One architecture, every modality

The field is done maintaining a separate model per modality pair. What replaces it is harder to evaluate than what it replaces.

Lance unifies multimodal modelling through multi-task synergy; InstructX routes visual editing through language-model guidance. Seedance 2.0 frames video generation as a world-complexity problem rather than a rendering one.

The motivation is maintenance, not elegance

A separate model per modality pair produces a combinatorial burden: inconsistent behaviour across paths, no shared learning between obviously related tasks, and a release process that multiplies with every new pairing.

Consolidation is an engineering decision before it is a research one, and that is usually a sign it will actually happen.

Multi-task synergy has to be tested, not assumed

Training tasks together sometimes helps and sometimes yields a model that is mediocre at everything. The difference depends on whether the tasks share structure, and that is an empirical question per combination — not a property of the approach.

Routing image editing through language-model guidance is the more immediately practical idea, because it converts editing into instruction-following, and an instruction can be checked against the result.

World complexity is the right target and the hardest to score

Moving from appearance to behaviour — objects persisting, causes preceding effects, physics surviving a cut — is the correct reframing. Resolution and artefact rate stopped being the interesting failures a while ago.

But there is no accepted metric for whether a generated world is internally consistent. Progress on the stated goal is currently assessed by watching, which means it is assessed by whoever is watching.

The measurement literature has spent this year auditing its own instruments. Video generation has not started, and it is arriving at exactly the moment when the thing being generated stopped being checkable by eye.

arXiv — Lance: unified multimodal modeling by multi-task synergy → · arXiv — InstructX: towards unified visual editing with MLLM guidance → · arXiv — Seedance 2.0: advancing video generation for world complexity →