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.
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.
qwen3.5_9b_*.safetensorstext encoder,moss_image_*.safetensorsdiffusion 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).
| 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).
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.
cd ComfyUI/custom_nodes
git clone https://github.com/ThetaCursed/ComfyUI-MOSSImageTextEncoder.gitRestart ComfyUI — that's it, the node is installed. Then drag workflow.json in and go.
MIT — see LICENSE.