2x Faster Qwen-Image-2.1 using 4 Steps LORA — ComfyUI Low-VRAM Workflow
retained text
I’ve just finished testing **Qwen-Image-2.1 in ComfyUI**, using a custom low-VRAM workflow with the **INT8 ConvRot model, ComfyUI Kitchen Attention, and a 4-step LoRA**. One of the biggest improvements I noticed is the generation speed: with the **4-step LoRA, the workflow can be around 2× faster** compared to running the model without it, while still producing usable results. Where Qwen-Image-2.1 really surprised me, though, is **image editing**. The editing results are quite impressive and, in some cases, can **rival Flux 2 Klein**, while Qwen-Image-2.1 is also noticeably faster in my testing. For image editing, it handled things like **style changes, character consistency, clothing changes, and character sheet generation** surprisingly well. However, the results are not as impressive when it comes to pure **text-to-image generation**. I noticed some weaknesses with **skin quality, finger/hand anatomy, fine textures, and text rendering**, so there is definitely room for improvement in these areas. Overall, I think Qwen-Image-2.1 is much more interesting to me as an **image-editing model** than as a pure image-generation model right now. ***Workflow Link*** [https://drive.google.com/file/d/15OgWK1aXBwTTAOhuFKdUQC6vWCQQIjR6/view?usp=sharing](https://drive.google.com/file/d/15OgWK1aXBwTTAOhuFKdUQC6vWCQQIjR6/view?usp=sharing) [https://civitai.com/articles/35828/2x-faster-qwen-image-21-using-4-steps-lora-comfyui-low-vram-workflow](https://civitai.com/articles/35828/2x-faster-qwen-image-21-using-4-steps-lora-comfyui-low-vram-workflow) ***Video tutorial*** [https://youtu.be/JCH3uCil9Nk](https://youtu.be/JCH3uCil9Nk)
document fetched 2026-09-28 00:38 · canonical https://reddit.com/r/comfyui/comments/1wrfhwj/2x_faster_qwenimage21_using_4_steps_lora_comfyui