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

This is the updated V2 release of Anima Tile &amp; Repair ControlNet-LLLite.Compared with V1, this version is trained with a much larger and more diverse dataset, with a stronger focus on anime image upscaling, tile repair, artifact removal, and detail-preserving restoration.V1 was more like an early test version. V2 is a more practical release for daily ComfyUI workflows.What is new in V2Better anime upscalingV2 adds a large number of upscale task image pairs.This means the model can now work better when the input anime image is small, blurry, compressed, or lacking local details. It is especially useful for workflows where you want to enlarge an anime image while keeping the original composition, character structure, line art, and color feeling stable.Typical use cases include:<ul><li>anime image upscaling;</li><li>tiled detail enhancement;</li><li>repairing low-resolution anime images;</li><li>restoring soft or blurry line art;</li><li>improving local texture and fine details without changing the original image too much.</li></ul>Better artifact removal and detail preservationV2 also adds many new training pairs for tasks such as:<ul><li>removing filtering artifacts;</li><li>removing over-sharpening artifacts;</li><li>reducing compression noise;</li><li>repairing damaged details;</li><li>cleaning up unstable or unnatural local textures;</li><li>preserving the original image details during restoration.</li></ul>Because of this, V2 should have much better generalization than V1. It is less likely to destroy the original structure, and it should be better at keeping the original character, pose, layout, and visual identity stable.Much larger training setThe V2 training dataset is about 5× larger than the V1 training dataset.The new data covers more degradation types, more repair situations, more upscale cases, and more real user workflow scenarios. This should make the model more robust across different anime styles and different input quality levels.Community workflow samplesFor this release, I collected many sample results and workflow examples from community users.You can use ComfyUI to read the shared workflows directly. I hope this makes the model easier to test, easier to reproduce, and easier to integrate into your own image restoration / upscale pipeline.Special thanks to the volunteers who contributed samples, workflow feedback, and testing support:<ul><li>poi氵</li><li>小李飞刀</li><li>黑泡芙</li><li>韶韵</li><li>幼刀</li><li>卡柚</li></ul>Thank you very much for helping improve this release.Recommended usageThis model is designed for Anima-compatible tile / repair / upscale workflows.It is not a standalone image generation model. Please load it together with the Anima base model and a compatible ControlNet-LLLite workflow.For most cases, you can start with a moderate ControlNet strength and adjust it based on the result:<ul><li>If the repair effect is too weak, increase the strength slightly.</li><li>If the result becomes too sharp, too constrained, or starts changing fine details too much, lower the strength.</li><li>For upscale workflows, it is recommended to compare different strengths, because different input images may need different levels of guidance.</li></ul>Suggested prompt style:<pre><code>restore this anime image with clean details, sharp line art, natural texture, and stable original composition
</code></pre><pre><code>upscale and repair this anime image, reduce blur and artifacts, preserve the original character structure, pose, colors, and scene layout
</code></pre><pre><code>enhance local anime details, clean up filtering and sharpening artifacts, keep the original image identity unchanged
</code></pre>LimitationsV2 is much stronger than V1, but it is still not the final version.It may still have limitations on extremely damaged images, very unusual styles, or inputs with heavy structural errors. In some workflows, very high strength may still over-sharpen the image or slightly change small details.Please adjust the workflow settings based on your own input image.RoadmapAfter I finish my business trip, I plan to add another dataset that is roughly 2× larger than the V2 dataset, and use it to train the future V3 version.I believe V3 will be a more complete and polished version of this model.Please look forward to it.The business trip is really long...

Files with SHA256: be5a9ead6527329237a66a0eb234ccea4cddb140c1d1e1b7d0776e43f5ff1330