References describe appearance; frames describe endpoints
Choose Reference mode when the new shot should draw on several pictures: a product view, a detail, a person’s appearance or an environment. Choose Image to Video when the composition of an existing image should be the opening frame. A reference set is useful when no single source photo contains the whole intended shot. It does not require you to design an exact opening layout in advance.
According to MiniMax’s H3 documentation, first/last-frame conditioning and reference conditioning are distinct workflows. According to ByteDance, images are supported as references in Seedance 2.0. Those descriptions support the general distinction; SeaDance’s model limits and charges below come from its current offering. This guide concerns still-image inputs only. Do not assume a capability on a vendor’s broader product page is a selectable input on SeaDance.
| Model family on SeaDance | Reference stills | Frame-based option | Reference charge rule |
|---|---|---|---|
| Seedance 2.0 / Fast / Mini / 2.5 | 1–9 | 1–2 images | No added per-image charge |
| MiniMax H3 | 1–9 | 1–2 images | First 5 included; $0.04 calculation input per extra image |
| MiniMax H3 Max | 1–9 | 1–2 images | First 2 included; $0.074 calculation input per extra image |
| Kling 3.0 | Unsupported | Exactly 1 first frame | No Reference mode |
Choose useful pictures before uploading
Begin with the clearest view of the subject. For a product, look for a visible silhouette, consistent packaging and lighting that reveals its shape. Add another angle only if it resolves something the main view cannot show. A background photo should explain the intended setting without introducing a competing product. Avoid mixing old and new packaging, contradictory colors or heavily obscured views of the same item.
You do not need to fill the image limit. Each additional photo should answer a practical question: what is the product, what detail matters, or where should it appear? Nearly identical photos can add clutter without adding information. If the prompt asks for an object that is missing from every source, describe it clearly and recognize that the model must invent that part. Review the result against the real product before publication.
Order the set and identify each role
Use a deliberate order and keep it stable while revising the prompt. A useful working arrangement is the main subject first, supporting details next, then the setting or style cue. This is an organizational recommendation, not a documented weighting rule: being first does not guarantee a stronger reference influence. For frame-based input, order does have a defined meaning here: first frame followed by the optional last frame.
In the prompt, refer to “image 1,” “image 2” and so on, then explain the relationship in plain language. For example, request the object from the first picture in the setting shown by the next picture. Avoid assigning the same image incompatible roles or asking the model to merge different products into a supposedly unchanged item. Image labels are descriptive prompt wording, not a promise of exact attachment syntax or perfect identity preservation.
Check the quote after the reference count changes
The following requests are all 10 seconds with sound and 16:9 output. Seedance Mini is sold at 480p, while H3 and H3 Max are sold at 768P. Extra-image amounts in the support table enter the calculation before it is converted into credits and rounded. They are not separate retail cash add-ons. Frame-based images do not incur these reference surcharges.
| Model | 1 reference image | 5 reference images | 9 reference images |
|---|---|---|---|
| Seedance 2.0 Mini | 7 credits / $1.40 | 7 credits / $1.40 | 7 credits / $1.40 |
| MiniMax H3 | 17 credits / $3.41 | 17 credits / $3.41 | 20 credits / $4.01 |
| MiniMax H3 Max | 17 credits / $3.41 | 21 credits / $4.21 | 27 credits / $5.42 |
All generation dollar estimates allocate prepaid credits at Basic monthly: $29.90 divided by 149 credits, approximately $0.200671 per credit. They are rounded to cents after calculating the whole request. They are not separate checkout prices. Other plans have slightly different purchase rates; compare the pricing page before purchasing. New accounts start with 0 credits.
Original prompt examples to adapt
These examples are starting points for a supported 10-second request. Upload the described stills in the stated order, then choose a model that accepts that set. Each prompt separates the identity source from the action. The wording is intentionally specific without asking for a complicated sequence of unrelated events. Replace the fictional subjects with material you have permission to use.
- Tabletop kettle: “Use the matte kettle in image 1. Use the wooden counter and pale wall in image 2 as the setting. Show the kettle standing still while the camera slowly approaches. Preserve its rounded handle and dark finish. Add quiet kitchen ambience; no speech.”
- Travel pouch: “Use the canvas pouch in image 1 and its zipper detail in image 2. Place it on the uncluttered desk in image 3. A hand gently opens the zipper while the camera stays close. Keep the fabric color consistent. Include soft fabric and zipper sounds.”
- Reading lamp: “Keep the lamp shape and brushed finish from image 1. Use the reading corner from image 2. Show the lamp glowing steadily beside a closed book as the camera moves sideways. Preserve the shade silhouette. Use quiet room ambience without dialogue.”
- Running shoe: “Use the shoe in image 1, preserving the sole pattern shown in image 2. Place it on the neutral display surface in image 3. Show a slow camera orbit around the stationary shoe. Keep the background simple and the shoe fully recognizable.”
- Ceramic bowl: “Use the blue bowl from image 1 and the glaze detail from image 2. Use the table arrangement in image 3. Show a hand setting the bowl down gently, followed by a still close view. Preserve the rim shape and blue glaze; include a soft contact sound.”
Review the references as well as the output
If the product changes, first check whether the source images disagree. If the background dominates, clarify which picture supplies only the setting. If the movement is difficult to follow, simplify the action while keeping the reference set unchanged. Changing all the images and the prompt together makes it harder to learn which correction helped. Avoid treating additional uploads as an automatic improvement.
Before submission, confirm the model, mode, order and quote. After delivery, compare the important product features with the sources and listen to the generated sound. A failed generation returns its deducted credits, while a successful backup keeps the original quote. Use support for job or balance questions and the refund policy for purchase requests.
Frequently asked questions
Can I upload the full image limit on every model?
No. Reference-capable Seedance and H3 variants accept up to 9 stills, but Kling has no Reference mode here. Always check the selected model before uploading.
Is the first reference an exact starting frame?
No. For an intended starting composition, use Image to Video. Reference images describe visual material for a new shot rather than guaranteed chronological endpoints.
Do more reference images cost more?
For H3 and H3 Max, images beyond their included counts increase the request calculation. The Seedance variants listed here have no separate reference-image surcharge.
Can I use a product photo and a room photo together?
Yes, on a reference-capable model. Tell the prompt which photo supplies product identity and which supplies the setting, and remove any contradictory objects from the brief.
Do the examples guarantee matching labels and shapes?
No. They are original creative instructions. Inspect text, product proportions and other important details in the delivered result before using it to represent a real product.
Sources
Related articles
Kling and H3 inputs · Choose a product-ad workflow · First-frame prompt tips
