Seedance 2.5 is live, and the headline for us isn't just the 30 seconds generation, it's that we got this in 2 generations. @seedance We've been building freeze-frame and dialogue-driven shots for a while now and getting used to burning 10-15 attempts before something is usable, wrong eyelines, physics that drift, camera moves that ignore the prompt. This interrogation scene was usable on the first try, and I ran one more just to compare. That's the whole cost conversation right there: the compute per usable shot dropped, not just the compute per generation. It cuts a lot of timer off what this normally takes. What actually held up: the suspect's micro-hesitation before he reaches for the photo, the way the practical overhead light falls off across the table, and camera movement that feltchosen rather than random drift. It stayed coherent across a single unbroken 30-second take with real dialogue, no stitching, no seams to hide. We used Seedream for key frame planning, which made the reference alignment much smoother.Also worth knowing: multimodal references up to 50 assets, localized editing so you can fix one region instead of rerolling the entire shot, and multilingual support. Prompt 👇 Photorealistic psychological crime drama set in a police interrogation room. A detective presents crime scene photographs to a suspect. When one photo accidentally flips upside down, the suspect instinctively turns it the right way, unknowingly revealing knowledge only the culprit would have. The detective quietly realizes she has caught him. Seedance 2.5 API is about to launch on @BytePlusGlobal, and now you can try it directly on @lumina_ai_aiart.