Two things this workflow fixes: plastic skin, and the oil-paint filter.
Most 2.5D prompts describe the style with one global adjective. The model then applies that texture to every surface, and the face ends up reading like knitted fabric. This workflow ships the description split by material instead: skin matte with soft gradients, hair as individual strands, wool with a visible weave, ceramic with one hard highlight.
It also ships the sampler settings that let the negative prompt actually do something on Qwen-Image's 4-step Lightning LoRA. At 4 steps / CFG 1.0 the negative is ignored entirely, which is why extra hands keep showing up. At 8 steps / CFG 2.0 with the LoRA at 0.5 the negative works and the illustration look survives.
WHAT IS INSIDE
· workflow.json — drag into the ComfyUI canvas
· style-blocks.txt — the style sentences, including a flat ukiyo-e print variant that runs from the same graph
· four sample outputs
This is the workflow behind every image on this profile. 1024×1408, 8 steps, about 80 seconds on an 8 GB laptop card.