ComfyUI-Photoshoot

Photoshoot

You built a person. Now shoot 40 photos of her.

Photoshoot is a set of ComfyUI nodes in two halves. The Person Builder describes someone across 44 fields — body, face, hair, make-up, clothing — and saves them under a name. The Photoshoot turns that person into a whole series: every photo a different framing, pose, placement, expression and aspect ratio, while the person stays the same. Set the count, press the button, and the node queues all 40 runs itself.

Building a person

44 fields across six tabs. You pick labels; the node writes English. Related fields share a row — hair and its colour, eye shape and eye colour, lips and finish — because they are merged into one phrase and you need to see both while setting either.

A person with nineteen fields set:

Basics    Woman · Late 20s · Scandinavian · Very fair · Natural pores
Body      Tall · Athletic · Athletic shoulders
Head      Long waves + Copper red · Almond-shaped + Greyish green
Face      High, pronounced cheekbones · Straight nose
Make-up   Dusty rose + Matte · Freckles
Clothing  Bare legs · Combat boots + Black
Details   wearing a charcoal wool coat

That produces one string, and the eye shape and colour have become a single phrase rather than two competing ones:

a woman, in her late 20s, Scandinavian features, very fair skin, with natural
texture and visible pores, tall, athletic, toned physique, athletic defined
shoulders, long wavy copper red hair, high, pronounced cheekbones, a straight
nose, almond-shaped grayish green eyes, matte dusty rose lipstick, freckles,
wearing a charcoal wool coat, bare legs, wearing black combat boots

The same person seen from further away — this is what the Photoshoot sends for a full-body shot, and for a wide one:

[figure]    … long wavy copper red hair, almond-shaped grayish green eyes,
            freckles, wearing a charcoal wool coat, bare legs, black combat boots
            (cheekbones, nose and lipstick dropped — 303 chars instead of 376)

[identity]  … long wavy copper red hair, grayish green eyes, freckles,
            wearing a charcoal wool coat, bare legs, black combat boots
            (eye shape and shoulders dropped too — 261 chars)

Nothing is lost: the builder keeps all nineteen fields and still shows them. Only what gets sent shrinks with distance, because tokens the camera cannot resolve still pull the composition onto the head.

<img src="docs/clothing.png" alt="The Person Builder on the Clothing tab, with per-tab counters, paired fields and the composed sentence" width="520">

The Person Builder, Clothing tab. The superscripts count what is set on each tab; hosiery and shoes carry their colour as a separate field on the same row; the reset buttons name what they would clear. The composed English sentence sits at the bottom and updates as you click.

The photoshoot

Six axes. Each can be switched off — then it holds still — or restricted to a family: only standing poses, only calm moods, only close-ups.

| Axis | What varies | |---|---| | Camera | 7 framings, extreme close-up through wide shot | | Pose | 18 postures in 4 families, plus placement in the room, orientation, arms, legs, tension — placement is coupled to the framing and tension to the posture, so no photo asks for a figure that is leaning against a wall and curled up at once | | Expression | 90 moods in 9 families, plus eyes, gaze, mouth, brows, head tilt | | Focus | 12 focal points, face to feet — coupled to the framing, since a wide shot does not focus on the lips | | Ratio | 9 aspect ratios at a constant pixel count, coupled to the framing so a full-body shot never lands in 16:9 landscape | | Noise | a fresh seed per photo, or one seed for the whole series to keep the setting recognisable |

<img src="docs/photoshoot.png" alt="The Photoshoot node: six axes, a start button, and a preview listing the first runs" width="400" align="right">

Two things make this a series rather than 40 random images.

It counts, it does not roll dice. Run 7 always produces the same photo, so a series is reproducible and you can extend it later. Counting uses a Kronecker sequence — each field steps by the fractional part of a different square root — because plain integer factors made fields with related list lengths march in lockstep: the same body turn kept arriving with the same mouth.

The person shrinks with distance. The prompt for a wide shot drops lipstick, eyeliner and brow shape and keeps the silhouette. Those tokens are invisible at that range but still pull the composition back onto the head. More on that.

