--- language: - en tags: - moss-image - tile - controlnet - lllite - image-to-image - image-restoration - comfyui - adapter base_model: - OpenMOSS-Team/MOSS-Image-Base - OpenMOSS-Team/MOSS-Image-SFT base_model_relation: adapter library_name: diffusion-single-file --- # MOSS-Image Tile Control — V1 **V1 is the recommended final release**, developed through four design iterations. It uses MOSS-LLLite to refine image details while preserving subject features and overall structure, with good color retention in the author's visual tests. **V1 provides the best clarity and detail enhancement among these releases in the author's comparisons, making it the author's personal choice for overall quality.** Use it for refinement after upscaling and image restoration guided by an existing image. Generated details are plausible reconstructions, not guaranteed recovery of the original information. **[GitHub: ComfyUI-MOSS-LLLite custom nodes](https://github.com/X-absurd/ComfyUI-MOSS-LLLite)** If you would like to support this project: [Ko-fi](https://ko-fi.com/xabsurd). ## Which version should I use? | Version | Best use | Observed behavior | Relative sampling speed | Setup | | --- | --- | --- | --- | --- | | **V1** | **Recommended for detail refinement and restoration** | Best clarity and detail enhancement in the author's comparisons, with subject-feature preservation and good color retention | Between beta2 and beta3 in the author's workflows | MOSS-LLLite custom nodes; MOSS-Image-SFT | | **beta2 — LoRA** | High control strength and minimal changes | Strongest preservation of original features; changes the image least. Preserves existing detail rather than repairing it | Slowest | Built-in control node and LoRA loader; MOSS-Image-Base | | **beta3 — residual adapter** | Speed first | Strong structural control, but tends to smooth away high-frequency detail; currently not generally recommended | Fastest | Legacy residual nodes included in this package; MOSS-Image-SFT | These comparisons reflect the author's visual evaluation and workflows, not a standardized benchmark. The versions use different control paths and Base/SFT families; the same numeric strength does not imply the same control intensity. ## Downloads | File | Load with | | --- | --- | | [V1/MOSS-LLLite-Tile-V1.safetensors](V1/MOSS-LLLite-Tile-V1.safetensors) | **Load MOSS-LLLite** | | [beta2/epoch1/adapter_model.safetensors](beta2/epoch1/adapter_model.safetensors) | **Load LoRA (Model Only)** | | [beta3/epoch1/adapter_model.safetensors](beta3/epoch1/adapter_model.safetensors) | **Load MOSS Tile Control** | | [beta3/epoch2/adapter_model.safetensors](beta3/epoch2/adapter_model.safetensors) | **Load MOSS Tile Control** | | [ComfyUI-MOSS-LLLite.zip](ComfyUI-MOSS-LLLite.zip) | Custom node installation archive | The beta2 and beta3 files are retained unchanged for existing users. V1 is the final design established through four iterations of the control approach. ## V1: installation and usage Use a recent ComfyUI with native MOSS-Image, Anima LLLite, and the V3 node API. No extra Python packages are required by these custom nodes. 1. Extract [ComfyUI-MOSS-LLLite.zip](ComfyUI-MOSS-LLLite.zip) into `ComfyUI/custom_nodes/`, producing `custom_nodes/ComfyUI-MOSS-LLLite/`. Alternatively, copy the included [ComfyUI-MOSS-LLLite](ComfyUI-MOSS-LLLite/) directory there, or install from [GitHub](https://github.com/X-absurd/ComfyUI-MOSS-LLLite). 2. If the old `ComfyUI-Moss-Tile-Control` custom node directory is installed, remove it before installing this package to avoid duplicate node IDs. The updated package includes the beta3 nodes and preserves existing workflow IDs. 3. Place `MOSS-LLLite-Tile-V1.safetensors` in `ComfyUI/models/model_patches/`. 4. Restart ComfyUI. Add **Load MOSS-LLLite** and **Apply MOSS-LLLite** from the **MOSS-LLLite** category. 5. Select V1 in the loader. Connect the MOSS-Image-SFT `MODEL`, controller, and reference `IMAGE` to the apply node. Connect its output `MODEL` to the sampler. 6. Keep normal text conditioning and VAE connections. The apply node accepts RGB images directly and has **no VAE input**. Resize the reference image to the intended output dimensions before applying control. | Parameter | Suggested starting point | | --- | --- | | `strength` | `0.5`–`1.0`; adjust for the desired preservation/refinement balance | | `start_percent` | `0.0` | | `end_percent` | `1.0` | Strength is continuously adjustable. Higher values generally increase the image's influence, but visual changes need not be linear or strictly monotonic. At `strength = 0`, the apply node returns the input model directly, without encoding the control image or adding control patches. Other LoRAs already applied to the input model remain active. **Load MOSS-LLLite** and **Apply MOSS-LLLite** were previously displayed with a `Tile` suffix. Their internal IDs are unchanged. The nodes are not task-locked to Tile, but other control tasks require appropriately trained MOSS-LLLite weights; V1 itself is a Tile checkpoint. ## beta2: maximum preservation Choose beta2 when high control strength and minimal modification matter more than detail repair. In the author's tests, it preserves original features most strongly and changes the image least, but does not repair fine details. Its behavior is closer to a conservative upscaler: strong input preservation leaves less room for generative detail refinement. It is also the slowest option in the tested workflows. 1. Put beta2 weights in `ComfyUI/models/loras/` and use **Load LoRA (Model Only)**. 2. Use the built-in **MOSS-Image Color/Depth Control** node. Connect the actual reference image to `color_map`, leave `depth_map` empty, and use the Flux2 VAE. No color-region or depth preprocessing is needed for this Tile adapter. 