Text Encode (MOSS Image)

ComfyUI Moss Text Encoder

Vision-aware text conditioning for MOSS-Image.

MOSS-Image conditions only through its Qwen3.5-9B text encoder: there is no reference latent, no VAE input and no second conditioning channel. This package exposes that path as a normal text-encode node that can also feed reference images into the encoder, so they can guide the prompt instead of merely being described.

Installation

Clone into ComfyUI/custom_nodes and restart ComfyUI:

cd ComfyUI/custom_nodes
git clone https://github.com/thezveroboy/ComfyUI-MossTextEncoder

No extra Python dependencies are required.

Nodes

Text Encode (MOSS Image)

Category: model/conditioning/moss image.

Inputs:

• prompt / negative_prompt — text (dynamic prompts supported).
• system_prompt (optional, STRING, forceInput) — extra system instruction. Connect MOSS-Image System Prompt or any string node.
• image1 / mask1 (optional) — reference image and its mask. More imageN/maskN pairs appear as slots get connected (up to 16).
• resolution — reference images are resized to about resolution × resolution pixels, preserving aspect ratio and rounding both sides to a multiple of 32. 0 keeps each reference at its own size (also rounded to 32).
• mask_padding — context kept around the mask before cropping, as a fraction of the image size added on each side. 0 = tight crop, 0.1 = about a 10% margin. Ignored when no mask is connected for that image.
• vision_position — before prompt (image tokens, then your text) or after prompt.
• print_prompt — print the assembled system + vision + user prompt to the console.

Outputs: positive, negative (CONDITIONING).

Both outputs use the same reference images and the same chat template; only the text differs. Masks are used per image for cropping, not as a separate conditioning signal.

MOSS-Image System Prompt

Emits an experimental system instruction that tells the model how to read reference images and merge them with the user prompt. MOSS-Image has no trained default system prompt, so treat this as an optional experiment: clear the text field (or do not connect the node) to send no system message.

Notes

• The text encoder is selected in CLIPLoader with type moss_image.
• Vision prompting needs a non-FP8 text encoder. A FP8 encoder cannot run the vision tower and raises add_stub not implemented for Float8_e4m3fn; load bf16/fp16 (or the supported int8 checkpoint). Text-only prompts do work with FP8.
• Prompts describing a reference image work best as structured JSON, for example:

{
  "summary": {"text": "Studio portrait on a dark background"},
  "items": [
    {"label": "subject", "caption": "woman", "text": "short dark hair, neutral expression", "bbox": [120, 80, 640, 900]}
  ]
}

License

MIT