How does it handle large outputs? If I pipe a big config file through, does the full content travel back to local for LLM processing, or is there any streaming/chunking?
curious about the first-frame / last-frame guided generation. That's been the weakest part of most video models — you either get smooth interpolation with no creative control, or full control with ugly artifacts. If 2.7 actually nails this, the instruction-based editing on top of it could be a real workflow shift. Any word on what resolution and length they're targeting?