Skip to main content

Character-consistent outfit change photo set

Lock identity in the first sentence and only let the model change clothing and scene, batch-producing different looks for the same person.

Disclaimer: All content shown on this page is generated by AI models and is for reference only; it does not constitute any guarantee of results. Actual output and Credits consumption are subject to "Usage Details" in the console.
This playbook runs an outfit-change photo set from a single reference portrait: swapping outfits and scenes one set at a time while keeping the face consistent throughout. If you just let the model redraw, every face comes out different. The right approach is to provide a reference image, write "identity lock" as the highest priority, and only allow it to change clothing, scene, and lighting. Call chain: wan2.7-image text-to-image (produce the reference portrait) → wan2.7-image reference-image editing (swap outfits and scenes one set at a time, writing identity lock at the very front each time).

Result

From the same reference portrait, the photo set after swapping outfits and scenes one set at a time:

Reference portrait (beige knit · solid gray background)

Reference portrait

Navy blazer + office floor-to-ceiling window

Outfit look one

Cream-white sweater + scarf + autumn street

Outfit look two

Step: Put identity lock in the first sentence

Once the reference image is produced, for each look, pass the reference image to wan2.7-image for one edit, with identity lock written at the very front of the prompt:
Strictly keep the facial features, face shape, hairstyle, hair color, and skin texture of this {person} in the frame completely unchanged—it is the same person.
Only change her outfit to {clothing, e.g. navy professional blazer over a white shirt},
change the scene to {scene, e.g. in front of a bright modern office floor-to-ceiling window},
with {lighting, e.g. soft early-morning natural light}.
Keep a high-realism portrait quality, {framing, e.g. frontal half-body} composition.
Do not change the appearance, do not cartoonify, do not add extra limbs.
Hand the reference image and the two photos to qwen3.7-plus for cross-comparison. Its judgment:

Second image (business attire): same person
  "The facial features (eye shape, nose bridge height, lip shape), face shape (a slightly long oval face), and hairstyle
    (black, shoulder-length, center-parted straight hair) are highly consistent with the first image. The glasses style (thin metal round frames)
    is also fully preserved."

Third image (casual wear): same person
  "It also preserves the core facial features of the first image, including brow-to-eye spacing, nostril width, and the curve of the mouth."

Flaw check: neither shows obvious facial drift, body-type mutation, or extra limbs;
  the blazer shot has firmer shoulder lines, the sweater shot has a bulkier upper body—normal differences due to garment cut and thickness.

(Note: this is a cross-comparison done by a vision model, not authoritative face recognition. For important use cases, verify again yourself.)

Notes

  • Consumption: This playbook involves image generation + image editing. The reference image is produced once, and after that, each look is one edit call. To produce a set of 5 looks is 1+5 calls—far cheaper than video, and well suited to batch image production.
  • Consistency: The identity-lock line must be placed at the very front of the prompt, with an explicit "this is the same person, do not change the appearance." Placing it at the end makes it easy for the clothing and scene descriptions that follow to override it. When generating multiple looks across multiple passes, always use the same reference image, rather than taking the previous output and editing it further—otherwise errors accumulate shot by shot and drift further from the original.
  • Compliance: When using a real person's photo for outfit changes, you must obtain the person's consent; it may not be used to forge identity or for misleading promotion. The demo images are AI-generated fictional characters.
Model Playground
Fine-tuning
Model Compression
Statistics and Monitoring
Support