Qwen-Image renders legible, correctly spelled text as part of the scene. Not a caption overlay and not a watermark — the letters are lit by the same light as everything else, they sit on a surface, they get wet, they peel.
Three samples here: a Chinese neon sign, an English print poster, a carved wooden plaque. Different carriers, different materials, same workflow.
WHAT MAKES THE TEXT LAND
· Quote the exact string in curly quotes instead of describing the letters: “深夜食堂”, “STAY CURIOUS”.
· Give the text a physical carrier — neon, print, carved wood, enamel. Text floating in air is where it falls apart.
· Give that carrier a condition: weathered, rain-wetted, sun-bleached, peeling at one corner. The imperfection is what sells it.
· Keep the string short. Four to twelve characters is reliable; paragraphs are not, in any language.
· Never ask for a font by name. Ask for the material instead.
25 steps / CFG 1.0 / euler / simple at 1024×1024, about 80 seconds on an 8 GB card. Note that CFG 1.0 means the negative prompt is ignored — fine for signage, but if you put people in frame, raise CFG to 2.0 and use the LoRA route on the 2.5D profile.
Description
Nothing is trained. A workflow plus the prompting rules that make text render legibly inside the scene.
Drag workflow.json into ComfyUI; models and settings are in README.md inside the archive.

