| name | wanx-best-practices |
| description | Wanxiang/Bailian image generation guidance for WanxImage prompts, negative_prompt usage, styles, sizes, editing, and multi-image references. |
| metadata | {"provider":"bailian","family":"wanxiang","version":"1.0.0"} |
Wanx Best Practices
Use this skill before calling Video Production Buddy WanxImage, the Bailian/Wanxiang
image tool. This guidance is provider-specific to Wanxiang image models and
must not be replaced with FLUX-only rules.
When To Use
- Generating still images with
wan2.7-image-pro, turbo variants, or other
Wanxiang image models exposed by WanxImage.
- Editing an existing image with natural-language instructions.
- Using one or more reference images to preserve product, character, or style
consistency across a production set.
- Creating Chinese visual styles such as ink, watercolor, flat illustration,
anime, portrait, photography, film, or product-packshot looks.
Prompt Structure
Write direct, visual prompts:
[subject/product] + [scene/action] + [style] + [camera/framing] +
[lighting] + [color/material details] + [composition constraints]
For physical products, anchor exact geometry and visible brand details in the
positive prompt, then use a reference image when product fidelity matters.
Text-only prompts are not enough for recognizable hardware, packaging, logos,
or UI.
Negative Prompts
Wanxiang supports negative_prompt; use it when the scene has clear failure
modes. Keep it concise and specific.
Good negative_prompt examples:
extra fingers, distorted logo, wrong product shape, unreadable text,
low resolution, watermark, duplicate object, melted edges
Do not copy FLUX guidance that says negative prompts are unsupported. For
WanxImage, negative_prompt is part of the tool contract and should be used
when it reduces predictable artifacts.
Model And Parameter Notes
- Use
wan2.7-image-pro for higher-quality image generation or editing when
cost and latency are acceptable.
- Use lower-cost/turbo variants only for drafts, bulk exploration, or cases
where small artifacts are acceptable.
- Use explicit supported sizes from the tool contract, such as
1024*1024,
1920*1080, 1080*1920, or other WanxImage-supported width*height
values.
- Use
style only when it matches the brief. Prefer explicit prompt language
for brand-sensitive output instead of relying only on a preset.
- Use
seed when a production needs repeatable variations.
Editing And References
For image editing, state what must change and what must remain unchanged:
Change the background to a moonlit studio surface. Keep the phone shape,
camera island, logo placement, color, and reflections unchanged.
For multi-image references, assign each reference a role in the prompt:
- Product reference: preserve geometry, material, color, and branding.
- Style reference: borrow lighting, palette, texture, or illustration style.
- Pose/layout reference: borrow composition only, not identity.
Quality Checks
Before accepting Wanxiang output, inspect for:
- Product or logo drift.
- Extra objects, duplicate limbs, or warped hands.
- Text artifacts and unreadable labels.
- Aspect-ratio mismatch against the deliverable.
- Watermarks or provider artifacts.
If any brand-critical geometry is wrong, retry with a clearer prompt and the
reference image rather than explaining the product in prose alone.