Seedance 2.5 Review: 30-Second Clips, Voice Casting, and Morphing Bugs
A hands-on look at Seedance 2.5's 30-second generations, omni-reference prompting, and voice quirks, based on producing a real short film.
What is Seedance 2.5 and what’s new in it?
Seedance 2.5 is the latest version of the AI video generation model available through Dreamina, best known for extending generation length up to 30 seconds and expanding its “omni reference” system to support up to 50 reference images in a single generation. Early hands-on testing, including a full short film produced almost entirely in the new 30-second mode, suggests it’s a meaningful step up from Seedance 2.0 in acting and scene performance, though it comes with new prompting quirks, voice inconsistencies, and occasional morphing artifacts that creators should plan around.
TL;DR
- Thirty-second generations turned out more useful than expected for dialogue-heavy scenes, cutting down on retakes even though each generation costs more.
- Omni reference with as few as three images (two character portraits and one location shot) was enough to carry an entire two-minute short film with consistent characters and setting.
- Prompting rules changed meaningfully between 2.0 and 2.5. Prompts that worked well on the previous model can produce broken, glitchy output on 2.5, and the new model resists tightly cut, rapid-edit prompting in favor of scenes that “breathe.”
- Voice consistency remains unpredictable. The model appears to cast a voice based on what it infers a character should sound like from reference images, sometimes assigning accents (a British accent showed up unprompted) that require explicit prompt correction like specifying “American” rather than “American English.”
- Morphing and decoherence still show up in fast action sequences, even with well-structured prompts, and no amount of upscaling removes it. This mirrors issues Seedance 2.0 had shortly after launch, which were later patched.
- Output resolution is capped at 720p for now, which makes a secondary upscaling pass (tested here with the standard, non-creative Topaz Desktop upscaler) a practical necessity for finished work.
- Cost runs meaningfully higher than traditional production would suggest at small scale, but is still a tiny fraction of what a comparable live-action shoot would cost.
How well does Seedance 2.5 handle acting and dialogue?
The core test in this round of hands-on use was whether Seedance 2.5 could hold a scene, act through dialogue, and maintain performance quality across a longer take. For a neon-noir short film built around a tense bar conversation between a hitman and a personification of Death, the model handled dramatic pacing and character performance well enough to carry the whole piece. The 30-second generation window turned out to be particularly suited to “talky” scenes like this one, where characters need time to deliver lines and react, rather than rapid-fire action cuts.
That said, longer generations introduce a mixed bag inside a single take. In one shot, a character’s book prop rendered at a comically oversized scale, more like a suitcase than a book, and it took several attempts to prompt the issue away. The workaround that emerged: with 30 seconds of footage per generation, even a flawed take usually contains multiple usable segments, so editors can mine the good parts rather than discarding the whole clip.
How does omni-reference prompting work in practice?
Omni reference lets creators feed the model images of characters, props, or locations to keep them visually consistent across generations. For the short film, the entire visual identity of the project rested on just three reference images: one Midjourney-generated portrait for each of the two main characters and a single environment shot for the bar setting. No detailed character sheets or location maps were needed.
Seedance 2.5 also expands this system to support up to 50 reference images at once, which was stress-tested by feeding the model an entire wardrobe of outfits for a recurring test character to see how well it could vary clothing while keeping the character recognizable. The result held up well across multiple outfit changes in a single generation, suggesting the expanded reference limit is genuinely useful for wardrobe or prop-heavy scenes, not just a marketing number.
A more experimental technique combined first-frame image-to-video with a detailed omni reference, feeding the model a full set of first frames for a sequence alongside reference images. The output largely delivered the intended shots, but not always in the correct order, indicating that sequencing may need to be reinforced directly in the prompt text rather than assumed from the order of inputs.
Why does Seedance 2.5 need different prompts than 2.0?
One of the clearest findings is that prompts optimized for Seedance 2.0 don’t reliably transfer to 2.5. Running an established 2.0 prompt through 2.5 produced visibly broken, glitchy results in at least one repeated test case. Seedance 2.5 also appears to resist prompt styles built around rapid, one-second-style cuts. Scenes prompted to breathe a little longer, with less crammed into each beat, produced noticeably better output than prompts trying to replicate quick-cut editing within the generation itself.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Practically, this means prompts for 2.5 tend to run longer and require more explicit structure than before. Official documentation (read and distilled with the help of Claude) provided the backbone for a more technically correct prompting structure, since dense “wall of text” prompts, while functional, aren’t actually the recommended approach for the model.
What’s going on with voice consistency and accents?
Voice behavior in Seedance 2.5 is one of the murkier areas. The working theory from hands-on testing is that when a generation relies on image references or omni references, the model infers a voice for the character based on what it thinks that character should sound like, rather than pulling from any fixed voice profile. This can be directed, but it’s not fully predictable.
In one test, a character was unintentionally given a British accent despite no such instruction. Prompting the character to speak “American” fixed it, but prompting “American English” did not, apparently because including the word “English” nudged the model back toward a British-sounding delivery. It’s a small but telling example of how sensitive voice casting is to specific wording rather than broad intent.
What are the current limitations and bugs?
Two issues stand out as the model’s rough edges right now. First, decoherence and morphing still occur, particularly in fast action sequences, showing up as a kind of visual spikiness or instability in character models. This happened even with carefully structured, documentation-based prompts, doesn’t occur throughout an entire generation, and can’t be cleaned up with upscaling, creative or otherwise. Seedance 2.0 had similar stability issues after its initial release that were addressed through subsequent patches, so this is expected to improve rather than remain a permanent ceiling.
Second, output currently caps at 720p resolution. The practical workaround is generating at 720p and running a standard upscale afterward (tested with Topaz Desktop’s regular upscaler rather than its creative mode) rather than trying to force higher native resolution out of the model.
Features like SmartEdit and longer video functions weren’t available during this testing window, so they remain untested for now.
Is Seedance 2.5 worth the cost?
Cost figures are approximate given the usual fog around credits, API pricing, and platform markups, but rough estimates for the short film’s production (about 15 generations at 30 seconds each, totaling roughly 450 seconds of raw footage for a final 2-minute-22-second cut, a 3.2-to-1 shoot ratio) landed somewhere between $90 and $355 depending on whether pricing is calculated at raw cloud API rates or marked-up platform rates.
For comparison, a traditionally shot equivalent, with a full crew and production budget, was estimated at $55,000 to $85,000, with a scrappy indie version still running $20,000 to $35,000. Even at the higher end of AI generation cost estimates, the gap is enormous. The 30-second generation mode specifically is worth using deliberately rather than by default: it’s well suited to dialogue-driven scenes but adds cost, so matching generation length to the actual needs of a shot matters more than defaulting to the maximum.
Frequently Asked Questions
What platform was used to test Seedance 2.5?
The testing and short film production were done through Dreamina, which offers access to Seedance 2.5 for generation.
How many reference images does Seedance 2.5 support?
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Up to 50 images can be used in a single omni-reference generation, though the tested short film used only three: two character portraits and one location image.
Does Seedance 2.5 generate video in full HD?
No, native output is currently capped at 720p. A common workaround is generating at 720p and upscaling afterward with a separate tool.
Why did a character end up with an unintended accent?
Seedance 2.5 appears to infer voice characteristics from image references when no explicit voice direction is given, which can produce unexpected accents unless corrected with specific wording in the prompt.
Are Seedance 2.0 prompts compatible with 2.5?
Not reliably. Prompts built for 2.0 can produce glitchy or broken results on 2.5, which favors longer, more deliberately paced prompts over rapid-cut style instructions.
