← sources

Qwen-Image 2.1 prompt enhancer in ComfyUI: 1.4–1.7× faster with MTP, plus uncensored (heretic) versions

source: reddit (reddit:StableDiffusion)published 2026-09-24 16:11ingested 2026-09-25 02:40

retained text

I really like using Generate Text node in ComfyUI for end to end VRAM management by ComfyUI. But using the Qwen-Image 2.1 prompt enhancers from [https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/text\_encoders](https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/text_encoders) in ComfyUI is quite slow.

Above prompt enhancers are a fine-tune of Qwen3.5-9B, but its checkpoints don't include an MTP head. So, I tested if attaching an MTP head to the above models increase the generation speed and was quite happy to find out that it does.

Below are my benchmarks on RTX 4090 Laptop, 16 GB, 1 input image for editing:

|Mode|max\_length|Comfy-Org checkpoint|With MTP head|Speed-up|
|:-|:-|:-|:-|:-|
|Text-to-image|8192|35 tok/s|47 tok/s|1.36×|
|Text-to-image|16256 (official)|27 tok/s|38 tok/s|1.38×|
|Editing|8192|31 tok/s|52 tok/s|1.65×|
|Editing|24000 (official)|21 tok/s|36 tok/s|1.67×|

**Compared with the Comfy-Org checkpoint at the official max\_length, MTP with max\_length 8192 is 1.7× faster for t2i and 2.5× faster for editing**. In my tests MTP didn't change the quality, but the same seed gives different text than the Comfy-Org checkpoint.

Now, the issue was I could not use these models as is in ComfyUI as it required some mtp specific fixes for processing image input for Qwen 3.5. So, I was left with no choice but to make a custom node to use them. Once, it was all setup and everything worked, I was like why not add heretic versions as well.

If someone wants to try it, it's available at: [https://github.com/mozophe/ComfyUI-Qwen-Image-2.1-PromptEnhancer-MTP](https://github.com/mozophe/ComfyUI-Qwen-Image-2.1-PromptEnhancer-MTP)

**Requirements:**

* ComfyUI v0.37.0 or newer
* VRAM: 16 GB recommended. Smaller GPUs work, but much more slowly. Peak use was about 14 GB for t2i and 16 GB for editing
* System RAM: 32 GB recommended
* Disk space: About 10 GB of disk per model (t2i or editing). Heretic versions download about 19 GB in bf16, but the loader converts them to int8 as they download, so they also take about 10 GB

**What the node does:**

* The loader downloads the model you pick and attaches the MTP head the first time you use it
* Uses Qwen's official system prompts and sampling settings
* Heretic versions for both t2i and editing, same speed-up
* The rewritten prompt comes out separately from the reasoning
* Edit with up to 10 input images
* Sample workflows included: drag the PNG/JSON into ComfyUI

document fetched 2026-09-25 00:33 · canonical https://reddit.com/r/StableDiffusion/comments/1wp5vix/qwenimage_21_prompt_enhancer_in_comfyui_1417