Seedance adds acting-style extraction for scene-to-scene reuse

• _OAK200's acting transfer demo shows the same audition performance being carried across three new scenes, with dialogue kept deliberately similar so the comparison stays focused on delivery, emotion, and pacing.
• According to _OAK200's warning about video reference, feeding the audition clip back in as a reference often drags unwanted framing, camera movement, lighting, and composition into the new shot.
• The reusable trick in _OAK200's acting-style method is to ask the system prompt to extract only the acting style, then paste that distilled direction into a fresh scene prompt.
• _OAK200's extraction prompt narrows the handoff to performance variables like restraint, pauses, breath control, eye behavior, and emotional arc, while explicitly banning dialogue, plot, costume, and camera carryover.
• The broader workflow in _OAK200's audition pipeline treats Seedance like a casting room first and a scene generator second, which is the part creators keep circling back to.

You can watch the three-scene transfer demo and compare it with the warning against raw video references. The deeper shift is in the full audition prompt, which turns character design into role options, audition lines, and voice triggers before any final scene gets generated. Elsewhere in the Seedance crowd, techhalla's continuity thread is chasing a different bottleneck, stretching clips past 15 seconds, while one commenter notes that chaining generations can quietly degrade image quality.

The starting point is simple: a character sheet can lock appearance, but it does not tell you how the character speaks, reacts, or holds tension. That is why _OAK200's casting-agent framing centers the process on an audition instead of going straight to the final scene.

The workflow shown in _OAK200's tool chain breaks into four steps:

1. Midjourney generates the character.
2. GPT Image 2 builds the character sheet.
3. A custom system prompt creates role options, audition lines, and voice triggers.
4. Seedance 2.0 generates the audition.

That workflow matters because the final scene prompt is not asked to invent performance from scratch. In _OAK200's follow-up, the stated goal is to figure out how the character moves, speaks, reacts, and performs before the real scene generation starts.

The clearest limitation in the thread is that a reference clip can transfer far more than acting. _OAK200's breakdown says Seedance may inherit medium-shot framing, camera motion, lighting, color, or composition from the audition, which makes the model imitate the whole clip instead of just the performance.

That is a useful distinction for anyone treating reference video as a universal continuity tool. In this setup, raw reference footage is described as less reliable precisely because it preserves too much of the old scene.

The replacement is prompt distillation. Instead of passing the audition back in as a clip, _OAK200's extraction prompt tells the system to read the approved Seedance prompt, casting read, character sheet, assigned role, and voice triggers, then output only an actor-directable style block.

The exclusions do most of the work. _OAK200's extraction prompt explicitly bans copying:

• dialogue
• scene or setting
• costume
• camera
• plot
• specific story events
• placeholders or templated output

The allowed variables are the performance layer:

• emotional starting state
• restraint level
• vocal tone
• pacing, silence, and pauses
• verbal fracture or stutter
• loudness shifts and breath control
• eye behavior and facial micro-movements
• body movement
• emotional arc
• what to avoid

By the end, the system is asked for 6 to 9 concise lines under an "ACTING STYLE TO APPLY" header, ready to paste into a different scene prompt _OAK200's output format.

The longer prompt underneath the workflow is more structured than the short extraction post makes it sound. _OAK200's full audition prompt has the model move through seven stages before it writes the final Seedance prompt:

1. Casting read
2. Role options
3. Audition purpose
4. Audition lines
5. Voice triggers
6. User choice
7. Final Seedance audition prompt

The interesting part is how specific those stages get. The prompt asks for objective, obstacle, tactic, emotional shift, and subtext, then forces three short audition lines and three compact voice triggers. It also includes a role reset rule, so switching from protagonist to antagonist is supposed to rebuild the scene logic rather than lightly remix the same beat pattern _OAK200's divergence rules.

That makes the acting-style extraction post feel less like a one-off hack and more like a second pass in a broader pipeline: audition first, distill second, then reuse the distilled performance in fresh scenes.

A parallel set of Seedance experiments is focused on longer clip continuity, not acting transfer. In techhalla's continuity thread, the claim is that getting past 15 seconds depends heavily on workflow, with Magnific used as the interface for text-to-video generation and extension.

The replies add two useful caveats. DavizCF7777's quality note says repeatedly feeding the latest generated videos back as references can degrade image quality over time, while techhalla's setup notes says duration changes need matching prompt changes and that this particular workflow skips image generation entirely before extending from the first Seedance clip.

Those posts are solving a different headache than the audition thread, but together they sketch the current Seedance split: one camp is trying to preserve performance across scenes, the other is trying to preserve coherence across time.

The payoff for all this prompt engineering is visible in the kind of clips people are shipping. MayorKingAI's Runway example pushes Seedance 2.0 toward large-scale VFX, while underwoodxie96's skier clip shows the model holding a stylized character design across four action beats with a very explicit shot-by-shot prompt.

Those examples are not about acting transfer directly, but they show why creators are getting picky about which parts of a generation should stay fixed. Once prompts are controlling shot order, motion style, palette, and continuity this tightly, extracting only the performance layer starts to look like a practical separation of concerns, not just a prompt-writing flourish.