What a series looks like

Eight runs of one series, unretouched. Same person throughout — the hair, the freckles, the charcoal coat and the black boots come from the Person Builder and do not drift. What changes is the framing, the pose, where she stands in the room, the expression and the aspect ratio; the two wide shots are landscape because the framing chose the ratio.

You set the count — the default is twelve, 40 is as easy. The person is fixed; everything else moves. This is the node's own preview of the first eight runs:

 1  Extreme close-up → Eyes      Standing                Foreground              Neutral      1328×1328 [full]
 2  Portrait         → Face      Squatting               Foreground              Calculating  1152×1536 [full]
 3  Full body        → Legs      Sitting on a table edge In a doorway            Skeptical     992×1776 [figure]
 4  Close-up         → Neckline  Walking                 By the window           Gentle smile 1184×1488 [full]
 5  Cowboy shot      → Waist     Lying on the stomach    Moving through the room Defiant      1088×1632 [figure]
 6  Extreme close-up → Lips      Kneeling                Foreground              Frightened   1328×1328 [full]
 7  Medium shot      → Back      Sitting on a stool      By the window           Mischievous  1088×1632 [figure]
 8  Wide shot        → Back      Leaning against a wall  Against the far wall    Seductive    1328×1328 [identity]

Every run is a plain prompt. Photo 1 comes out as:

extreme close-up shot, with the focus on the eyes, a sunlit loft with tall
windows, a woman, in her late 20s, Eastern European features, fair skin, tall,
hourglass figure, long wavy dark brown hair, almond-shaped green eyes, matte red
lipstick, wearing black opaque pantyhose, wearing black stiletto high heels,
standing, placed in the foreground, facing the camera directly, both arms held
behind the back, hands clasped together, legs closed together, with an upright
posture, a neutral expression, wide open eyes, looking directly at the camera,
relaxed brows, lips closed, head held straight, editorial photography, 85mm

The [full] / [figure] / [identity] tag is that detail level — compare photo 1 with photo 8 and the lipstick is gone.

Wiring

Person Builder ──person_data──► Photoshoot ──┬── person   ─┐
  44 fields, saved once          the series  ├── pose      │
                                             ├── expression├──► Build Prompt ──► your sampler
                                             ├── camera    │
Scene / Style ───────────────────────────────┴── width/height/seed

1. Build a person. 44 fields across six tabs. Save under a name and reuse across sessions — the point is that the same person comes back.
2. Send them to a photoshoot. Set the count, restrict the axes you care about, then press Start shooting in the node's own panel — that button, not the Queue button above the canvas. Queue renders one run, photo 1 of the series; Start shooting resets the counter and queues the whole series.
3. Assemble the prompt. Placeholders ({camera}, {person}, {pose} …) drop the parts into your own text, or the node appends them in a sensible order if you have none.

Pose and expression also work as standalone nodes when you do not want a series.

Already have your character?

Then leave step 1 out. The Person Builder is optional — unwire it and the {person} placeholder simply drops from the prompt, no stray commas, and the Photoshoot node varies framing, pose, placement, expression and aspect ratio around whoever your character already is:

close-up shot, with the focus on the neckline, walking, near a window, torso
turned away, looking back over the shoulder, arms wrapped around the knees,
knees together, feet apart, back arched, a soft gentle smile, eyes closed,
looking past the camera, furrowed brows, pursed lips, head tilted slightly

That is a whole series' worth of direction with no description of a person in it. Whatever holds your character — a LoRA and its trigger word, an IPAdapter reference, an identity-preserving edit model — supplies the who, and these nodes supply everything else. Put the trigger word in the {extra} placeholder if you need it first in the prompt.

Two things help there. Describing the face on top of a reference or a LoRA gives the model two sources for one thing, so leave those fields empty. And a reference photo shows no body below the shoulders, so the body tab is worth filling in even when the face comes from elsewhere — otherwise every image invents a different build.

You do need one of those mechanisms, though. Plain img2img is not a substitute: the latent carries the source image's geometry, so framing and aspect ratio cannot move, and with the description gone nothing holds the identity either. Measured at three denoise levels — at 0.45 the framing never changed and the face already had, at 0.85 the direction won and the person was someone else.

