Note: update your ComfyUI to the latest nightly release! The 'Load CLIP' node must be updated to include the 'moss_image' type.

V1.1 Update: Small update - cleaned up node placement and added a SkipLayerGuidanceDiT node in accordance with OpenMOSS recommendations. This node skips Layer 22, although OpenMOSS also recommends adding Layer 24 and 26 and experimenting. There definitely seems to be an improvement in output quality and reduction in body horror after testing dozens of renders.

Hints for MOSS Prompting:

• MOSS freaks out if there is a mismatch between the canvas shape and the subject's posture. For example, if you prompt a person "lying flat horizontally" on a narrow, tall portrait canvas, the model will drastically contort or stretch the limbs to force them into frame. Always use wide aspect ratios for horizontal compositions and vertical aspect ratios for standing portraits.

• If you find the generation looks sloppy or generic, switch from the Base model to the SFT (Superfine-Tuned) version. The SFT version handles dense, long text instructions up to 4K tokens much more gracefully.

• MOSS is very good at handling text - this is one of the model's main 'selling points'. Always wrap the exact words you want rendered inside straight double quotes (e.g., a t-shirt with the text "MOSS" or a sign reading "Open")

• If you choose to use Booru-style tag sequences instead of natural language, follow this specific hierarchical order for optimal decoding: Quality → Score → Meta → Year → Character → Series → Artist → General.

• MOSS-Image is highly sensitive to random seeds. Short prompts tend to vary wildly from seed to seed, whereas longer, highly descriptive prompts make the model consistent and predictable.

__________________________________________________________________________________________________

This is a simple T2I workflow for the MOSS model by OpenMOSS. By default, this workflow is designed to use the SFT INT8 quantization with the DMD 8-step turbo LoRA. However, it can be easily reconfigured to run the full-weight model or any turbo model without a turbo LoRA. This workflow uses only default ComfyUI nodes and the Lora Manager node - nothing fancy here! This is meant to get you up and running with the minimum essentials: a positive prompt, a negative prompt, and a selection of LoRAs. I can generate a 2.6MP (1312x1984) image in under 10 seconds with a 5070Ti. This model absolutely rips while maintaining excellent quality.

I'm looking at adding some additional features in future releases, such as detailers, upscaling, and I2I or BBOX editing.