Qwen-Image 2.1 edits without loading a text encoder in ComfyUI.
If you already keep a large language model running on your machine, this workflow uses that same model as the text encoder for Qwen-Image 2.1. ComfyUI never loads Qwen-Image 2.1's own 8B text encoder, which takes 6 to 10 GB depending on the file. That memory stays free for the image model, and nothing is swapped in and out between a chat and a render.
The same model also rewrites your short request into a full edit prompt before it is encoded, at no extra memory cost.
WHY
One language model in memory, not two. The model you chat with is the encoder. ComfyUI loads and unloads no text encoder, so the image model stays in VRAM between runs. A small adapter (269 MB) maps the language model's hidden states into what Qwen-Image 2.1 expects.
WHAT IT DOES
You load a picture and write what you want in a few words. The Prompt Enhancer looks at the picture and writes the edit prompt: what changes, what stays, and any quoted text exactly as you wrote it. It answers in the language of your request. The Universal Conditioner encodes that prompt and the picture for Qwen-Image 2.1. It runs 8 steps at cfg 1 with the turbo model.
WHAT YOU NEED
Node pack: ComfyUI-UniversalConditioner, at https://github.com/igorls/ComfyUI-UniversalConditioner
Adapter: unicond-qi21-v3.2-seq.safetensors, placed in ComfyUI/models/universal_conditioner/, from https://huggingface.co/igorls/universal-conditioner
Image model: Qwen-Image 2.1 and its VAE, as in any Qwen-Image 2.1 workflow. Keep the Qwen-Image 2.1 text encoder file (qwen3vl_8b_int8_convrot.safetensors) in models/text_encoders/. It is read from disk for its token table and is not loaded onto the GPU.
Server: one running Qwen3.8-27B that answers /v1/hidden_states and /v1/chat/completions. This workflow was built on NInfer Workstation, at https://github.com/igorls/ninfer. The hidden-states endpoint is in release 2026.11.22-v3port.25 and later.
Server model: recommended https://huggingface.co/neroued/Qwen3.8-27B-nvfp4-NInfer
HOW TO USE IT
Set your server's address on both the Universal Conditioner and the Prompt Enhancer nodes. Load your picture. Write your request, with any text that must appear in the picture in double quotes. Run. The Enhanced prompt node shows what was sent to the model. Turn enhance off to use your own words. Change the enhancer's seed for another wording. For a second picture, connect it to image_2 on both the enhancer and the text encode node, in the same order.
HOW IT COMPARES WITH THE NATIVE ENCODER
On 14 same-seed pairs (edits, signage and posters) the adapter and Qwen-Image 2.1's own text encoder gave near-identical pictures, and every quoted text was spelled correctly with the adapter. Small details differ between the two at a fixed seed. The adapter card on Hugging Face has the numbers.
LIMITS
This saves memory only if the language model is already running. It is for people who keep a large model resident; it is not a way to run Qwen-Image 2.1 on a smaller card. The enhancer can miss a detail you care about, so read the enhanced prompt and name that detail in your request if it is missing. Tested with Latin and Japanese text and up to two pictures.
LICENSE
The node pack is MIT. The adapter follows Qwen-Image 2.1's terms: Qwen Research License, non-commercial use only. Built with Qwen.
Description
First release. Tested with the Qwen-Image 2.1 adapter v3.2 (unicond-qi21-v3.2-seq.safetensors) and NInfer Workstation 2026.11.22-v3port.25.




