are not endorsed by CircleStone Labs, the LoRA author, or NVIDIA.

Files

File	Size	Role
Anima-Aesthetic-1.1-Q8_0.gguf	2.24 GB	Diffusion model, 8-bit. The default.
Anima-Aesthetic-1.1-Q4_K.gguf	1.20 GB	Diffusion model, 4-bit K-quant. For low memory.
Qwen3-0.6B-Base-Q8_0.gguf	633 MB	Text encoder, 8-bit.
Qwen-Image-VAE-F16-Image.safetensors	94 MB	VAE, 16-bit, single-image kernels.
Anima-Turbo-4Step-v3-LoRA.safetensors	201 MB	Turbo LoRA v3, BF16.
SigLIP2-Base-Patch16-512-Vision-F16.safetensors	173 MB	Optional. Image encoder for reference pictures: the vision half of Google's siglip2-base-patch16-512, converted to 16-bit.
Anima-IP-Adapter-Character-Reference-10.safetensors	503 MB	Optional. Anima IP-Adapter (Character Reference), BF16, unchanged from LuciferTC/Anima-IP-Adapter.
Anima-3D-Pose-LLLite-6200.safetensors	51 MB	Optional. ControlNet-LLLite for a shaded 3D body mesh as the control picture (fixes the pose and framing), unchanged from toyxyz/Anima_ControlNet-LLLite (anima_3dpose_6200.safetensors), renamed.
SHA256SUMS		Checksums for every file above.

The SigLIP2 and IP-Adapter files are only needed to draw a character from a reference picture (about 0.7 GB more). The pose adapter is only needed for the fixed profile framing (51 MB).

A full set is the diffusion model, the encoder, the VAE and the LoRA: about 3.2 GB with Q8_0 or 2.1 GB with Q4_K.

Which diffusion model

	Q8_0	Q4_K
Size	2.24 GB	1.20 GB
Closeness to the 16-bit model	nearly identical	same style and composition, slightly softer shading
Prompt following	best	a little looser on small details

Both were compared with the original at fixed seeds on portraits and scenes. Plain q4_0 was tested and left out because it shifted the look of the images toward semi-realistic.

Recommended settings

With the turbo LoRA: 4 steps, CFG 1, sampler er_sde.

License

• Anima and the turbo LoRA are under the CircleStone Labs Non-Commercial License. The model and derivatives are for non-commercial use only. The restriction applies to the weights, not to generated images, which the source page says may be used commercially. Commercial use of the weights needs a separate license from CircleStone Labs.
• Anima is also subject to the NVIDIA Open Model License Agreement as a derivative of Cosmos-Predict2. Built on NVIDIA Cosmos.
• The IP-Adapter weights are listed on Hugging Face with license "unknown"; its code repository is Apache-2.0 and says the weights follow the CircleStone Labs Non-Commercial License. Treat it as non-commercial.
• The 3D pose adapter (toyxyz) declares the license field "upstream-anima-and-source-data-terms" and says its weights are derivatives of Anima under the CircleStone Labs Non-Commercial License. The page does not state redistribution terms. Treat it as non-commercial.
• SigLIP2 is released by Google under Apache-2.0.
• Qwen3-0.6B-Base is released by the Qwen team. See its model card for the license.

This summary comes from the source pages, not from the full license texts. Read the originals before using or redistributing these files. The Anima page does not spell out terms for redistributing quantized versions.