"Wash" is the local term for a specific operation: take an image, analyse what is actually in it, then re-render it at higher quality without changing the content. Done badly it is just an upscale. Done well it recovers detail the source never had.

This episode builds that pipeline with Qwen Image 2.1 and documents fifteen test sets.

The three stages

1. Vision analysis (Gemini). The source image is described in structured terms — subject, composition, materials, lighting. This is not captioning for a human; it is producing a specification the generator can follow.

The quality of this stage bounds everything after it. A vague description produces a generic re-render.

2. Qwen Image 2.1 re-render with Detail LoRA. The image is regenerated from the specification at target resolution. The Detail LoRA is what makes the re-render add structure rather than just resolution — it biases the model toward plausible fine detail (fabric weave, surface texture) instead of smooth gradients.

3. 2K output. The final upscale. Done last, because upscaling before the re-render means the model works from interpolated pixels rather than from the specification.

What the 15 test sets showed

The interesting result is about failure modes by content type:

  • Product shots — the pipeline excels. Clean geometry, controlled lighting, and the model reconstructs material convincingly.
  • People — good at portrait scale, degrades on full-body at distance. Facial structure is preserved; hands and small accessories are re-invented.
  • Text in image — fails predictably. Any lettering gets re-rendered into plausible-looking but wrong glyphs. Composite text after the pass, never through it.
  • Architecture — very good, because the geometry is regular and the model has strong priors for it.

That mapping is more useful than a quality score: it tells you what to feed the pipeline.

Files

Video walkthrough (Chinese): https://youtu.be/-80Jf67eOF8
Bilibili: https://www.bilibili.com/video/BV1SxaT6nEvk/

Notes

  • Single ComfyUI workflow export; Qwen Image 2.1, the Detail LoRA, and a vision-model API key are required.
  • The vision stage calls an external model; costs and rate limits apply.
  • Tutorial material, not a product.
  • Independent AI-KSK release; not affiliated with the model authors.