The model layer became broader while the production evidence became longer: FLUX 3 joined four modalities in one architecture, Luma demonstrated ensemble character work, and OVERGROWN documented a 300-plus-shot film process.
Black Forest Labs introduced FLUX 3 as one multimodal architecture, with video entering early access. Unifying media and action can reduce translation loss between ideation, motion and sound, but a broad demo does not guarantee controllable production output. Benchmark reference adherence, timing and edit locality separately before replacing specialized models.
Luma highlighted an ensemble scene whose characters read as a cast with lives beyond the presented moment. Multi-character coherence is harder than a single hero identity because wardrobe, eyelines and spatial relationships must all hold. Test cast workflows with reverses and regrouping shots, not only one tableau, and track which reference owns each character.
Google Flow highlighted Henry Daubrez's 13-minute OVERGROWN, built over six months with more than 300 shots, each beginning as an image. The numbers expose the real cost of long-form AI film: continuity, selection and finishing compound across hundreds of decisions. Teams should budget by approved shot and revision pass, not by generation count, and preserve the first-frame lineage for every clip.
