Seedance 2.5 Reference Images: How @Image1 and @Image2 Get Assigned

What this tool does

• The automatic @Image mechanic applies to Seedance 2.5 only. On Seedance 2, Seedance 2 Fast and Seedance 2 Mini the effect's prompt is sent to the model unmodified.
• The automatic reference sentence is prepended only when the effect's stored prompt contains no @Image tag. If the effect author wrote their own tags, that authored mapping is used exactly as written.
• The exact text prepended is: "Reference @Image1 for the person's face and appearance; reference @Image2 for the scene and visual style." followed by the effect's own prompt. The tag numbers expand to cover however many images sit in each group.
• Numbering is positional and deterministic: your uploads are numbered first, starting at @Image1, and the effect's stored reference images are numbered after them.
• Your first uploaded photo is therefore always @Image1 — the slot the model is told to read the subject's face and appearance from. Upload order changes the result even though it never produces an error.
• On the Effects tab you upload the effect's total reference-image count minus the reference images the effect already carries. Sending the wrong number is rejected before any credits are spent, with a message stating exactly how many images that look needs.
• Effect photos are sent as reference images, not as a first frame. Seedance 2.5's forced-"adaptive" rule applies only to first-frame or last-frame tasks, so on the Effects tab your aspect-ratio choice — 16:9, 4:3, 1:1, 3:4, 9:16, 21:9 or adaptive — is the one used.
• The Image to Video tab reaches the same reference-image input: with a Seedance model selected, the Images panel offers Frames (I2V) and Multimodal, and Multimodal sends up to 9 reference images. Frames is the first-frame path, and that one does force adaptive on 2.5.
• A reference-image list is capped at 9 images, and uploads outside the accepted 300-6000 pixel range are downloaded, resized and re-hosted automatically before the job is submitted.
• Seedance 2.5 accepts prompts up to 30,000 characters against 20,000 for Seedance 2, so the prepended reference sentence never crowds out the effect's own instructions.

How it works

1. 1.Open the effect and read the upload count

   Go to /dashboard/video-creation/seedance-effects/ and open a look. The interface states how many images it needs from you, which is the effect's total reference count minus whatever reference images the effect already carries. Some looks need one photo, others more, and a few need none because the effect supplies all its own references.

2. 2.Order your uploads before you drop them in

   The first file you add becomes @Image1, the second @Image2, and so on, ahead of any references the effect contributes. Since @Image1 is described to the model as the person's face and appearance, the clearest face photo goes first. Uploading a scene plate first does not fail — it just tells the model to read a face out of a photo that does not contain one well.

3. 3.Give @Image1 something to work from

   A single subject, face unobstructed, evenly lit, facing roughly toward the camera. Keep the short side above 300 pixels: anything smaller is upscaled automatically to meet the minimum, which passes validation but does not recover detail the file never had. Crowd shots and heavy filters make the identity slot ambiguous.

4. 4.Let the effect handle scene and style

   You are usually not uploading the style plate. Reference images attached to the effect fill the later slots and are the ones the prepended sentence points at for scene and visual style. That is why the upload count is often lower than the effect's total reference count, and why you should not try to pad the list with extra images — the request will be rejected.

5. 5.Set model, duration, resolution and aspect ratio

   Switch the model to Seedance 2.5 explicitly, since the tab opens on Seedance 2 and the automatic @Image behaviour only applies to 2.5. Duration runs every second from 4 to 30, resolution is 480p or 720p, and your aspect-ratio choice is honoured because these images travel as references rather than as a first frame.

6. 6.Use Image to Video when you want to write the mapping yourself

   Pick Seedance 2.5 in /dashboard/video-creation/image-to-video/ and switch the Images panel to Multimodal. Your photos go in as reference images — up to nine — and the prompt is yours, up to 30,000 characters, so you describe which photo carries what instead of relying on the automatic sentence. Leave the panel in its default Frames mode instead and your photo becomes the first frame, which forces the output ratio to match it.

7. 7.Test cheap, then re-run with the same order

   Run 4 seconds at 480p for 37 credits to check that identity carried the way you wanted. If it did not, swap which photo is first before changing anything else — that single change alters what the model is told to read the face from. Credits are refunded automatically if a render fails.

Frequently asked questions

Related tools