Photoshoot Stil

The most-asked question after ComfyUI-Photoshoot's 2.2.0 release wasn't about poses or framings. It was "how do I get black and white?" - and the correct answer, type black and white photograph, monochrome into the style box, was unhelpful, because nothing in the kit told you so and a bare "b&w" loses against a person block full of colour words at CFG 1. Krea2StyleBuilder (canvas name Photoshoot Stil) is that question turned into a dropdown. It composes the whole photographic look - colour or black and white, genre, lens, finish - into an English phrase your sampler can actually read.

What it actually does

Like every node in this pack, it emits text, nothing else. The author renders his posters against Krea 2 Turbo (8 steps, CFG 1, no LoRA), so the style words here were measured, not guessed: 10 black-and-white and 7 colour looks, each verified to land differently, plus 13 genres, 10 lenses and 8 finishes.

The interesting engineering is what the measurements taught him. Film stocks named as such - "shot on Ilford Delta 3200" - produced no grain at CFG 1, while described properties ("extremely grainy black and white photograph, heavy coarse pushed film grain") did. So the looks ship as what they look like, with the famous name as a hint: the Tri-X entry is worded as visible grain, the Leica one as crisp grain-free micro-contrast, the "Selenium" print as a cool-toned finish. Teal-and-orange, cross-processed and the rest got cut entirely - they rendered as the neutral image. What survived moves brightness, saturation or warmth by a measurable margin.

The phrase is assembled in a deliberate order - look, genre, lens, finish - because the look has to come first. "Copper red hair" and "green eyes" are colour words too, and a style competing with them only wins from the front of the prompt. The author measured that too: the same black-and-white wording at the end of the prompt still produced a colour image; at the front it took.

The inputs that matter

The graph-exposed input is just seed (INT) - the dropdowns you actually use live on the node's own panel and are carried in a hidden state field. Here's what's on that panel, in order:

• Look - the big one. Pick a black-and-white family member (classic, high contrast, film noir, grainy, high-key, coarse grain, fine & grain-free, infrared, sepia, selenium) or a colour one (Portra, Cinestill 800T, Kodachrome, Polaroid, warm golden, cool desaturated, pastel).
• Genre - editorial through paparazzi, passport photo and film still included.
• Lens - 85mm f/1.4, 50mm, 35mm, 135mm, 24mm, macro, anamorphic, tilt-shift, vintage, or f/8 deep focus.
• Finish - natural skin, grain, halation, vignette, soft focus, motion blur.

Defaults are Editorial + 85mm f/1.4 with no look and no finish - deliberately, that's exactly the text the example workflow used to carry by hand. Each field has a die; roll one and the value comes from a seed-based generator (random.Random(f"{seed}-{field}")), so a rolled look is reproducible run-to-run, and if you've chosen a black-and-white family, the roll stays inside that family.

Output is one stil STRING. Wire it into the style input of the pack's Photoshoot Build Prompt node and drop a {style} (or {stil}) placeholder in your prompt text - and put it at the front, as the README's recommended order does.

Installing

Shared with the rest of the pack - no dependencies beyond ComfyUI itself. ComfyUI Manager, search Photoshoot; or

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

then restart. The heavy model files only matter if you run the example workflow, which loads Comfy-Org's Krea 2 Turbo (fp8), the Qwen3-VL-4B text encoder (CLIP type krea2) and the Qwen-Image VAE. The node itself never touches a model - it writes a string, so any model that reads prompts can consume it, though the long, sentence-like phrasing is built for LLM-style encoders (Krea 2, Flux, Qwen-Image). On a 77-token CLIP model like SDXL the style still works; it just gets chunked.

Where people get burned

If you pick a black-and-white look and still get colour, the style isn't at the front of your prompt text - position is the whole trick at CFG 1. And remember this node sets the look, not the pixels: distilled models keep a hint of brown in the lips no matter what you write, which is exactly what the pack's Monochrome node (between VAE Decode and Save Image) exists to strip. Also set expectations: the posters were rendered on Krea 2 Turbo specifically, so a different checkpoint will read the same phrase differently. Don't blame the dropdown when your SDXL output isn't the poster - the author measured the words, not your model.