Emulsion

ComfyUI-Emulsion

JSON-preset color grading, Lightroom-preset style. Two nodes under image/postprocessing:

• Emulsion Grade applies a preset in a single pass: levels, tone curve, contrast, white balance, HSL, .cube LUT, clarity, bloom, halation, vignette, grain, and more.
• Emulsion Extract does the reverse: from a before/after pair, it recovers the grading parameters by optimization and writes a JSON preset reusable by Emulsion Grade.

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/24latents/ComfyUI-Emulsion.git

Restart ComfyUI. No dependencies to install (torch is provided by ComfyUI).

Usage

Add the Emulsion Grade node, connect an image, pick a preset:

• preset — a file from presets/*.json. The folder is rescanned on browser refresh: no restart needed for a new preset to show up.
• strength — 0 = original image, 1 = full grade, > 1 = extrapolation (clamped to 0-1).
• seed — reproducible grain.

On an invalid preset or a missing LUT, the error is logged to the console and the image passes through unchanged (the workflow never crashes).

Emulsion Extract

Connect image_source (ungraded) and image_target (the same image, graded); the node fits by gradient descent (Adam, pure torch, CPU or GPU) a 6-point tone curve per RGB channel, saturation and an optional vignette, then estimates grain from the residual high-frequency noise. Per-channel curves capture color casts — e.g. a vintage look that lifts red blacks while crushing the blue channel; when the three fitted curves agree, they are collapsed into a single shared curve in the written preset.

• preset_name — the "name" stored in the JSON; the file itself is written under a slugified version of it (My Look! → my_look.json).
• iterations — optimization budget (default 400). More iterations = tighter fit but slower. The fit stops early on its own once the error stops improving for 50 iterations, so high values only cost time when they actually help: ~100 for a quick draft, 400 for most grades, 800-2000 for stubborn ones. The console shows the final RMSE and the number of iterations actually run.
• fit_vignette — also fit a vignette (amount, radius, feather). Turn it off when the pair has no vignette, so the optimizer cannot abuse it to compensate for other differences (e.g. a naturally darker corner). The preset is written to presets/<slugified_name>.json (with a _2, _3… suffix on collision) and returned as a STRING; it shows up in the Emulsion Grade dropdown on browser refresh. The JSON only contains the significantly non-neutral parameters, plus the achieved RMSE in description. The two images may have different resolutions (e.g. an upscaled target) as long as the aspect ratio matches: the target is resampled to the source before fitting.

Technical note: during the fit, the tone curve is evaluated with piecewise linear interpolation between the 6 points (differentiable); at render time, grading.py applies a Catmull-Rom spline through those same points. This controlled linear-fit / spline-render gap is on the order of the fit noise (the reported RMSE is measured on the spline render, at native resolution).

Example workflows

Two ready-to-use workflows ship in example_workflows/ — they also show up in ComfyUI's template browser (Workflow → Browse Templates → ComfyUI-Emulsion):

• emulsion_grade.json — one image graded with four presets side by side (screenshot above).
• emulsion_extract.json — extracts a preset from a before/after pair; load your own two images into the LoadImage nodes. The Display Any node (from rgthree-comfy) only shows the returned JSON and can be removed if you don't have it installed.

Preset format

All fields are optional (absent = neutral). The full schema is documented in the grading.py docstring. Minimal example (presets/moody_garden.json):

{
  "name": "Moody Garden",
  "levels": {"input_black": 57, "input_white": 217, "gamma": 1.04,
             "output_black": 38, "output_white": 210},
  "contrast": 1.15, "contrast_pivot": 0.5,
  "saturation": 0.96,
  "grain": {"amount": 0.008, "size": 1.0, "mono": true},
  "strength": 1.0
}

"tone_curve" accepts either a single list of points (applied to all channels) or a per-channel form {"R": [...], "G": [...], "B": [...]}; out values may go below 0 / above 255 to crush or clip a channel.

.cube LUTs referenced by a "lut" field are looked up in presets/luts/.

Tests

pytest tests/

The real-image regression tests expect tests/A.png and tests/B.PNG (not versioned); they are skipped when absent.

License

MIT — see LICENSE.