I would like to bring up Seedance 2.5, a multimodal audio‑video generation model developed by ByteDance’s Doubao team, for technical discussion on this list. This note shares practical observations from exploratory testing rather than commercial promotion; further technical details are available at https://seedance25ai.cc/. Many existing open‑ and closed‑source text‑to‑video systems struggle with short maximum clip length, weak multi‑reference control, and the requirement for full re‑generation when adjusting small parts of a sequence. Seedance 2.5 addresses several of these pain points. It supports single‑pass generation of clips up to 30 seconds, doubling the 15‑second limit of its prior iteration. Longer native output reduces manual stitching of many tiny fragments and makes it feasible to render complete short narrative beats within one generation run. Its multimodal reference capacity is expanded to a combined total of 50 assets: up to 30 images, 10 video snippets, and 10 audio files. Notably, audio‑only reference is supported; users can feed music or sound tracks to guide motion rhythm, beat alignment and lip‑sync behaviour without image or video inputs. References can constrain character styling, product appearances, scene aesthetics and camera movement, though exact visual replication is not guaranteed. Another practical feature is targeted timestamp‑based local editing. Instead of re‑rendering the whole video to fix minor defects, users can modify selected regions or time segments while keeping original camera motion, lighting and audio intact. This lowers iteration overhead during creative trials. The model natively works with more than ten human languages for prompt input, which is useful for cross‑language generative‑video workflows. Of course, the model still has known limitations. Artifacts can appear in complex multi‑character crowd scenes. Real human faces are disallowed as reference inputs; only virtual or AI‑generated character references are permitted. Commercial usage is bound by platform terms of service, so anyone building production work needs to review those carefully. I am curious whether others on this list have experimented with Seedance 2.5 or comparable closed‑source video‑generation models. Have you found the multi‑reference or local‑editing capabilities interesting for your own prototyping? Are there specific shortcomings you would highlight? Feel free to share experiences or comparative viewpoints.