MMnode/bench
2026-09-28 01:43:15 +00:00
..
app feat: Moxie bench - standalone python/web harness for the fixed pipeline 2026-09-28 01:42:58 +00:00
tauri feat: Moxie bench - standalone python/web harness for the fixed pipeline 2026-09-28 01:42:58 +00:00
tests feat: Moxie bench - standalone python/web harness for the fixed pipeline 2026-09-28 01:42:58 +00:00
ui-dist feat: Moxie bench - standalone python/web harness for the fixed pipeline 2026-09-28 01:42:58 +00:00
base_workflow.json feat: Moxie bench - standalone python/web harness for the fixed pipeline 2026-09-28 01:42:58 +00:00
moxie_bench.spec feat: Moxie bench - standalone python/web harness for the fixed pipeline 2026-09-28 01:42:58 +00:00
README.md feat: Moxie bench - standalone python/web harness for the fixed pipeline 2026-09-28 01:42:58 +00:00
requirements.txt feat: Moxie bench - standalone python/web harness for the fixed pipeline 2026-09-28 01:42:58 +00:00
run.py feat: Moxie bench - standalone python/web harness for the fixed pipeline 2026-09-28 01:42:58 +00:00

Moxie Multimedia Bench

A standalone Python + web harness for the Moxie Multimedia pipeline — the first step toward running this node pack outside ComfyUI as a sidecar for a Tauri 2 desktop app.

The bench locks to the fixed pipeline of the base workflow (bench/base_workflow.json):

MoxieMultimediaLoader -> LoraLoaderModelOnly -> MoxiePreviewOverride
  -> easy multiTrackEditor -> MoxieProject -> MoxieCombine
  -> [native branch]   CreateVideo -> SaveVideo
  -> [upscaled branch] RTXVideoSuperResolution -> CreateVideo -> SaveVideo

The multitrack timeline editor is the real thing: the bench bundles the same React editor components the ComfyUI node uses, mounted by a thin shim (frontend/src/bench/) that talks REST + WebSocket instead of the ComfyUI runtime.

Run from source

pip install -r bench/requirements.txt      # fastapi, uvicorn, numpy, pillow
python bench/run.py --port 8787            # auto backend (mock here)
# open http://127.0.0.1:8787

The UI needs a one-time build (a prebuilt copy ships in bench/ui-dist/):

cd frontend && bun install && bun run build:bench

Backends

backend needs behaviour
mock ffmpeg, numpy, pillow deterministic synthetic frames per (prompt, seed, frame), sine audio bed, lanczos "upscale" to the RTX target, h264 encode. Both output branches produce real mp4s.
comfy a running ComfyUI + downloaded models substitutes the plan into the base workflow, converts it to the /prompt API format, submits to ComfyUI, polls progress, downloads outputs.
auto — comfy when reachable, else mock.
python bench/run.py --backend comfy --comfyui-url http://127.0.0.1:8188

HTTP API

  • GET /api/health, GET /api/config — liveness, backend availability
  • GET /api/workflow — base workflow summary + editor track_data
  • GET /api/plan — resolved render plan (segments, prompts, frames)
  • POST /api/render — queue a render: {"track_data": <editor JSON>, "overrides": {"width": 1920, "height": 1080, "steps": 4, "seed": 123, "project_name": "MM-x", "backend": "auto"}}
  • GET /api/jobs, GET /api/jobs/{id}, DELETE /api/jobs/{id}
  • WS /api/ws — push events: job_queued, execution_start, job_progress, execution_success, execution_error (mirrors ComfyUI's push model)
  • GET /api/view?filename=&type=output&job= — fetch rendered files
  • ComfyUI-compatible aliases: /view, /upload/image, /easy-media/*

Outputs land in bench/work/outputs/<job_id>/ (MOXIE_BENCH_WORK overrides the root). Every render produces both branches: native (editor canvas) and upscaled (RTX target dimensions).

Tauri 2 sidecar

See bench/tauri/README.md — PyInstaller onedir packaging (bench/moxie_bench.spec), tauri.conf.json resource wiring, spawn + health poll snippet, and the endpoint table.

Tests

python3 -m pytest bench/tests -q        # contract tests (no ComfyUI needed)
python3 scripts/test_bench_pipeline.py  # full API + WS + mock render e2e
cd frontend && bun run test             # editor suite incl. bench shim

Phase map

  • phase 2 (this) — bench: fixed pipeline, dual backend, timeline UI port, sidecar packaging.
  • phase 3 (planned) — in-process execution backend: load the pack's node classes directly (torch + ComfyUI modules on sys.path) instead of driving a separate ComfyUI server; bench/app/backends/comfy.py keeps the same interface.