3. Route both positive and negative conditioning through that node to the sampler. Use MOSS-Image-Base and match the reference and output dimensions. No additional custom nodes are needed. Control-image strength and LoRA strength are separate: setting control strength to zero does **not** remove the LoRA's effect on the backbone. Bypass both for a fully disabled beta2 comparison. ## beta3: speed first, legacy option beta3 is the fastest option in the author's workflows and provides strong structural control. It tends to smooth high-frequency textures and fine details, so V1 is recommended for general restoration. Put beta3 weights in `ComfyUI/models/model_patches/`. Use **Load MOSS Tile Control** and **Apply MOSS Tile Control**, included in the current package. Connect the MOSS-Image-SFT `MODEL`, controller, Flux2 `VAE`, and reference `IMAGE`. At strength zero, control is bypassed. beta3 weights cannot be loaded with the LLLite nodes. ## V1 examples and comparisons These four examples use the released V1 weights. The left column is the workflow's upscaled control image, aligned to the output dimensions; the right column is the actual V1 output. **Color Match and sharpening were not applied.** Images below link to their full-resolution files. | Example | Input control | V1 output | Workflow | | --- | --- | --- | --- | | Flying cat | ![Flying cat input](examples/v1/flying-cat-control.png) | ![Flying cat V1](examples/v1/flying-cat-v1.png) | [JSON](examples/v1/flying-cat.workflow.json) | | Grass photograph | ![Grass photograph input](examples/v1/grass-control.png) | ![Grass photograph V1](examples/v1/grass-v1.png) | [JSON](examples/v1/grass.workflow.json) | | Jetpack cat | ![Jetpack cat input](examples/v1/jetpack-cat-control.png) | ![Jetpack cat V1](examples/v1/jetpack-cat-v1.png) | [JSON](examples/v1/jetpack-cat.workflow.json) | | Cat photograph | ![Cat photograph input](examples/v1/cat-photo-control.png) | ![Cat photograph V1](examples/v1/cat-photo-v1.png) | [JSON](examples/v1/cat-photo.workflow.json) | Each output PNG also embeds its cleaned ComfyUI workflow: drag it into ComfyUI to load the graph. Copy the matching `*-input.png` from [examples/v1](examples/v1/) into `ComfyUI/input/`, install V1, then select your installed backbone, text encoder, VAE, and acceleration LoRA in the loaders. The separate `*.workflow.json` files are UI workflows; `*.api.json` files are API prompts. The recorded setup uses **MOSS-Image-SFT INT8/ConvRot**, the **Qwen3.5-9B BF16** text encoder, **Flux2 VAE**, and **Coffmeteor-sft-15step-cfg3** acceleration LoRA at strength `1.0`. Sampling: **15 steps, CFG 5, Euler / SGM uniform, denoise 1, shift 3, seed 994529790995974**. V1 strength is `1.0`, active from `0` to `1`. The acceleration LoRA is separate from V1 and is not included in this repository. The flying-cat and jetpack-cat inputs are enlarged 4× with Lanczos. The two photographs are deliberately reduced to 10% and then enlarged 5× to test restoration; their original high-resolution sources are included for reproducing that preprocessing. Small border adjustments follow the model's 16-pixel grid. These are prompt-guided restoration examples, not a standardized benchmark or proof that all generated details match the original scene. See [example instructions](examples/v1/README.md) and [exact settings](examples/v1/examples.json) for the files and preprocessing. ## Legacy beta2/beta3 examples and workflows The following examples are **legacy beta2/beta3 comparisons**, not V1 results. Each comparison shows the input on the left and output on the right. | beta3 · epoch2 | beta2 | | --- | --- | | ![beta3 comparison](examples/beta3-epoch2-comparison.png) | ![beta2 comparison](examples/beta2-comparison.png) | | ![beta3 comparison 2](examples/beta3-epoch2-comparison2.png) | ![beta2 comparison 2](examples/beta2-comparison2.png) | Existing [workflows](workflow/) are retained for beta2 and beta3. Follow the V1 instructions above or load one of the V1 example workflows; the old graphs use different load/apply nodes. ## Compatibility and limitations - V1 uses MOSS-Image-SFT. MOSS-Edit reference-image and cached-reference-KV paths are currently unsupported by the LLLite nodes. - V1, beta3, beta2, and Anima LLLite weight files are not interchangeable. Remove old beta2 concat Tile LoRAs when switching to V1 or beta3. - V1 retains colors well in the author's comparisons, but color, contrast, or saturation shifts can still occur. **Color Match** can optionally correct the output using the original image as the color reference. - Fine detail may be invented or changed. Evaluate results when exact identity, text, patterns, or faithful document reconstruction matters. - The prompt, base model, sampling settings, quantization, and other LoRAs affect results. Compare versions with the same image and controlled settings where possible; higher strength is not automatically better. ## Development notes and credits See [RESEARCH_REPORT.md](RESEARCH_REPORT.md) for a concise development report in English, including the architectural choices, color investigation, and checkpoint selection. [SHA256SUMS.txt](SHA256SUMS.txt) records hashes of the distributed weights and node ZIP. Built on [MOSS-Image](https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT), [ComfyUI](https://github.com/Comfy-Org/ComfyUI), and ComfyUI's native Anima LLLite modules. This is a community-developed control adapter. Inference code attribution and GPL-3.0 terms are included in [NOTICE](ComfyUI-MOSS-LLLite/NOTICE) and [LICENSE](ComfyUI-MOSS-LLLite/LICENSE); model weights have their own applicable terms.