A READING PATH / 2025–2026
Move from the source record to a production decision. 5 researched reads across two subject years.
2025 SUBJECT / Model research
Wan2.1’s 2025 release gave filmmakers code and weights, not a guarantee that an unprepared local machine is a production pipeline.
Reader question: What should a filmmaker establish before treating an open video model as a repeatable local workflow?
Read the evidence ↗2025 SUBJECT / Model research
Wan2.1-VACE combined creation and editing tasks in 2025. Its useful production lesson is to label every condition instead of treating inputs as interchangeable inspiration.
Reader question: How should a crew organize text, images, video and masks when an open model accepts several kinds of condition?
Read the evidence ↗2025 SUBJECT / Model research
Tencent’s 2025 I2V release and its early identity-fix record offer a useful warning: a reference frame is an anchor, not a promise.
Reader question: How should a filmmaker use one image reference without assuming character consistency is solved?
Read the evidence ↗2025 SUBJECT / Model research
HunyuanCustom’s 2025 paper is a research claim about multimodal conditioning; a filmmaker still has to decide what counts as the same subject in a scene.
Reader question: How can a production define subject consistency before using a multimodal customized-video workflow?
Read the evidence ↗2025 SUBJECT / Model research
Tencent’s 2025 Avatar paper describes multi-character, audio-driven animation. The responsible craft response is to bound speech, speaker and scene before generation.
Reader question: What should be locked before a team uses an audio-driven character-animation model for dialogue?
Read the evidence ↗