@-references in Seedance versus R2V endpoints
Seedance 2.x @-files are multimodal references in one generate. /models/*-reference-to-video is a dedicated R2V endpoint. Different APIs, same job family.
Do not mix both on one shot. Pick the endpoint the catalog exposes for the model.
Seedance 2.x @Image1 / @Video1 / @Audio1 tokens are multimodal references in one generate: you attach files and you point at them from the prompt. /models/*-reference-to-video is a dedicated R2V endpoint — Seedance 2.5 reference-to-video, Wan 2.7 reference-to-video, Veo 3.1 R2V, and the rest. Different APIs. Same job family: identity as an input, scene as a prompt.
The job family is image-to-video vs reference, explained. This page is the routing rule inside that family: which door you open, and why opening both on one shot is how you lose the identity you meant to lock.
Two doors, one job
@-references (Seedance-shaped generate). The prompt language is part of the input contract. @Image1 is the presenter, @Image2 is the jacket, @Video1 is the motion you want continued, @Audio1 is the line or the room. The files sit on the same call as the text. Seedance 2.0 Fast documents this syntax; Seedance 2.5's R2V listing uses the same tokens and raises the ceiling (up to 30 images, 10 videos, 10 audio — a kit, not a mood board). If you write "the woman from the second photo" instead of @Image2, you are not using the contract. You are hoping.
Dedicated R2V endpoints. The catalog page is the mode. You pick Wan 2.7 reference-to-video (or Veo, Kling O3, Happy Horse) and you upload into that model's slots. Many of these have no token syntax. Veo, Kling O3, and Wan want a plain-language subject and let the files carry identity. Happy Horse wants character1…character9, which is not @Image1. Pasting Seedance tokens into Wan is flavour text. Pasting character1 into Seedance is flavour text.
Same job: recast this person / product / room in a scene I describe. Different address.
Pick the endpoint the catalog exposes
| Model you already chose | Door | How you point at files |
|---|---|---|
| Seedance 2.5 R2V | seedance-2-5-reference-to-video | @Image1, @Video1, @Audio1 in the prompt |
| Seedance 2.5 T2V / I2V | Those endpoints | You do not get the 50-slot kit. Do not @-token a T2V call and expect the files to attach |
| Wan 2.7 | wan-2-7-reference-to-video | No @ tokens. At least one reference image or video. Optional voice. |
| Veo 3.1 R2V | The Veo R2V page | Up to 3 images. Plain language. Negative prompt for exclusions. |
| Happy Horse R2V | That model's page | character1…character9 in file order |
| Kling O3 R2V | That model's page | Plain language. Audio often off until you turn it on. |
The failure is mixing doors:
- Running Seedance T2V with
@Image1in the prose, and also uploading the same still to a different R2V endpoint "for backup." You now have two identity events. They will not agree. - Uploading a kit to Wan R2V and writing
@Image1because a Seedance tutorial said so. Wan does not read that token. - Feeding a finished first frame to R2V (any flavour) and asking it to "just add motion." That is I2V. R2V will restage. The I2V vs R2V contract still holds regardless of syntax.
One shot. One door. One syntax.
When the @-generate is the right door
Use Seedance's @-kit when identity and motion and a sound reference have to travel together in one pass.
A presenter (@Image1) in a room (@Image2) who must walk the way @Video1 walks and land on a line from @Audio1, for up to ~30 seconds, is the job that ceiling exists for. Fifty slots is not a default. A product still plus a character still is the usual kit. Extra files past what the scene uses are not "more consistency." They are extra things the model can get confused by.
Do not open this door for a product turntable that already has a hero still. That is I2V. Do not open it for a Veo talking-head that needs always-on dialogue more than a large kit. That is Veo.
When the dedicated R2V page is the right door
Use the catalog R2V endpoint named for the model you already picked for other reasons.
You are on Wan because the file cannot look like Seedance, or because you want Wan's prompt_extend and negative prompt. Then you open Wan R2V, not Seedance's @-grammar. You are on Veo because the mouth is the shot. Then you open Veo R2V with three images, not fifty Seedance slots you will not fill. You are on Happy Horse because of multilingual lips. Then you number characters the way Happy Horse numbers them.
The model choice comes first (job, audio, duration, licence). The door is whatever that model's catalog page exposes. Do not pick a door and then smuggle another model's syntax through it.
The kit, either door
Each file does one job. Face. Wardrobe. Product. Room. A collage is not a kit. The prompt then names the role, in the syntax the door requires.
Caps are hard. Veo will not take a fourth image. Wan will not take a sixth in an array. Seedance 2.5's R2V will not silently blend file 51. Extras are dropped, and the output will not tell you which. Budget the kit to the door, not to the folder on disk.
FAQ
Can I use @-tokens on Seedance text-to-video and skip the R2V page?
Only if that T2V endpoint's schema actually accepts the files. On Versely, Seedance 2.5's multimodal kit is the reference-to-video listing. The T2V listing is a text prompt (and I2V is a first frame). Writing @Image1 into T2V prose without slots is a caption, not a reference. Open the R2V page when you have files.
Why not always use Seedance because the ceiling is highest?
Ceiling is not quality, and it is not audio character. A 50-file kit on the wrong model is a 50-file mess. Pick Seedance R2V when the job is Seedance's job (I2V lock, long single pass, multimodal kit). Pick Wan or Veo R2V when those jobs won. The @-grammar does not travel.
Is I2V a third door I should mix in?
I2V is a different contract: the image is the first frame. You may chain — R2V (or @-generate) to stage, then I2V the best frame for a hero. That is two shots, two doors, in sequence. It is not two doors on one shot.
What if I need a voice reference and the R2V endpoint has no audio slot?
Then that endpoint cannot take the voice. Seedance R2V and some Wan setups can; Veo's R2V is images. Do not paste @Audio1 into a model with no audio array. Drive the voice another way (native prompt on Veo, TTS + lipsync, a talking-shot model) or change endpoint. Syntax cannot create a slot.