Xabsurd's picture
Release V1 with updated nodes, comparisons and workflows
08d4d84 verified
|
Raw History Blame Contribute Delete
3.23 kB

V1 examples and workflows

All four examples use the released V1 weights. Outputs retain the original generated pixels. Their PNG metadata embeds a cleaned ComfyUI workflow using the public V1 filename. No Color Match or sharpening was applied.

Example Original input Upscaled control preview V1 output UI workflow API prompt
Flying cat PNG PNG PNG JSON JSON
Grass photograph PNG PNG PNG JSON JSON
Jetpack cat PNG PNG PNG JSON JSON
Cat photograph PNG PNG PNG JSON JSON

Run an example

  1. Install the accompanying ComfyUI-MOSS-LLLite nodes and restart ComfyUI.
  2. Copy MOSS-LLLite-Tile-V1.safetensors to ComfyUI/models/model_patches/.
  3. Copy the chosen *-input.png into ComfyUI/input/. Do not use the control preview as the source: the workflow already performs the recorded resizing.
  4. Drag the corresponding *-v1.png or *.workflow.json onto the ComfyUI canvas.
  5. Select the installed files in the model loaders. These examples use:
    • moss_image_sft_int8_convrot.safetensors — diffusion model;
    • qwen3.5_9b_bf16.safetensors — text encoder, type moss_image;
    • flux2-vae.safetensors — VAE;
    • Coffmeteor-sft-15step-cfg3.safetensors — acceleration LoRA, strength 1;
    • MOSS-LLLite-Tile-V1.safetensors — controller.
  6. Queue the workflow. The additional models above are not bundled with this repository. Replacing them or changing prompts/settings can change the result.

Only core ComfyUI nodes and MOSS-LLLite nodes are used. Inactive captioning, comparison, color-matching, and unrelated control nodes have been removed from the shared graphs. *.api.json files are for the ComfyUI API, not canvas import.

Recorded settings

  • Euler sampler, SGM uniform schedule, 15 steps, CFG 5, denoise 1, shift 3.
  • Seed 994529790995974; V1 strength 1, start 0, end 1.
  • Flying cat and jetpack cat: Lanczos enlargement by 4.
  • Grass and cat photographs: Lanczos reduction to 0.1, followed by enlargement by 5. These are deliberately degraded restoration tests, not enlargement of the full-resolution photograph. Original input pixels are included as PNGs.
  • VAE dimensions align to a 16-pixel grid; control previews are aligned to the resulting output size using the node's resize method.

The control preview reconstructs preprocessing for side-by-side viewing; it is not an additional model output. The V1 PNGs contain the actual generated pixels. Source-image EXIF and unrelated workflow metadata are omitted.

See examples.json for exact dimensions, parameters, and output pixel hashes. Results are visual examples rather than a standardized benchmark.