Anima Tile & Repair ControlNet-LLLite

A lightweight ControlNet-LLLite model for the Anima model family, designed for anime image repair, tile-based enhancement, and restoration-style image-to-image workflows. The training data includes part of the edited image-pair data from the Noob v2 project.

This model helps Anima restore blurry, low-quality, or degraded anime images while preserving the original composition, character structure, pose, and scene layout as much as possible.

This is not a standalone image generation model. It should be used together with an Anima-compatible checkpoint and inference workflow.

Model Information

What This Model Does

This model is intended to guide Anima during image-to-image repair and restoration tasks.

Typical use cases include:

• Repairing blurry anime images
• Restoring low-quality or degraded anime images
• Reducing compression artifacts and local visual noise
• Enhancing tiled image details
• Improving anime line art and local texture consistency
• Preserving the original composition, pose, and scene layout
• Producing cleaner and sharper anime-style outputs without heavily changing the source image

The v1 release is mainly optimized for blur-damaged images and general low-quality anime restoration tasks.

Gallery / Visual Examples

The following images are sample results and visual examples included in this repository.

_{SAMPLE1.png
SAMPLE2.png
SAMPLE3.png}

Recommended Checkpoint

For regular inference, the recommended Anima checkpoint is:

anima-base-v1.0.safetensors

Earlier checkpoints may be used for comparison, but the final v1 Anima base checkpoint is recommended for normal use.

Usage

There are two main ways to use this model:

1. Python inference with the Anima ControlNet-LLLite script
2. ComfyUI workflow with ControlNet-LLLite_node

Python Inference Example

Example command:

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_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/

Batch Inference Example

You can also use a prompt file for batch inference:

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

Important Parameters

Recommended starting range:

--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 changes fine details too much, lower the multiplier.

ComfyUI Usage

This model can be used in ComfyUI with ControlNet-LLLite_node.

Basic workflow concept:

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

Some testers reported slight color shift when using the ComfyUI node. This issue was not observed in the Python inference path, so it may be related to the node implementation or specific workflow settings.

If you encounter color shift, compare the result with the Python inference path and report the issue together with your workflow settings.

Suggested Prompts

You can start with 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 QQ Groups:

1080876483

531021130

635772191

956810411

519382846

Discord: Laxhar Dream Lab SDXL NOOB https://discord.com/invite/DKnFjKEEvH

Training Information

The v1 model was trained with A100 GPU resources.

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

Limitations

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

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

Roadmap

A v2 version is planned / in training with a larger and more diverse dataset.

Future versions are expected to improve robustness on:

• Blurry images
• Low-resolution images
• Low-quality degraded images
• More diverse tile and repair tasks
• Stronger detail recovery while preserving the original layout

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.

Source

Original model page:

https://civitai.red/models/2708551/anima-tile-and-repair-controlnet-lllite