ComfyUI node rewrites prompts with the encoder's own LLM — 17.4s vs 24.8s in img2img

Stock_Mycologist1104 · reddit · 2026-11-20

A developer released the open-source ComfyUI node SingleEncoderEnhancer: it rewrites your instruction using the text encoder's own language model, then encodes the result with the same model — no second LLM needed, the model loads once, and each reference image passes the vision tower once for both rewrite and encode.

Benchmarks (RTX 5060 Ti 16GB, Qwen-Image-2.1): no node 20.8s, Generate Text + Text Encode 24.8s, this node 17.4s. The gain is mainly in image-to-image mode; text-to-image gains are minimal. Built for Qwen3-VL encoder, works with any CLIP ComfyUI's Generate Text node supports; images only (up to 16 refs), no video/audio. Published on ComfyUI Manager and GitHub.

Original post →

More from Multimodal

Multimodal channel →