PhotoLab

If you've ever rendered a portrait and thought "this is a skincare ad, not a person," PhotoLab is aimed squarely at that. The most reliable tell that an image is AI-made has been too-smooth, poreless, plastic skin since 2023, and no checkpoint fully fixes it in one pass. PhotoLab is the post-processing answer: it's a single node that bakes in film grain, compression artifacts, color grading, vignette, and a whole suite of procedural skin effects - pores, freckles, blemishes, acne, subsurface scattering, peach fuzz, even a little T-zone oil shine - so a clean render finally looks lived-in and human. It's the kind of finishing pass the "amateur photography" LoRAs try to teach the model; here it's deterministic and you control every dial.

The name is a lie in the best way: this is not a neural anything. It's pure classic image processing - PIL, numpy, and a couple of blur routines running on CPU in milliseconds. It's not a model, it downloads nothing, and it won't touch your VRAM. That's also what makes it worth keeping in a workflow: you can rerun it a dozen times tweaking skin_seed and it costs you nothing but time.

How it works

Every input image goes through a fixed pipeline, roughly in this order: optional lighting match (first, in CIE LAB space), saturation and color grade, pixelation, the masked skin effects, then film grain, blur, vignette, and finally JPG compression. That compression pass isn't an accident - it saves the image to JPEG into an in-memory buffer and reopens it passes times, so quality at 60 with two passes gives you authentic blocky artifacts, not a filter approximation.

The skin effects are the interesting part. Each one is a different synthesis trick: pores are a high-frequency Difference-of-Gaussians texture layer blended in Overlay mode, peach fuzz is procedurally drawn micro-strokes, subsurface scattering is a low-opacity orange-pink blurred layer Screen-blended in. Before applying anything, the node samples the central 40% of the image in LAB space to estimate a Fitzpatrick skin-tone bucket (1–6) and scales the melanin-aware effects (freckles, blemishes) to match. That's genuinely thoughtful - the same freckle setting on a fair face and a dark face doesn't look identical.

The inputs that matter

You don't touch most of these on day one. The node ships with collapsible Global Photo Effects and Face Skin Effects preset buttons in its UI - Film Snapshot, Darkroom B&W, Golden Hour, plus face presets like Natural Skin and High-Detail Skin - and you can save your own to localStorage. Start there.

When you go manual, three things matter most:

• mask_mode - "Face Only" applies skin effects inside a connected face_mask, "Inverted" applies them outside (for body skin), and "Disabled" skips all skin effects so the node becomes a pure global-effects pass. If a face preset seems to do nothing, this is the first place to look.
• skin_seed - one seed drives every procedural pattern (pores, freckles, spots, fuzz). Same seed + same image = same face every time; bump it to reshuffle blemishes.
• quality / passes - your compression-artifact dials. Lower quality, more passes, more degradation.

The two inputs worth wiring from elsewhere: reference_image (only read when lighting_match_mode isn't Disabled - pick Histogram (L-channel) as the best all-rounder for matching another photo's light and shadow) and face_mask, typically from an Impact Pack or similar face-detect node. Outputs are images and a passthrough face_mask (full white if you never connected one), so you can reuse the same mask downstream without a second mask node.

Installing it

Here's the gotcha: the PhotoLab repo's own README is now just a stub that says moved to ComfyUI-ialhabbal, install the suite instead. The author folded PhotoLab into that larger suite, so don't chase the standalone repo - install the suite:

cd ComfyUI/custom_nodes
git clone https://github.com/ialhabbal/ComfyUI-ialhabbal

or search ComfyUI-ialhabbal in ComfyUI Manager. The old ialhabbal/PhotoLab path still installs and works if you really want it, but it's no longer where updates land.

Where people get burned

The one real trap is the lighting match doing nothing. It relies on scikit-image, which isn't in the pack's dependency list - if it's missing, the node prints a console warning and silently skips the match. If you set lighting_match_mode and see no change, pip install scikit-image and restart. Everything else is just sliders being at 0 - which, in this node, means "off," by design.