Seedance 2.5 demo clones a creator’s voice and likeness from one prompt

• A single Seedance 2.5 prompt produced a talking office-tour avatar from a photo, a voice clip, and office images, according to venturetwins' demo, and their reply said it was one continuous generation.
• The creator spec is now built around longer, reference-heavy shots: AIwithSynthia's feature summary listed 30-second generation, videos up to 3 minutes, 50 multimodal references, timestamp control, green-screen assets, and Maya and Blender plugins.
• Indie film workflows are already shifting from clip stitching to reference packing, with techhalla's CapCut breakdown using 30-second takes and up to 50 reference images, while PJaccetturo's Nexus thread used Seedance 2.5 inside a feature-film pipeline.
• The friction is access, price, and resolution: venturetwins said Dreamina required a European IP, venturetwins called it a pricey model, and techhalla said the current downside was 720p output.

ByteDance's Seed Team launch post says Seedance 2.5 accepts 30 images, 10 video clips, and 10 audio clips in one pass. Dreamina's launch page says the model is rolling out first to subscriber accounts over 16 in Europe, Asia, the Middle East, and South America. The weirdest creator detail came from venturetwins' voice prompt reply: the voice instruction was just, “use the voice from the video as a reference.”

venturetwins gave Seedance 2.5 three kinds of reference material: a photo of himself, a clip of himself talking, and public images of the a16z SF office. The result was a synthetic host giving an office walkthrough in his image and voice.

The prompt apparently did not need a production stack around it. venturetwins said the clip was “one continuous generation, not different clips,” and a later reply said the voice prompt was simply, “use the voice from the video as a reference.”

That makes the demo more useful than the usual character reel. It tests likeness, voice reference, scene references, camera cuts, and talking-head performance inside one short commercial-style output.

ByteDance's launch post says Seedance 2.5 extends single-pass generation from 15 seconds to 30 seconds, with multiple rounds of extension for longer videos. It also says the model can organize multiple connected shots inside those 30 seconds, including setup, development, turn, and resolution.

Creators immediately wrote prompts like timed mini-screenplays. techhalla's diner sequence maps the action from 0 to 30 seconds across a spill, freeze, orbit, rewind, and close-up, while another techhalla prompt uses the same structure for a roller-coaster gag.

The practical change is simple: a 30-second generation can now hold a whole bit, not just a transition.

Seedance 2.5's official reference ceiling is 50 assets in one pass. ByteDance breaks that into 30 images, 10 videos, and 10 audio clips, while Dreamina's launch page describes references across images, video, audio, and text.

Creators are treating those references as a production bible:

• techhalla's short-film thread started with three images: environment, main character, and bike with trailer.
• techhalla's continuity step added the last 15 seconds of the previous video as a new reference, plus character and prop sheets.
• techhalla's loop repeated the process by cutting useful 15-second clips and loading reference material so the model had less room to invent.
• techhalla's survival-TV test used 42 image references, including islands, gear, props, tents, parachutes, and character sheets.

The best examples look less like prompting and more like art department work compressed into an upload tray.

PJaccetturo's Nexus thread is the cleanest production breakdown in the evidence pool. The team used Seedance 2.5 for a fantasy feature-film sprint, then published the ingredients behind the dragon battle.

The workflow was structured like this:

1. Story meeting, transcript, screenplay formatting, revisions, then a shot list, according to PJaccetturo's script step.
2. Pinterest look development and location inspiration, shown in PJaccetturo's look-dev post.
3. Character design with a rule of “80% standard and 20% unexpected,” from PJaccetturo's character note.
4. Scene grids that attach characters to each scene as ingredients, shown by PJaccetturo's organization board.
5. One 30-second Seedance 2.5 run to generate multiple angles from the shot-list moments, described by PJaccetturo's coverage step.
6. Screenshot the strongest frames from the first generation, then upload them back as anchor images, according to PJaccetturo's cost-saving step.

The long prompt in PJaccetturo's prompt post reads like a shot list, character bible, location brief, and sound-design note fused into one generation request. In a reply, PJaccetturo said time prompts were too restrictive, a useful counterpoint to the timestamp-heavy examples elsewhere.

Dreamina's launch page says Seedance 2.5 adds granular editing for timestamps, characters, and specific video elements. ByteDance's post also calls out green-screen editing, camera perspective editing, and reference-based editing.

The creator demos already split into two editing patterns:

• HalimAlrasihi's Smart Edit test replaced dragon fire with wasps, ice, and flowers while preserving the underlying motion.
• venturetwins' Rednote find showed a Timberland-style ad replicated with the shoes swapped out, and venturetwins' follow-up connected that pattern to video-to-video features in newer models.

For ad people, this is the spicy part: the model is not only generating footage, it is starting to behave like post-production software.

Dreamina's official rollout says Seedance 2.5 will reach subscriber accounts over 16 in Europe, Asia, the Middle East, and South America in the coming week. hasantoxr's launch thread said US creators were expected to wait roughly one more week, and venturetwins said he used Dreamina with a European IP.

Other surfaces are forming around it:

• higgsfield_ai said Seedance 2.5 was coming soon to Higgsfield.
• higgsfield_ai advertised 14 days of unlimited Seedance access around a K-pop one-shot demo.
• magnific said Seedance 2.5, powered by BytePlus, was coming soon to Magnific.
• ByteDance's launch post says API access is coming soon through BytePlus ModelArk.

The cost picture is still messy. venturetwins called it “a PRICEY model,” venturetwins said a Rednote creator described it as about twice the cost of Seedance 2, and QualitativAi said Dreamina accounts appeared to show different credit costs.

Resolution is the other hard limit in creator reports. techhalla said the only downside in his testing was 720p output, and 0xInk_ said native 1080p would be better even if upscaling is possible.

The office-tour demo is a consenting self-clone, but it lands directly on the most sensitive feature in the release: reference-based voice and likeness transfer. ByteDance's launch post says Seedance 2.5 can preserve the appearances and voices of multiple characters in complex scenes.

Dreamina's launch page says its safeguards include invisible watermarks, C2PA Content Credentials, visible AI labels, prompt blocking, moderation, reporting, and policies against unauthorized use of likeness. That policy line matters more after a demo where one casual voice-reference instruction was enough to make the system speak as the creator.