CivArchive
    Moss-Image Tile Control - workflow and node
    NSFW
    Preview 148730284
    Preview 148730286

    If this project helps you, feedback and example images are welcome!

    Support my work on Ko-fi

    MOSS-Image Tile Control

    Tile control for MOSS-Image, with three released options for different priorities: beta2 for maximum preservation, beta3 for speed, and V1 for detail refinement and overall quality.

    Use an existing image to guide composition, subject outlines, and local structure during generation. Depending on the version, you can prioritize minimal changes to the source or give the model more room to refine textures and details after upscaling.

    V1 is my recommended final version, developed through four design iterations. In my comparisons, it delivers the best clarity and detail enhancement of the released versions, with good feature and color preservation. beta2 and beta3 remain available for users who prefer their particular trade-offs.

    Hugging Face: downloads, documentation, examples, and workflows

    Color preservation update

    If preserving the reference image's colors matters to you, updating to ComfyUI-MOSS-LLLite v0.1.3 or later is recommended. The update adds the optional experimental Apply MOSS Color Guide node. Existing model weights remain compatible; there is no need to download V1 again.

    Which version should you use?

    • V1 — recommended: The best balance in my tests for clarity, detail refinement, subject features, and color retention. Uses MOSS-LLLite with MOSS-Image-SFT.

    • beta2 — maximum preservation: Suited to high control strength and minimal changes. It preserves original features most strongly and behaves more like a conservative upscaler, rather than repairing fine details. It is the slowest option in my workflows. Uses a LoRA with MOSS-Image-Base and built-in ComfyUI control nodes.

    • beta3 — speed first: The fastest option in my workflows, with strong structural control, but it tends to smooth high-frequency textures and fine details. Currently not my general recommendation for restoration.

    These are observations from my own workflows, not standardized benchmark rankings. Equal strength values do not mean equal control across versions.

    beta2: strong preservation, no extra custom nodes

    Choose beta2 when you want high control strength and the smallest changes to the original image. It preserves input features most strongly, behaving more like a conservative upscaler. It does not repair fine details and is the slowest option in my tested workflows.

    Base model: MOSS-Image-Base. Format: LoRA.

    1. Put the beta2 weights in ComfyUI/models/loras/.

    2. Load them with Load LoRA (Model Only).

    3. Add the built-in MOSS-Image Color/Depth Control node. Connect the actual reference image to color_map, leave depth_map empty, and connect the Flux2 VAE. No color-region or depth preprocessing is required for this Tile adapter.

    4. Route both positive and negative conditioning through the control node to the sampler. Match the reference image and output dimensions.

    Control-image strength and LoRA strength are separate. Setting control strength to 0 does not remove the LoRA's changes to the backbone; bypass both the LoRA and control node for a fully disabled comparison.

    beta3: speed first

    beta3 provides strong structural control and is the fastest option in my workflows. However, it tends to smooth high-frequency textures and fine details, so it is currently not my general recommendation for restoration.

    Base model: MOSS-Image-SFT. Format: independent residual adapter.

    1. Install the current custom node package under ComfyUI/custom_nodes/ and restart ComfyUI.

    2. Put beta3 weights in ComfyUI/models/model_patches/.

    3. Use Load MOSS Tile Control and Apply MOSS Tile Control, included in the package. Connect the model, controller, Flux2 VAE, and reference image, then send the model output to the sampler.

    Start around strength 0.5 and adjust. At strength 0, the node returns the input model unchanged. beta3 uses its residual nodes, not the MOSS-LLLite nodes or a standard LoRA loader.

    V1 is my personal choice for overall quality. It lets the model refine details while preserving subject features and overall structure, with good color retention in my visual tests.

    Base model: MOSS-Image-SFT. Format: standalone MOSS-LLLite adapter.

    V1 is a standalone control adapter. Load it with the MOSS-LLLite nodes, not a standard LoRA loader.

    1. Download the custom node ZIP and extract it into ComfyUI/custom_nodes/.

    2. Put MOSS-LLLite-Tile-V1.safetensors in ComfyUI/models/model_patches/.

    3. Restart ComfyUI. Add Load MOSS-LLLite and Apply MOSS-LLLite.

    4. Connect your MOSS-Image-SFT model, loaded controller, and reference image to the apply node. Send its model output to the sampler. The apply node takes RGB images directly and has no VAE input.

    If you have the old ComfyUI-Moss-Tile-Control node directory installed, remove it before installing the updated package to avoid duplicate nodes. The new package also supports existing beta3 workflows.

    Start with strength 0.5–1.0, start 0, and end 1. Adjust strength to balance preservation and refinement. At strength 0, V1 is bypassed and the input model is returned unchanged; other LoRAs already loaded on that model remain active.

    Examples and workflows for each version

    Follow the instructions for the selected version. The three weight formats and loading nodes are not interchangeable. Remove the beta2 Tile LoRA when switching to beta3 or V1.

    The published V1 examples use no Color Match or sharpening. Their workflows include an acceleration LoRA and the exact sampling settings used for those images.

    Color and detail preservation

    Color, saturation, and contrast can change during diffusion sampling. V1 retains colors well in my tests, but exact matching is not guaranteed. Color Match can be used with any version after generation, using the original image as the color reference.

    Fine details are generated and may differ from the source; they should not be treated as exact recovery of missing information. Results depend on the image, prompt, sampling settings, and additional LoRAs.

    Description

    FAQ

    Controlnet
    MOSS Image

    Details

    Downloads
    37
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/12/2026
    Updated
    11/22/2026
    Deleted
    -

    Files

    workflow and node.zip

    Mirrors

    CivitAI (1 mirrors)