Wan2.2 makes short clips, trails frontier models

Hot take: open video gen is no longer a toy, but it is still an assistant for making fragments, not a replacement for a top private model when you need polished, coherent scenes.

• –Wan2.2 is a meaningful open release: the repo documents T2V/I2V/TI2V models, 720p support, 24 fps for TI2V, and ComfyUI/Diffusers integration, so the ecosystem is real and improving.
• –The user’s 6-10 second ceiling is normal; short duration is still where most local models are strongest, especially when you want something you can actually steer.
• –Prompt extension matters: Wan’s own docs recommend it, and in practice it helps more than adding extra adjectives manually.
• –Best local use cases today are:
• –stylized b-roll
• –image-to-video motion from a strong keyframe
• –character animation/replacement
• –stitched montage-style sequences with repeated look and camera language
• –What frontier private models still do better:
• –longer temporal coherence
• –better physical plausibility
• –stronger prompt following
• –cleaner motion, faces, hands, and camera transitions
• –native audio and lip sync
• –What is feasible right now locally:
• –5 to 10 second shots with a controlled look
• –storyboards built from multiple clips
• –“history video” style content if you cut around drift and use recurring references
• –What is still out of reach:
• –truly long, single-take narratives with stable identity
• –dialogue-heavy scenes with convincing synced speech
• –complex multi-character blocking without drift
• –consistent scene-to-scene continuity without human editing help