A little support goes a long way! If you’d like to help me keep creating, you can do so at <a target="_blank" rel="ugc" href="https://ko-fi.com/xabsurd">https://ko-fi.com/xabsurd</a> huggingface:<a target="_blank" rel="ugc" href="https://huggingface.co/Xabsurd/MOSS-Image-Tile-Control">Xabsurd/MOSS-Image-Tile-Control · Hugging Face</a><h1 id="moss-image-tile-control">MOSS-Image Tile Control</h1>Tile control for MOSS-Image, designed to preserve composition, subject outlines, and local structure while allowing the model to regenerate textures and details.Use it for detail enhancement, refinement after upscaling, and image variations guided by an existing image. Outputs are generated rather than copied, so exact pixel reconstruction and recovery of authentic original details are not guaranteed.Two beta versions are available: beta 2 for a simpler setup and beta 3 for finer control over how much the output changes. The author has manually tested a variety of images with positive overall results.<h2 id="which-version-should-i-use">Which version should I use?</h2>Featurebeta 2beta 3Main advantageNo additional custom nodes requiredFiner control over structure preservation and detail regenerationControl methodConcatenated image conditioning with a backbone LoRAIndependent lightweight residual control branchComfyUI supportBuilt-in nodes in a ComfyUI version with MOSS-Image control supportRequires the accompanying <code>ComfyUI-Moss-Tile-Control</code> custom nodesWeight location<code>ComfyUI/models/loras/ComfyUI/models/model_patches/</code>Strength adjustmentControl-image strength and LoRA strength affect the result separatelyOne strength parameter scales the control residuals across layersAt control strength 0The Tile LoRA can still affect the outputTile control is bypassed, preserving the input model and its existing settingsBase model familyMOSS-Image-BaseMOSS-Image-SFTChoose beta 2 for convenience. Start with beta 3 if you want finer adjustment of the modification strength.The weight formats and loading methods are different. Files are not interchangeable, and equal strength values do not imply equal control across the two versions.<h2 id="beta-2:-built-in-nodes-no-extra-installation">beta 2: Built-in nodes, no extra installation</h2>beta 2 uses ComfyUI's existing MOSS-Image control interface. No additional custom nodes are required for this adapter.<h3 id="setup">Setup</h3><ol><li>Use a ComfyUI version that supports MOSS-Image and includes <code>MOSS-Image Color/Depth Control</code>.</li><li>Place the beta 2 weights in <code>ComfyUI/models/loras/</code>.</li><li>Load them with <code>Load LoRA (Model Only)</code> and connect its MODEL output to your sampling chain.</li><li>Add the built-in <code>MOSS-Image Color/Depth Control</code> node.</li><li>Connect your reference image to <code>color_map</code>, leave <code>depth_map</code> empty, and connect the Flux2 VAE used by MOSS-Image.</li><li>Connect the node's positive and negative conditioning outputs to the sampler.</li></ol>For this adapter, <code>color_map</code> accepts the actual reference image. You do not need to convert it into a color-region map or a depth map. Resize the reference image to your intended output dimensions first.<h3 id="strength-behavior">Strength behavior</h3>The control node's <code>strength</code> scales image conditioning. It does not disable the Tile LoRA already applied to the model. Setting control strength to 0 can therefore produce a different result from running without beta 2.For a comparison with beta 2 fully disabled, bypass both its LoRA loader and its control-conditioning node. The LoRA strength slider should not be treated as equivalent to beta 3's residual strength control.beta 2 targets the Base model family. Results on SFT should be evaluated separately.<h2 id="beta-3:-finer-strength-adjustment">beta 3: Finer strength adjustment</h2>beta 3 uses an independent residual control branch. Its <code>strength</code> parameter scales the control residuals across layers, letting you adjust how closely generation follows the reference image.It does not merge Tile weights into the backbone as a LoRA or replace the model's input projection.<h3 id="setup">Setup</h3><ol><li>Install the accompanying <code>ComfyUI-Moss-Tile-Control</code> directory under <code>ComfyUI/custom_nodes/</code>, then restart ComfyUI.</li><li>Place beta 3 weights in <code>ComfyUI/models/model_patches/</code>. Version and epoch subfolders can help keep checkpoints organized.</li><li>Select the weights with <code>Load MOSS Tile Control</code>.</li><li>Add <code>Apply MOSS Tile Control</code> and connect the MOSS-Image MODEL, control weights, Flux2 VAE, and reference image.</li><li>Connect its MODEL output to the sampling chain. Connect positive and negative text conditioning as in a normal MOSS-Image workflow.