Skip to content

DOCKER.md

Latest commit

 

History

History
102 lines (80 loc) · 5.81 KB

DOCKER.md

File metadata and controls

102 lines (80 loc) · 5.81 KB
Raw

Dockerized ComfyUI test environment

A reproducible, CPU-only ComfyUI with this pack installed, so the OCIO nodes can be driven programmatically and the color math can be verified end-to-end. Builds native arm64 on an Apple-silicon (M1/M2/M3) Mac and amd64 in CI from the same docker/Dockerfile. No GPU and no ML model downloads are required — the OCIO nodes are pure color math.

OCIO config: the image bundles the classic ACES 1.0.3 config (sparse-cloned from imageworks/OpenColorIO-Configs, config + LUTs, ~a few MB) and sets $OCIO to it, so every node uses it. Set OCIO="" (or override the env) to fall back to the opencolorio wheel's built-in ACES 2.x config — the test resolves colorspace/display/view names from whichever config is active, so both work.

Quick start

docker compose build                 # build the image
docker compose run --rm roundtrip     # the flagship round-trip color-accuracy test (below)
docker compose run --rm test          # standalone tools/test_*.py + node-registration smoke
docker compose up comfyui             # interactive headless server at http://localhost:8188

./data on the host is bind-mounted to /data in the container, so the downloaded input and the written outputs land in ./data/input and ./data/output where you can inspect them.

The round-trip color-accuracy test

docker compose run --rm roundtrip downloads the Kodak "Marcie" test image, feeds it through a real ComfyUI graph built from the OCIO nodes, round-trips it back to the starting space, and compares the output to the input with OpenCV histograms (cv2.calcHist / cv2.compareHist) plus a max/mean absolute pixel-error backstop. Read and write are done in raw 32-bit float EXR (no implicit colorspace conversion on IO, no fp16 quantization) so the measurement isolates exactly the node chain under test.

Three chains run by default. Marcie is HDR (scene-linear values up to ~4.35), which is the whole point: it separates the color math (invertible) from the display-range encodings (which clip):

Chain Path Expectation
colorspace (gated) ACEScg → ARRI LogC → linear Rec.709/sRGB → LogC → ACEScg, via OCIOColorSpace Perfect — histogram correlation ≈ 1.0, max pixel error ~1e‑6. The "color math is perfect" proof. Uses the linear Rec.709/sRGB primaries (Utility - Linear - Rec.709), which invert exactly even for HDR values >1.
srgb (info) same, but through the display-range sRGB texture encoding (Input - Generic - sRGB - Texture) Residual expected. The sRGB texture encoding clamps to [0,1], so HDR values >1 are crushed and can't round-trip. A property of the encoding, not the math.
display (info) ACEScg → LogC → the ACES display / sRGB view (fwd) → (inverse) → LogC → ACEScg Residual expected. The ACES output view (RRT+ODT) tone-maps and clips, so it's not losslessly invertible.

Only colorspace gates the exit code (GATE_CHAINS=colorspace); srgb and display are reported to show where display-range encodings lose the HDR highlights. Colorspace / display / view names are resolved at runtime from the live OCIO config (docker/ocio_names.py), so the test works on the bundled ACES 1.0.3 and the built-in ACES 2.x alike.

Representative result (ACES 1.0.3, marcie_whacked.exr):

[GATED] colorspace : PASS  corr=1.000000  maxpix=4.29e-06
[info ] srgb       : FAIL  corr=0.240     maxpix=3.35
[info ] display    : FAIL  corr=0.244     maxpix=0.031

Configuration (env vars)

Var Default Meaning
INPUT_IMAGE_URL sobotka marcie_whacked.exr Direct-download URL for the input EXR.
CHAINS colorspace,display Which chains to run.
GATE_CHAINS colorspace Which chains' PASS/FAIL sets the exit code.
TOL_CORR 0.99999 Min histogram correlation to pass.
TOL_PIX 1e-4 Max absolute pixel error to pass.
RUN_ACCURACY 0 1 also runs the tools/accuracy suite in the test service.
OCIO bundled aces_1.0.3/config.ocio Path to the OCIO config; set "" to use the wheel's built-in ACES 2.x.
COMFYUI_REF (build arg) v0.3.68 Pin a different ComfyUI version.
OCIO_CONFIGS_REF (build arg) master Branch/tag of imageworks/OpenColorIO-Configs to pull the ACES 1.0.3 config from. 
# Example: only the invertible control, tighter pixel tolerance
CHAINS=colorspace TOL_PIX=1e-5 docker compose run --rm roundtrip

Using the official AMPAS Marcie

The default is the verified, directly downloadable marcie_whacked.exr. The canonical AMPAS DigitalLAD.2048x1556.exr (true ACES2065-1) is only distributed via a Dropbox folder link that can't be curl'd cleanly — to use it, download it yourself and drop it in as ./data/input/marcie.exr (the job uses an existing input as-is and skips the download).

Interactive / API use

docker compose up comfyui
curl -s localhost:8188/object_info | grep -o '"OCIOColorSpace"'   # confirm the nodes are live

You can then POST API-format graphs to http://localhost:8188/prompt (see docker/build_workflow.py for how the round-trip graphs are constructed).

Files

  • docker/Dockerfile — the environment (ComfyUI + pack + ffmpeg + CPU torch).
  • docker/comfyui_client.py — stdlib-only ComfyUI HTTP-API client.
  • docker/ocio_names.py — resolves colorspace/display/view names from the live OCIO config.
  • docker/build_workflow.py — builds the API-format round-trip graphs.
  • docker/compare_histograms.py — OpenCV histogram + pixel-error comparison.
  • docker/roundtrip_test.py — the flagship orchestrator.
  • docker/run_tests.sh — standalone tests + node-registration smoke.
  • docker-compose.yml, .github/workflows/docker-tests.yml — local + CI wiring.