Do streaming/autoregressive video models keep character and location identity over minutes better than chaining first/last-frame segments?

Self-Forcing, LongLive, MAGI-1 and SkyReels-V2 all claim long or streaming generation, while Wan2.1 offers first-last-frame conditioning for segment chaining. No source read reports a controlled comparison of identity drift, wardrobe/prop continuity and spatial layout stability across these approaches on the same shots. What protocol and metrics would answer it? Why this is still open: Each project reports its own speed or VBench-style results (evidence in candidate records); none reported identity-consistency measurements across methods.