AfterDark Film Grain & Emulsion

The thing that separates "film look" from "grain filter from 2010" is that real grain is not a flat overlay. It's denser in the shadows and midtones, sparser in bright skies, and it varies with the stock - Tri-X is a different texture than Portra, which is different from Velvia. AfterDark Stochastic Noise is the node in this pack where the author clearly spent the most time: 32 named film emulsions (Portra 400, Tri-X 400, Velvia 50, CineStill 800T, Kodachrome 64, Polaroid 600, and a couple dozen more), each wired to grain size, noise type, color behavior, and warmth. Pick a stock, and the look follows.

It's in the HackAfterDark category. This is your "make it read as captured, not generated" layer - the single cheapest realism move there is, and one the community has converged on: a subtle grain pass does more for believability than most prompt tweaks.

How it works

Every preset is a dict of parameters - noise distribution, base grain size, jitter, chromatic aberration, warmth, noise level, gamma shift, softening, and a luminance response. A film_format scaler (35mm / medium / large format) shrinks grain for the bigger negatives. The genuinely nice part is the luminance response: instead of uniform noise, it weights grain by local brightness - Film (Midtone & Shadow) peaks grain in shadows and midtones with a soft roll-off in highlights (default), Digital (Shadow Heavy) concentrates it in the darks like sensor read noise, and Uniform (Flat) spreads it everywhere.

Four noise distributions back it: gaussian (standard), poisson (shot noise, signal-dependent), multiplicative (dye-cloud clumping), and laplacian (heavy-tailed halide spikes). Beyond grain, the node layers on the "anti-detector" extras - sub-pixel micro_jitter, a down/up spatial_resample that strips VAE lattice artifacts, gamma_shift, and edge_softening. Treat those as the experimental half: the mainstream win is the grain itself.

Inputs that matter

You can get 90% of the value from three knobs:

• film_preset - pick a stock. Choose None (Manual) and the manual widgets below become your controls.
• noise_level (0–0.2, default 0.025) - keep it 0.01–0.03 for realism; higher is heavy texture.
• luminance_response - leave Film (Midtone & Shadow) on unless you specifically want the digital-shadow or uniform look.

grain_size, noise_type, channel_mode, tone_warmth, and the anti-detector group are all exposed for manual tuning, plus an optional seed. Output is a single image (IMAGE), after VAE decode and before save. Nice touch from the pack's front-end: picking a preset auto-fills every manual widget, and touching any of them flips you back to manual mode.

Installing it

Same one-time setup as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/hackafterdark/ComfyUI-HackAfterDark-Nodes

Restart ComfyUI, or install via ComfyUI Manager by searching "HackAfterDark" (pack title AfterDark Film AR Selector). No models to download, no API keys. The google-generativeai in the pack's requirements is for its Gemini node only - this one is pure torch.

Where people get tripped up

Two things. First, the defaults on the node are the manual values - a preset overwrites them when you select it, which surprises people who tweaked a slider first and then picked a stock. Set the preset, then fine-tune. Second, don't crank everything at once; it's a 13-parameter node and they interact (the resample, jitter, and gamma shifts stack). Pick a stock, set noise_level, leave the rest alone, and you'll get the film look without turning your image into static. And if a stock reads too heavy, that's normal - the presets are tuned loud, and dialing noise_level down is the intended workflow.