Use MOSS-Image with the smaller Qwen3.5-4B text encoder. This lightweight adapter projects 4B text features into the original 9B conditioning space, helping reduce prompt-encoding memory requirements. Fitted on 1,600 prompts. Highly experimental.

Requirements

• MOSS Smaller Encoder ComfyUI node

• Qwen3.5-4B INT8 ConvRot text encoder

• Optional: Lodestones’ 8-step DMD2 LoRA for faster generation. Place it in ComfyUI/models/loras and load through Load LoRA.

Instructions

1. Clone the custom node into ComfyUI/custom_nodes.

2. Download this adapter and place it in ComfyUI/models/moss_encoder_adapters (create the folder if missing).

3. Place the Qwen3.5-4B encoder in ComfyUI/models/text_encoders.

4. Restart ComfyUI.

5. Select the 4B encoder in Load CLIP, with type set to moss_image.

6. Connect Load CLIP → MOSS Smaller Encoder Adapter → CLIP Text Encode, selecting this adapter in the adapter node (workflow available in images as well)

This is a projection adapter, not a LoRA. Results will differ from the original 9B encoder.

Reuploaded due to a mistake with the uploaded files