No broad theory here — just specific habits confirmed through extensive testing to actually improve results.
1. Match Your Scene to the Product's Register
This is the most common failure mode. A luxury perfume on a beach looks like a souvenir; workwear in a fine dining restaurant looks out of place. Every scene element in your prompt needs to match the product's market position.
❌ luxury perfume bottle, beach setting ✅ luxury perfume bottle, private yacht deck, white linen tablecloth, golden sunset
2. Specify Camera Movements Concretely
"Show the product," "make it move," and "camera movement" are too vague. The same prompt generates completely different motion on each run. Use specific cinematography terms:
slow push-in — gradual zoom in orbit around — rotate around subject overhead tilt — overhead angle tilting to eye level static camera — fixed camera (subject moves)
3. Don't Stack Quality Modifiers
"ultra high quality, best quality, masterpiece, perfect, stunning" stacked together adds no value and can cause oversaturation. What works is describing quality specifically:
❌ ultra high quality, best quality, perfect ✅ leather texture clearly visible, metal edges sharp and clean, no blurring
4. Tell the Model What to Do With Your Reference Image
If you upload a reference without explaining it in the prompt, the model's use of it is inconsistent. Be explicit about what to keep and what to change:
Based on the reference image, maintain the exact product color and logo placement. Change: background to white studio, lighting to soft overhead.
5. Use Negative Descriptions to Exclude Unwanted Elements
Adding what you don't want at the end of your prompt measurably improves consistency:
..., no text overlay, no watermark, no extra hands or fingers, no motion blur, no color distortion, no background clutter.
6. Keep Seedance Video Prompts Under 60 Words
Beyond 60 English words, Seedance models (especially 2.0) start ignoring parts of the prompt — typically whatever appears later. For complex scenes, break into multiple shorter clips. More predictable than one long prompt.
7. Save Your Successful Prompts
When you get an output you're happy with, immediately save the prompt to a local document with the model name and use case noted. Next time you have a similar need, modify that prompt instead of starting from scratch. This single habit saves more testing time than anything else.