</li></ol>Load beta 3 with its dedicated nodes. It cannot be applied through a standard LoRA loader or the built-in <code>MOSS-Image Color/Depth Control</code> node. Remove any older beta 2 Tile LoRA from the model chain when using beta 3.<h3 id="parameters">Parameters</h3>ParameterDescription<code>strength</code>Multiplier for the control residuals. Start at <code>0.5</code> and adjust for your desired balance of structure preservation and regeneration<code>start_percent</code>Sampling progress at which control begins. Default: <code>0.0end_percent</code>Sampling progress at which control ends. Default: <code>1.0</code>Try <code>0 / 0.25 / 0.5 / 0.75 / 1.0</code> with a fixed seed. Higher strength generally increases the influence of the reference image, while the final result also depends on the prompt, base model, sampler settings, and other LoRAs.Strength is continuously adjustable, but visual changes are not guaranteed to be linear or strictly monotonic. A strength of 1 applies the control residuals at their nominal scale; it does not lock the output to the input image.At <code>strength = 0</code>, the node returns the input model directly and skips control-image encoding. In a paired real-workflow test with other settings held constant, the zero-strength output was pixel-identical to bypassing the Tile node. Existing style or acceleration LoRAs on the input model remain in effect.<h3 id="epoch-1-or-epoch-2">Epoch 1 or epoch 2?</h3>Both checkpoints provide effective structural control. In a fixed-seed portrait comparison, their results were similar at strength <code>0.5</code>. At strength <code>1.0</code>, epoch 2 preserved some structures and local features more closely.Start with epoch 2, and keep epoch 1 as an alternative. These observations do not establish a universal ranking across all images.<h2 id="testing-and-feedback">Testing and feedback</h2>Beyond the initial portrait comparisons, the author completed another round of manual testing across a variety of images and was satisfied with the overall results.Evaluation currently relies on generated examples and subjective visual assessment. No quantitative benchmark on a standardized test set is provided. The epoch comparison above refers specifically to the controlled portrait test.Feedback from other images and workflows is welcome. Please include the version and epoch, base model, control strength, prompt, seed, sampling settings, and any additional LoRAs so results can be reproduced.<h2 id="compatibility-and-sampling">Compatibility and sampling</h2><ul><li>For beta 3, MOSS-Image-SFT is the recommended base model family.</li><li>Actual generations have been tested locally with SFT INT8 and a 15-step acceleration LoRA. This does not establish compatibility with every quantization or acceleration method.</li><li>When investigating color, texture, or structural issues, compare against a BF16 base model with additional style and acceleration LoRAs disabled, using the base model's recommended sampling settings.</li><li>Different checkpoints, base models, and LoRA combinations may need different control strengths.</li></ul><h2 id="known-limitations">Known limitations</h2><ul><li>Color and tonal shifts can occur. Portrait tests showed warmer skin tones, deeper shadows, and contrast changes. Evaluate the output carefully if preserving the original colors is important.</li><li>Fine details are regenerated. Hair, skin, fine lines, and other textures may change. Generated details should not be treated as accurate recovery of missing source information.</li><li>Results depend on the image and workflow. Manual testing across varied images has been positive, but systematic benchmark results are not available.</li><li>Higher strength does not always mean better quality. Choose a value that balances structural fidelity, natural detail, and the amount of change you want.</li></ul><h2 id="base-models-and-acknowledgments">Base models and acknowledgments</h2><ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base">MOSS-Image-Base</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT">MOSS-Image-SFT</a></li><li><a target="_blank" rel="ugc" href="https://github.com/Comfy-Org/ComfyUI">ComfyUI</a></li></ul>Thanks to the OpenMOSS team and the open-source projects that make this work possible. This is a community-developed Tile control adapter.The official MOSS-Image base models use the <a target="_blank" rel="ugc" href="https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT/blob/main/LICENSE">Apache-2.0 license</a>. Licensing for this project's released weights and accompanying code is specified in their respective license files.

Files with SHA256: 21856898240e6f21e2aee8e856726948cedfde35a8d899c11b05c2105d4af4b9