PUBLIC COPY: local account and storage paths are placeholders; configure them before running. IP fields in reports use example addresses. Original benchmark timings are unchanged.

MiniMax H3 — Mac Heretic MLX Q4 + Ubuntu RTX PRO 4500

These are the actual 3D and 2D benchmark graphs: 37 nodes, 47 links each.
Use the *workflow.json files in ComfyUI. The *api.json files are API format.
PNG diagrams are rendered from the graph, not screenshots; their workflow metadata can also be imported.

Requirements: the existing H3 models, KJNodes, T8 audio pack, latent upscaler, six input images and experimental bridge node. This package does not include model weights or reference photos.

Experimental bridge source:
Ubuntu: copy bridge/__init__.py, node.py, wire.py into ComfyUI/custom_nodes/h3_mlx_reference_trial, then restart your ComfyUI.
Mac: server.py and wire.py must be adjacent. Run server.py manually with the Python environment in /path/to/storage/Backups/codex/i2v_minimaxh3/.venv.
The server depends on that backup's app and vendor libraries, and the existing Heretic Q4 model in experiments/heretic-int8-mlx/target/model. Paths are specific to this Mac.
SSH reverse tunnel: Ubuntu 127.0.0.1:8189 -> Mac 127.0.0.1:8787.
No autostart; stop your own foreground processes when done. Do not disable memory guards.

The workflow writes a native 294-frame, 1344x768, 24fps video (~12.25 seconds).
For the delivered exact 12-second version, trim to 288 frames with ffmpeg:
ffmpeg -i native.mp4 -t 12 -vf fps=24 -c:v libx264 -pix_fmt yuv420p -c:a aac -movflags +faststart final.mp4

Measured generation time (encoding and roundtrip included):
3D: 303.54 seconds; Mac encoding+roundtrip 28.3341 seconds.
2D: 292.31 seconds; Mac encoding+roundtrip 28.9209 seconds.
Excludes Mac preload/settling, download and final trimming.

This is an experimental integration for this two-machine setup. The bridge node was archived outside custom_nodes after the run. Importing the JSON alone does not install it or start the Mac encoder.
