H3 Martial-Arts and TaoMate Workflows: Read the Schedule, Not the Filename

Two H3 exports showcase the distinction between filename formats and actual execution traces, as well as scheduler configurations without proving speedup. The package features two JSON workflows, one with 113 nodes and another with 239 nodes, counting a total of 396 explicit links along with named Set/Get relationships. The guide focuses on a static walkthrough, not a new inference, audio review or hardware benchmark, with source claims regarding frame rate and 1080p resolution remaining unverified.

The martial-arts graph employs SetNode/GetNode pairs for many inputs, bypassing other image and audio providers, while the TaoMate example collection follows a similar structure with a similar prompt and scheduler. The scheduler receives 20 steps, overriding its saved local value of 10, leading to a sigma schedule passing through H3SigmaRefiner with specific settings.

The selected base model and LoRA models are distinct, with the final stack including model modifications and prompt adjustments, but not an explicit pose-map lock or proof of precise motion transfer. The duration control produces a nominal 10.125-second output, with two sampling stages and two decoders carrying audio and video latents, requiring separate testing to verify the final audiovisual output.