Need a character to use elsewhere?

The other direction: a handful of consistent images of one person, to feed a reference input, a LoRA training set, or the next shot of a video.

Switch the noise axis off before you go looking, not after. This is the one place where the obvious order is the wrong one. With noise on, every run reinvents everything the description leaves open, so no two runs share a face. Switching the axis off replaces that with a single series seed and puts a seed field with a die next to it — roll it until you like who turns up, and from then on she is fixed. Camera, pose and expression keep varying, because those come from the prompt rather than from the noise.

Doing it the other way round is the trap: find someone you like at run 7, switch the axis off, and you get a stranger, because the seed changed from 7 × 2654435761 mod 2³¹ to whatever the series seed says. If you do want to keep a particular run, that formula is the answer — run 7 is seed 1401181143, typed into the field.

Then restrict the camera axis to the framings you need, set the count, and press Start shooting. Which framings depends on what has to stay fixed: a close-up carries the face and nothing below the shoulders, a full-body shot carries the clothes and the build but no facial detail. Keep both and pick per use.

A character LoRA fits in here too, if you have one. The workflow below already carries the loader, wired between the UNETLoader and both samplers and shipped bypassed, so it runs untouched as it is: pick your file, press Ctrl+B, and the model flows through it instead of around it. You do not need one, though. The noise seed alone holds a person still.

example_workflows/photoshoot-anchor-set.json is all of that already set up: same graph as the series workflow, but with the noise axis off and a seed in place, the ratio fixed, the mood held to one family and the count at six. Load it, press Start shooting, and you have your anchor set — without having to know about the trap above first. Images go to krea2/anchor so they do not mix with a series.

Example workflow

example_workflows/photoshoot-series.json is the whole thing wired up — drag it onto the ComfyUI canvas.

It contains the three Photoshoot nodes above, a scene and a style field, and a two-pass sampler: 8 steps at full denoise, a 1.5× latent upscale, then 4 steps at denoise 0.44. Keeping the upscale but dropping the refinement is the one combination to avoid — that takes 11.4 s instead of 24.8 s and comes out unusably soft, because latent upscaling invents no detail, it only pulls things apart. Dropping both is a different thing — see "Draft before you shoot" below.

width and height come from the Photoshoot node, so the aspect ratio follows the framing. bildseed feeds both samplers, which is what makes a run reproducible.

Draft before you shoot

Before shooting a series you usually want to see the first five or six runs: is the person right, the scene, the style, the axes you restricted? Rendering those at full quality is a waste, because you are going to change something and shoot them again.

Select LatentUpscaleBy and the second KSampler, press Ctrl+B. That is bypass, not mute: both nodes take a LATENT and hand back a LATENT, so the latent passes straight through them and the decode reads the first sampler instead of the second. Measured here against local Krea 2 weights, that is 10.2 s per image instead of 21.5 s, at the base resolution rather than 1.5×. Sharp, just smaller. Press Ctrl+B again and you are back at full quality with every setting untouched — which is the point of doing it this way rather than in a second file: the person, the scene and the axis restrictions live in the nodes, and moving them across workflows is work.

Because the series counts through combinations rather than drawing at random, runs 1 to 6 in the draft are runs 1 to 6 in the finished series. You are previewing the real thing, not a stand-in. Six calibration images cost 61 s instead of 129 s, and you rarely go round only once.

If you would rather keep the drafts out of the finished set, change the SaveImage prefix while the two nodes are bypassed.

Drafting a whole series works too, and there it depends on how many you keep, since everything renders twice: forty full runs cost 860 s, forty drafts plus k re-renders cost 408 + 21.5·k, even at k ≈ 21.

The three loaders name the files Comfy-Org publishes for Krea 2 — krea2_turbo_fp8_scaled, qwen3vl_4b_fp8_scaled as CLIP type krea2, and qwen_image_vae — so the example matches what you already have if you followed the official ComfyUI tutorial. If yours are named differently, ComfyUI flags the loader nodes on load and you pick your own; the rest of the graph is unaffected.

