Qwen 2.1 in ComfyUI: low-VRAM setup and editing tests

This tutorial covers a ComfyUI workflow for Qwen 2.1, a 7-billion-parameter model for text-to-image generation and image editing. The presenter uses INT8 weights, ComfyUI kitchen attention and a Viggle turbo LoRA to reduce memory use and inference time for local AI image work.

Setup starts with updating ComfyUI to the version identified in the tutorial as 0.37, then downloading the diffusion model, text encoder and VAE into their designated folders. The presenter describes the INT8 model as 7.2 GB. Bypass controls switch between generation and editing. The workflow also supports transparent output through an alpha-aware save node and prompts that request an alpha channel.

The walkthrough explains resolution selection, GPU caching and sampler settings. With the turbo LoRA, the presenter reduces sampling from 25 steps to four and reports roughly 30 seconds per generated image on an RTX 3060 with 6 GB of VRAM. These are reported test results, not a hardware guarantee.

Editing examples cover style changes, object removal, pose transfer and character sheets. The presenter finds character consistency useful but notes removal artifacts and neck defects in head transfers. Generation tests show uneven text rendering, artificial skin texture and finger errors. For character sheets, the tutorial recommends selecting a custom output resolution rather than inheriting the input image size.