This model is a Tile & Repair ControlNet-LLLite model for Anima, trained with part of the edited image-pair data from the Noob v2 project, and designed for anime image restoration, tiled detail enhancement, and repair-style image-to-image workflows.

It is trained as a lightweight ControlNet-LLLite guidance model for the Anima model family. It is not a standalone image model. You should load it together with an Anima-compatible inference pipeline or workflow.

What this model does

This model is intended to help Anima restore and improve degraded anime images while keeping the original layout and character structure stable.

Typical use cases include:

repairing blurry anime images;
restoring low-quality or low-detail images;
reducing visible noise and compression artifacts;
improving local details in tile / repair workflows;
preserving the original composition while making the final image cleaner and sharper.

In the current v1 release, the model already performs well on blur-damaged images and low-quality degraded images. I have also added a large amount of new training data and I am currently training v2, which is expected to be released next week.

Recommended checkpoint

Use:

anima-base-v1.0.safetensors

Earlier checkpoints are also provided for comparison, but the final v1 checkpoint is recommended for normal use.

How to use

There are two main ways to use this model:

Python inference with the Anima ControlNet-LLLite script
ComfyUI workflow through the ControlNet-LLLite_node

Python inference usage

You can use the provided Anima ControlNet-LLLite inference script:

python anima_minimal_inference_control_net_lllite.py \
  --dit /path/to/anima_dit_or_model \
  --vae /path/to/qwen_image_vae \
  --text_encoder /path/to/qwen3_text_encoder \
  --lllite_weights /path/to/anima_tiled_lllite_v1.safetensors \
  --control_image /path/to/input_or_control_image.png \
  --prompt "restore this anime image with clean details, sharp line art, and natural texture" \
  --image_size 1024 1024 \
  --infer_steps 50 \
  --guidance_scale 3.5 \
  --lllite_multiplier 1.0 \
  --save_path ./outputs/

For batch inference, you can use a prompt file:

python anima_minimal_inference_control_net_lllite.py \
  --dit /path/to/anima_dit_or_model \
  --vae /path/to/qwen_image_vae \
  --text_encoder /path/to/qwen3_text_encoder \
  --lllite_weights /path/to/anima_tiled_lllite_v1.safetensors \
  --control_image /path/to/default_control_image.png \
  --from_file prompts.txt \
  --save_path ./outputs/

Example prompts.txt line:

restore this blurry anime image with clean line art and improved details --w 1024 --h 1024 --d 42 --cn images/input_001.png --am 0.8

Useful parameters:

--lllite_weights      Path to the ControlNet-LLLite .safetensors file
--control_image       Control / reference image path
--lllite_multiplier   ControlNet-LLLite strength
--cn                  Per-prompt control image override in batch mode
--am                  Per-prompt multiplier override in batch mode

A good starting point is:

--lllite_multiplier 0.8 ~ 1.0

If the repair effect is too weak, increase the multiplier slightly. If the result becomes too sharp, too constrained, or starts changing fine details too much, lower the multiplier.

ComfyUI usage

You can also use this model in ComfyUI with the ControlNet-LLLite_node workflow.

Basic idea:

Anima base model
+ ControlNet-LLLite_node
+ anima_tiled_lllite_v1.safetensors
+ input/control image
= repaired Anima output

Please note: some testers reported a slight color shift when using the ComfyUI node. I have not observed the same issue in the Python / diffusers-style inference path, so this may be a Comfy node-side bug or workflow-specific issue. If you encounter this, please compare your result with the Python inference path and feel free to report the issue with your workflow settings.

Suggested prompts

For repair / tile workflows, you can try prompts like:

restore this anime image with clean details, sharp line art, and natural texture

repair the low-quality anime image, reduce blur and compression artifacts, preserve the original composition

enhance the image details, clean up artifacts, keep the character structure and scene layout unchanged

restore fine anime line art and local details while keeping the original pose, composition, and colors stable

Communication

Training information

This v1 model was trained on A100 GPU resources.

The model focuses on anime-style restoration and tile / repair guidance. It is optimized for Anima workflows rather than general photographic restoration.

Limitations

This is a v1 release, so there are still limitations:

It is mainly designed for anime images.
It may not work well on realistic photos.
Very strong guidance may over-sharpen details.
Some heavily degraded images may still require stronger restoration or a future version.
ComfyUI node output may show a slight color shift in some workflows.

NODE INPUT&OUTPUT

Roadmap

I am currently training v2 with a larger and more diverse dataset. The next version is expected to improve robustness on blurry, low-resolution, and low-quality images.

Expected release: next week.

Credits

Special thanks to Comfy.org for providing GPU sponsorship.

Thanks also to the volunteers who contributed testing and feedback:

Yidhar
GHOSTLXH
年糕特工队
轻松
Free Will

Their feedback helped improve the training and release process.

License / Usage

Please follow the license and usage terms of Anima and the related ecosystem components. This model is released as an auxiliary ControlNet-LLLite guidance model for Anima-compatible workflows.

Search AI Models & Addons

very few results