This model is a Tile &amp; 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.<h2 id="what-this-model-does">What this model does</h2>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:<ul><li>repairing blurry anime images;</li><li>restoring low-quality or low-detail images;</li><li>reducing visible noise and compression artifacts;</li><li>improving local details in tile / repair workflows;</li><li>preserving the original composition while making the final image cleaner and sharper.</li></ul>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.<h2 id="recommended-checkpoint">Recommended checkpoint</h2>Use:<pre><code>anima-base-v1.0.safetensors
</code></pre>Earlier checkpoints are also provided for comparison, but the final v1 checkpoint is recommended for normal use.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e25fd9d4-4ec2-45a5-9525-4f88aa81664b/original=true/e25fd9d4-4ec2-45a5-9525-4f88aa81664b.jpeg" /><h2 id="how-to-use">How to use</h2>There are two main ways to use this model:<ol><li>Python inference with the Anima ControlNet-LLLite script</li><li>ComfyUI workflow through the <code>ControlNet-LLLite_node</code></li></ol><h2 id="python-inference-usage">Python inference usage</h2>You can use the provided Anima ControlNet-LLLite inference script:<pre><code>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/
</code></pre>For batch inference, you can use a prompt file:<pre><code>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/
</code></pre>Example <code>prompts.txt</code> line:<pre><code>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
</code></pre>Useful parameters:<pre><code>--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
</code></pre>A good starting point is:<pre><code>--lllite_multiplier 0.8 ~ 1.0
</code></pre>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.<h2 id="comfyui-usage">ComfyUI usage</h2>You can also use this model in ComfyUI with the ControlNet-LLLite_node workflow.Basic idea:<pre><code>Anima base model
+ ControlNet-LLLite_node
+ anima_tiled_lllite_v1.safetensors
+ input/control image
= repaired Anima output
</code></pre>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.<h2 id="suggested-prompts">Suggested prompts</h2>For repair / tile workflows, you can try prompts like:<pre><code>restore this anime image with clean details, sharp line art, and natural texture
</code></pre><pre><code>repair the low-quality anime image, reduce blur and compression artifacts, preserve the original composition
</code></pre><pre><code>enhance the image details, clean up artifacts, keep the character structure and scene layout unchanged
</code></pre><pre><code>restore fine anime line art and local details while keeping the original pose, composition, and colors stable
</code></pre><h2 id="communication">Communication</h2><ul><li>QQ Groups:<ul><li>1080876483</li><li>531021130</li><li>635772191</li><li>956810411</li><li>519382846</li></ul></li><li>Discord: <a target="_blank" rel="ugc" href="https://discord.com/invite/DKnFjKEEvH">Laxhar Dream Lab SDXL NOOB</a></li></ul><h2 id="training-information">Training information</h2>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.<h2 id="limitations">Limitations</h2>This is a v1 release, so there are still limitations:<ul><li>It is mainly designed for anime images.</li><li>It may not work well on realistic photos.</li><li>Very strong guidance may over-sharpen details.</li><li>Some heavily degraded images may still require stronger restoration or a future version.</li><li>ComfyUI node output may show a slight color shift in some workflows.</li></ul><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/f0533d30-abf2-4133-a18b-c5b54435a318/original=true/f0533d30-abf2-4133-a18b-c5b54435a318.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/37b05d7d-8caa-42bd-bf9b-d2a1e918f787/original=true/37b05d7d-8caa-42bd-bf9b-d2a1e918f787.jpeg" /> NODE INPUT&amp;OUTPUT<h2 id="roadmap">Roadmap</h2>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.<h2 id="credits">Credits</h2>Special thanks to <a target="_blank" rel="ugc" href="http://Comfy.org">Comfy.org</a> for providing GPU sponsorship.Thanks also to the volunteers who contributed testing and feedback:<ul><li>Yidhar</li><li>GHOSTLXH</li><li>年糕特工队</li><li>轻松</li><li>Free Will</li></ul>Their feedback helped improve the training and release process.<h2 id="license-usage">License / Usage</h2>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.

Anima Tile & Repair ControlNet-LLLite

V1 is a test training version, while V2 will greatly expand both the types and the amount of training tasks

Files with SHA256: ed136d14594bd34c7968597a5d1846fb1cdcd0cfcbcaf8988ea0bc73e4865aa0