Scene References Do More for Video Success Than Longer Prompts
A Japanese AI drama studio ran about 50 generations on Seedance 2.5, trying detailed prompts, timestamps and layout sketches. The method that actually raised the success rate was a scene reference. Here is what that changes about how you write video requests, and what it does not fix.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
On this page

The Japanese studio CreativeEdge CL+ runs AI drama, anime and games in parallel, mostly solo, and publishes its test notes in detail. In a September post about lip sync, he reports what happened after roughly 50 generations on Seedance 2.5 spent comparing detailed prompts, timestamp beats and layout sketches: the technique that actually raised the success rate was not in the prompt at all. It was supplying a scene reference.
That result matches what anyone who has burned credits on a stubborn shot already suspects. Past a certain point, adding words does not add information the model can use.
What the lip-sync test showed
The post is not a controlled test. It sets one HeyGen Avatar V clip against two Seedance 2.5 variants built from different inputs. HeyGen's 47-second talking head was driven by ElevenLabs audio and behaves like the product it is: front-facing, steady tone, clean delivery. The author's own recommendation is blunt. For long monologue, use HeyGen without hesitation.
Seedance 2.5 generated 15 seconds instead, partly because his Runway plan was unlimited but capped, and the result got fast and compressed. Two variants were compared: one conditioned on a 23-second ElevenLabs clip, one where the dialogue sat in the prompt and Seedance generated the voice itself. The second felt more present on screen.
His verdict is that Seedance 2.5 is clearly the stronger image maker, and also the more expensive one to work with, because the success rate is low. Japanese pronunciation and intonation drift. Lines change between takes. He names MiniMax H3 as the more reliable model for hit rate and still uses Seedance 2.5 anyway, because the performance is the part he cannot get elsewhere.
That tradeoff is worth naming plainly, because people keep shopping for a model that does not have it. You are choosing between a stable presenter, a reliable generator and a model that gives you a face worth looking at. Those are three different products with three different failure modes.
Why a reference outperforms a longer prompt
A prompt describes the result in a compressed form. A reference shows it. When the model has the frame, the ambiguity that your adjectives were trying to resolve is already gone, and the generation has one target instead of a paragraph of constraints that may contradict each other.
The practical consequence: every extra clause you add to nail down a shot is another condition the model has to satisfy at once. If two clauses conflict, nothing in the system tells you which one wins. You get a take that is plausible and wrong, and no diagnostic for why.
His method is to hand the model a proper scene reference, and the cost is visible. Reference counts run 10 to 20 on average, occasionally hitting the spec ceiling of 30. That is not free. Each image has to be produced, chosen, and kept consistent across a sequence, and a large pack can fight itself if two files describe the same thing differently. The role-assignment discipline in our Seedance 2.5 multi-reference workflow exists exactly because of that second problem.
So the honest version of the advice is narrower than "use more references". It is this: invest in the reference set before generation, not after the third failed take.
A two-strike rule that saves credits
The most useful operational change I would take from this is a stop condition, with a diagnosis in front of it. If a shot fails twice with the same reference set, and the failure is something a reference is supposed to settle, the next attempt should change the reference, not the prompt. Rewriting prose feels like progress because it is fast, and it is usually the cheapest way to waste an afternoon.
The rule does not cover every failure. His own test is the counterexample: 23 seconds of ElevenLabs audio pushed into a 15-second generation came out rushed. No reference fixes that, because the problem is duration. Diagnose before you edit. A take that fails on timing, speech rate, or a line that mutates needs a shorter line or a longer generation. A take that fails on how something looks is a reference problem.
Text still earns its place in three situations:
- There is no reference yet, and you are describing a scene for the first time.
- The constraint is temporal. What happens, in what order, at what moment.
- You are directing performance or audio, where words carry intent that pixels cannot.
What text does poorly is specify what a thing looks like. That is what a photo is for.
Diagnosing lip-sync uncanny instead of guessing
The post also contains the clearest explanation of lip-sync uncanny valley I have read from a working studio, and it is useful because it is a checklist rather than a mood.
The mechanism is perceptual mismatch. An audience does not find a face unsettling because it is too realistic. They find it unsettling because some parts are realistic and others are not, and the eye integrates the whole chain unconsciously: lips, jaw, cheeks, around the nose, brows, eyelids, gaze, head, neck, shoulders.
Three specific mismatches are named, and each points at a different fix:
Mouth amplitude against vocal force. The voice is projected but the mouth barely opens. This is a performance-consistency problem. A performance or motion reference helps more than a sentence asking for a wider mouth.
The inside of the mouth changing between frames. Teeth and tongue shapes that mutate per frame. This is a hard limit of the generator at close range. The practical response is to shoot these lines wider, or shorten the take, rather than spend credits on a fix that will not hold.
Upper and lower face disagreeing. The mouth moves but the brows and eyes do not carry the emotion of the line. Either the emotional direction is missing, or it is buried under a reference pack the model is busy reconciling.
He notes that Seedance 2.5 can be directed down to facial muscle micro-expressions, and that eye saccades, small facial motion and body sway are what produce the sense of a living person. That is the argument for using a model with range even at a lower hit rate: on dialogue close-ups, the micro-motion is the product.
What I would not take from the post
Treat the 50 generations as one artist's experience, not a benchmark. There are no per-axis numbers, no prompt-to-prompt comparison table, and the workflow runs through a specific provider whose generation length shaped at least one of the tests. The conclusion that references beat prompt detail is credible because it is consistent with how conditioning works and with what production users report, not because 50 is a large sample.
The image-reference limit of 30 is a spec limit, not a target, and the same model also takes up to 10 video and 10 audio references. The 10 to 20 figure is what this workflow actually cost him across roughly 50 generations, not a general planning target, so size your own reference set to the shot and count the authoring time, which does not appear in a per-second price. If you are budgeting, the retry math in cost per usable second is the honest way to compare two models, and the camera vocabulary in cinematic video prompts is what to keep in text once the look is carried by images.
The takeaway that survives all of that: build the reference before you build the sentence. A detailed prompt is a description of an image you have not made yet. A scene reference is the image, and the model has far less room to be creative with it in the wrong direction.