ComfyUI-MOSSImageTextEncoder

Feeds a reference image through the Qwen3.5-9B vision tower and inserts its visual embeddings directly into the text prompt. The image conditions the generation natively through the text encoder.

Note: This node is for vision/conditioning, not image editing.

Workflow

A ready-to-use example workflow is included in the repository: workflow.json.

Simply drag and drop the workflow.json file directly into your ComfyUI workspace to load the entire setup.

Requirements

• qwen3.5_9b_*.safetensors text encoder, moss_image_*.safetensors diffusion model, flux2-vae*.safetensors (for decoding only).
• Load the text encoder through CLIPLoader with type moss_image.
• Latents: use the core Empty Flux 2 Latent node (128 channels).

Text Encode (MOSS-Image Vision)

Input	Type	Notes
clip	CLIP	CLIPLoader → type moss_image.
prompt	STRING	Your text prompt.
negative_prompt	STRING	Optional negative conditioning.
image	IMAGE	Optional reference image. A batch is treated as multiple references.
image_weight	FLOAT	Reference influence applied before the vision tower: 0 ignores it, 1 standard, <1 fades toward gray, >1 pushes away from gray.
vision_megapixels	FLOAT	Max size per reference before the vision tower. Oversized images are downscaled, smaller ones kept native (never upscaled). Default 1.0.
system_prompt	STRING	Optional system instruction prepended to the prompt. Empty matches MOSS-Image's normal conditioning.
print_prompt	BOOLEAN	Dump the assembled Qwen3.5 prompt to the console.

Outputs: positive, negative (CONDITIONING).

Typical wiring

UNETLoader   (moss_image_sft_*)
CLIPLoader   (type: moss_image, qwen3.5_9b_*)
VAELoader    (flux2-vae)                          # decode only
LoadImage ─┐
           ▼
  Text Encode (MOSS-Image Vision) ── positive/negative ─┐
Empty Flux 2 Latent ────────────────────────────────────┤
                                                        ▼
                                    KSampler ─► VAEDecode ─► PreviewImage

Start with image_weight = 1.0, vision_megapixels = 1.0, denoise 1.0.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ThetaCursed/ComfyUI-MOSSImageTextEncoder.git

Restart ComfyUI — that's it, the node is installed. Then drag workflow.json in and go.

License

MIT — see LICENSE.