The text encoder is the one part that is not interchangeable. Krea 2 taps twelve layers of Qwen3-VL-4B at a hidden size of 2560, so plain Qwen3-4B (no vision tower) and the 32B build (wrong width) will not load.

After changing anything under js/, bump the version in pyproject.toml and run python3 tools/setze_js_version.py. ComfyUI serves the .mjs files without a Cache-Control header, so without a fresh ?v= in the import a browser can sit on the old interface for a day or more.

The workflow is generated by tools/baue_workflow.py, which checks every node type, input name, output index and required input against a running ComfyUI. It was verified by actually executing it, not just by loading it.

Which models does this work with?

Any of them, in principle. The nodes emit text — the person, the pose, the expression, the camera phrasing — plus width, height and a seed. Nothing here touches a model, so whatever consumes a prompt can consume these.

Nothing goes in, either. These nodes take no image and no reference — you state the person, so nothing is guessed. That is the difference to a workflow driven by one portrait photo: such a photo holds no body below the shoulders, so the model fills the gap from its own prior. The two combine well, though. Since all that leaves here is a string, an IPAdapter can carry the face while these fields say what the photo cannot show.

What differs is how much of it survives:

• Models with a T5 or LLM text encoder — Flux, SD 3.5, Qwen-Image, Krea 2 — are the good case. A finished prompt here runs 760 to 1000 characters, median around 840, roughly 190–250 tokens, and these read all of it as connected language. The whole design assumes this: that "arms behind the back, hands clasped together" lands as a sentence rather than as loose keywords.
• CLIP-only models — SD 1.5, SDXL — cap out at 77 tokens per chunk. The prompt still works, but it gets split, and details near the end lose weight. Set fewer fields, and use the detail levels: a wide shot already sends a much shorter person block.
• Edge lengths run 1024 to 1536, sized for current models. For SD 1.5 at 512 they are too large; the ratio coupling still helps, the pixel counts do not.

Krea 2 is what the measurements were taken against and what the example workflow loads — that is the extent of the tie.

Installation

No dependencies beyond ComfyUI itself.

Through ComfyUI Manager: search for Photoshoot and install. Or by hand:

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

Restart ComfyUI afterwards; the nodes appear under the Photoshoot category.

What changed between versions is in the changelog, and every version is tagged as a release — watch the repository for releases only if you want to hear about new ones.

The interface follows ComfyUI's language setting: German if ComfyUI is set to German, English otherwise. Note that ComfyUI defaults to your browser language when you have never picked one, so a German browser gives you a German interface without you setting anything. Pick a language explicitly under Settings → Comfy → Locale. The prompt is identical either way — see Language.

The nodes

| Node | What it does | |---|---| | Photoshoot Person | 44 fields on six tabs; outputs the person as text and as JSON | | Photoshoot Series | the series: six axes, queues the runs itself | | Photoshoot Pose | posture, placement, orientation, arms, legs, tension | | Photoshoot Expression | 90 moods in nine families, plus eyes, gaze, mouth, brows, head | | Photoshoot Build Prompt | drops the parts into your prompt text | | Save / Load | persons, scenes, styles and prompts under ComfyUI/user/krea2_* | | Photoshoot Pick Blocks | four dropdowns, four outputs in one node |

Full reference → — every control, and why the awkward ones are the way they are: how focus is coupled to framing, why the person block shrinks with distance, how the series counts without repeating itself.

More

• Internals — layout, translation, tests, which ComfyUI versions this was built against
• Measurements — what actually got measured, including a conclusion that turned out to be wrong
• Changelog

Saved blocks

The stores live under ComfyUI/user/krea2_* and are not part of this repository — that is where personal content ends up. If you want them backed up, copy those folders deliberately.

Scope

Photoshoot describes adult subjects. The age presets start at the early 20s, and the exact-age field has a floor of 18 — type a lower number and both the preview and the prompt say 18.

That floor is a guardrail against the accident, not a content filter. These nodes only emit text, and text can be typed into any node.

License

MIT — see LICENSE.