Luma adds ByteDance's Seedance 2.5 for 30-second multi-shot video
By Steven Van ·
Luma Agents can now drive Seedance 2.5 with up to 50 text, image, video and audio references per generation.
Luma AI has added Seedance 2.5 to its platform, bringing ByteDance's new flagship video model into the same workspace as Luma's own Ray models and its agentic creative tools. The headline numbers: up to 30 seconds of multi-shot, high-definition video in a single generation, guided by up to 50 references spanning text, image, video, and audio. And because it lives inside Luma Agents, you can direct scenes, camera angles, and pacing conversationally instead of wrestling with one giant prompt.
Thirty seconds of multi-shot video in one generation
The multi-shot part matters as much as the length. Plenty of models can now hold a single continuous take, but Seedance 2.5 handles editorial structure inside one generation: cuts, angle changes, and scene transitions that read like intentional filmmaking rather than a camera left running. Thirty seconds of that is enough for a trailer beat, a full product story, or a short narrative sequence with an actual arc. ByteDance built the model to generate video and audio together in a single pass, and generations can be extended across multiple rounds when a project needs more runtime.
Fifty references across text, image, video, and audio
Consistency is the other half of the release. Seedance 2.5 accepts up to 50 references in a single generation, and on Luma that spans all four modalities: text, image, video, and audio. According to ByteDance, the mix supports up to 30 images, 10 video clips, and 10 audio files. That is enough material to pin down a protagonist's face, a supporting cast, a location, a color palette, and a sonic direction before generating, which is exactly what multi-shot storytelling demands. A character who looks different in every cut breaks the illusion; references are how you keep them on model across all of them.
Luma Agents put a director in the loop
What separates Luma's integration from a raw API is Luma Agents, the agentic creative platform Luma launched earlier this year to wrap its own models alongside partner models. With Seedance 2.5 plugged in, you can tell the agent what you want changed: tighten the pacing, move the camera lower, hold the wide shot longer, and it handles the model wrangling. That turns video generation from prompt roulette into something closer to a director-editor conversation. For creators who think in shots and beats rather than prompt syntax, this is the most natural way yet to drive a model of this size.
Luma's bet on being model-neutral
Luma built its reputation on its own technology, from 3D capture to the Dream Machine and Ray video models. Adding ByteDance's flagship alongside them signals a pragmatic stance: give creators the best available model for each job and compete on the workflow around it. Seedance 2.5 is landing on several platforms this week, but the combination of 50 multimodal references with agent-driven scene and camera control makes Luma's version particularly interesting for narrative work. If you have been waiting for AI video that can hold a story together across cuts, this is one to test.
