Qwen image model's prompt enhancer secretly swaps transparency for white backgrounds
Bill_Evans777 · reddit · 2026-10-20
A Reddit user found that the default prompt-enhancer model in the official Qwen image template (qwen3.59bqwenimage2.1pei2i.int8convrot and variants) carries an explicit instruction to "silently" rewrite transparency requests into a solid white background.
- Workarounds: skip the enhancer and write prompts manually, or "jailbreak" it with instructions forcing literal transparency.
- Still unreliable: even with explicit asks, results often come back solid white (not a checkerboard misreading of alpha). The most effective phrasing so far: "Remove the solid color background and the white outline stroke surrounding the character entirely and replace it with alpha channel, leaving her isolated on a fully transparent background" — though it still fails depending on the rest of the prompt.
- The poster is soliciting more reliable transparency prompting techniques.
More from Multimodal
- MOSS-Image finetune Kromoss trends on Hugging Face under Apache-2.0 — lodestones · 2026-10-20
- LMArena Benchmarks Meta's Muse Wonder Against Claude, GPT and Gemini Flagships — arena · 2026-10-20
- AI music generation shows dramatically improved beat and chord sync — AIandDesign · 2026-10-20
- NVAI-open: free open-source local AI image tool fills selections on your own GPU — No-Tap6022 · 2026-10-20
- Suno highlights three Studio 2.0 features users may have missed — suno · 2026-10-20
- AI video of Ben Franklin giving honest art reviews goes viral on Reddit — aeeiee · 2026-10-20