The Qwen Image 2.1 workflow I use every day for reference images and video covers: image editing, outfit swap, text rendering and the official prompt enhancer in one graph.
What's inside
Models: original bf16, 8-step and the 2-step fast version. 2 steps is genuinely good; if fine details are off, reroll, or fix the seed and rerun with 8/16 steps.
- 8 image inputs: 6 references + 2 libraries (outfits, poses). Turn on only what you need.
- Prompt expansion with the official template: turn it on if you don't have a polished prompt. It takes most of the run time, so turn it off when you already have one.
- A fixed box above the prompt for LoRA trigger words; the expander won't rewrite it. Leave it empty if you don't need it.
- Detail pass: VOSR2 for clean sources, SeedVR2 for blurry ones (e.g. video frames). Don't run both.
- Any Angle (camera angle editing), color grading, ControlNet (skeleton / grayscale / mannequin).
- Several ready-to-use prompt sets and notes for multi-angle and the color palette.
Tips
Plain-language prompts work, e.g. "the woman in image 1 wearing the outfit from image 7, 9:16 full body, keep her face and hair".
- Skin too yellow? Add "cool fair skin".
- Text: Qwen Image handles Chinese well, but spell out every word you want on the image.
- LoRAs and the timer node are not included; search the LoRA names in the workflow notes.
Video walkthrough: YouTube channel YJ LAB, EP07.

