Qwen Image 2.1: ComfyUI character sheets on 8GB VRAM

This tutorial covers Qwen Image 2.1 image generation and reference-based character sheets in ComfyUI. The presenter tests an RTX 5060 Laptop GPU with 8GB VRAM and 32GB system RAM. The setup uses ComfyUI Nightly, Python 3.12 and CUDA 13.2. The creator identifies the four-step adapter as Viggle Turbo v0.1; these tests do not cover the v0.2 preview.

For text-to-image generation, the presenter compares two, four and eight megapixels with a fixed seed. Four-step outputs show anatomy errors, unfinished backgrounds and overly smooth skin. Bypassing the LoRA and using 25 steps improves the demonstrated full HD results, with runs taking about 71 to 77 seconds. Increasing resolution alone does not resolve the defects in these tests.

The character-sheet workflow uses one reference image and requests an outfit change, four full-body angles and a close-up portrait. The presenter finds four steps usable at two and four megapixels. An eight-megapixel attempt produces an extra view and overlapping panels; the four-megapixel retest restores the requested layout at 2752 by 1536 pixels. The 25-step edit still leaves an outfit mismatch in one view.

This local AI workflow also has a licensing constraint: the creator notes that the Qwen Research License limits use to research or evaluation unless a separate commercial license is obtained.