ComfyUI Text Encode Reference Images

An advanced text encoding node for ComfyUI that supports a dynamic number of reference images.

Mainly for usage with Qwen Image / Flux 2 / Z-Image models.

Nodes

Text Encode Reference Images (v3.0.0)

This node allows you to provide a prompt along with reference images. It supports multiple Vision-Language (VL) encoding strategies, CLIP Vision encoding, and reference latents.

• Category: conditioning
• Features:
  • Native Batch Support: Supports image batches directly. To use multiple images, use the native ComfyUI Batch Images node to combine them.
  • Selectable VL Strategies:
    • none: Standard text encoding without vision-language tokens.
    • qwen image edit: Uses hardcoded system templates for Qwen-style image editing.
    • z-image base omni: Rebase logic from TextEncodeZImageOmni for advanced vision token handling.
  • CLIP Vision: Encodes images using an optional image_encoder (CLIP Vision) for all strategies.
  • Auto-Resize: Scalable image processing that resizes images to ~1MP (~1024x1024) before VAE encoding to maintain efficiency.
  • Reference Latents: Optionally encodes input images using a VAE and appends them as reference_latents to the conditioning data.
• Inputs:
  • clip: The CLIP model to use for encoding.
  • prompt: The text instruction/prompt.
  • vl_selection: Choice of VL strategy (none, qwen image edit, z-image base omni).
  • auto_resize_images: (BOOLEAN) If enabled, resizes images to ~1MP for VAE encoding.
  • vae (Optional): VAE for encoding reference latents.
  • image_encoder (Optional): CLIP Vision model for image encoding.
  • images (Optional): Reference image(s). Use Batch Images to provide multiple.
• Output:
  • CONDITIONING: The encoded conditioning data.

Installation

1. Place this folder in your ComfyUI custom_nodes directory.
2. Restart ComfyUI.
3. The node will appear in the conditioning category.