A structured prompt builder for photorealistic image generation with FLUX Klein 9B in ComfyUI. Turns dropdown selections into optimized natural-language prompts — no prompt engineering required.

Features

Prompt Builder — 20+ parameters (pose, scene, lighting, camera, lens, DoF, color grading, etc.) composed into flowing prose optimized for Klein 9B
Outfit Composer — Mix and match tops, bottoms, shoes, lingerie, outerwear, and accessories into natural outfit descriptions
Image Edit Composer — Structured editing instructions (add/remove/replace elements, quick IG effects)
VLM Prompt Refiner — Vision language model integration via Ollama or Claude Code CLI. Sees your reference images and merges character identity with scene settings in a single shot
List Nodes — Boolean-toggle selectors for batch generation via XY Plot Queue (scenes, poses, shots, lighting, outfits, actions, hairstyles, effects, and more)

Installation

Git Clone (Recommended)

cd ComfyUI/custom_nodes
git clone https://github.com/artokun/ComfyUI-Photoreal-Prompt-Builder.git comfyui-klein-photoreal-promptbuilder

Restart ComfyUI. No pip dependencies required.

ComfyUI Manager

Search for "Photoreal Prompt Builder" in the ComfyUI Manager install menu.

Nodes

All nodes appear under conditioning/klein in the node menu.

Prompt Builder (kppb)

The main node. Takes a subject description + structured settings and outputs an optimized positive prompt, negative prompt, and JSON dump of all settings.

Key inputs: | Input | Type | Description | |---|---|---| | subject | string | Character description (e.g. "blonde woman in her mid 20s") | | pose | dropdown | 53 poses from standing to yoga warrior | | scene_type | dropdown | 17 locations (bedroom, beach, cafe, etc.) | | lighting_setup | dropdown | 15 presets with rich prose expansions | | photo_style | dropdown | 15 styles (phone candid, cinematic, editorial, etc.) | | shot_type | dropdown | 16 framing options (close-up to wide full-body) | | camera_angle | dropdown | 12 angles (eye level, dutch angle, from below, etc.) | | lens | dropdown | 8 options (24mm wide to iPhone cameras) | | depth_of_field | dropdown | 6 aperture settings (f/1.4 to f/8) | | color_grading | dropdown | 12 grades (natural, cinematic teal-orange, film noir, etc.) | | hairstyle | dropdown | 36 styles | | hair_color | dropdown | 27 colors | | preserve_identity | boolean | Append identity lock phrase for character reference consistency |

Optional inputs: outfit (from Outfit Composer), edit_instructions (from Image Edit Composer), environment, lighting_custom, mood, extra_details, negative_prompt, exposure controls.

Outfit Composer (kppb)

Compose natural outfit descriptions from categorical selections.

Inputs: top, bottom, shoes, lingerie_top, lingerie_bottom, outerwear, 3 accessory slots, extra details.

Output: Natural prose like "Wearing lace bralette, crop top, high-waisted jeans, ankle boots, and choker"

Image Edit Composer (kppb)

Build structured edit instructions for inpainting workflows. Supports 3 stacked edit slots + IG quick effects (rain, snow, lens flare, bokeh, neon glow, fog, etc.).

VLM Prompt Refiner (kppb)

Vision language model integration for AI-assisted prompt composition. The VLM sees your reference images directly and writes the complete generation prompt.

Modes:

describe & enhance — Merges character identity from the reference image with your scene settings
image edit aware — Describes the final scene after edits
caption only — Direct image description

Backends:

Ollama — Local VLM inference (recommended: Qwen3-VL 32B). Auto-pulls models, auto-unloads from VRAM after inference
Claude Code CLI — Uses Claude as the VLM backend with native image support

Inputs: character reference image, optional scene/prop reference images, model settings, temperature, seed.

List Nodes (kppb)

Boolean-toggle selectors that output LIST + INT count for XY Plot Queue batch generation:

Scene List, Pose List, Shot Type List, Camera Angle List
Lighting List, Outfit List, Hairstyle List
Action List, Group Action List
Image Edit List, IG Effect List

NSFW Module

An optional NSFW expansion module is available as a separate submodule. It adds explicit pose, action, and group action expansions plus corresponding list nodes. See the NSFW module repo for details on what's included.

NSFW content is disabled by default and not downloaded during installation.

Enabling

Edit config.json in the node folder:

{
  "nsfw": true
}

Restart ComfyUI. The submodule will be fetched automatically and the NSFW nodes will appear in the node menu.

Disabling

Set the flag back to false:

{
  "nsfw": false
}

Restart ComfyUI. The NSFW module files will be cleaned up automatically and the nodes will be removed from the menu.

VLM Setup

Ollama (Local)

Install Ollama

The node auto-pulls models on first use, or manually:

ollama pull huihui_ai/qwen3-vl-abliterated:32b-instruct-q8_0

Connect the VLM Prompt Refiner node with ollama_url set to http://localhost:11434

Claude Code CLI

Install Claude Code:

npm install -g @anthropic-ai/claude-code

Enable use_claude_code on the VLM Prompt Refiner node
Select model (opus, sonnet, or haiku)

Examples

Example workflows are in the examples/ folder. Drag and drop the JSON files into ComfyUI to load them.

| Workflow | Description | |---|---| | xy_plot_minimal.json | Start here. Minimal, self-contained XY Plot — sweeps outfit × pose into a grid using a plain Flux Klein txt2img pipeline. Only needs this pack + ComfyLab-Pack + comfy-core. | | xy_plot_sfw.json | Advanced batch workflow with a Flux.2 Klein image-edit subgraph, face-detailer, and VLM refiner. Requires several extra custom node packs (see below). |

Required custom nodes

The example workflows use third‑party nodes in addition to this pack and comfy‑core. Install these from ComfyUI Manager → Custom Nodes Manager (search by the name below), then restart ComfyUI.

[!IMPORTANT] ComfyLab-Pack must be installed manually — search for "ComfyLab Pack" in the Manager. ComfyUI Manager's Install Missing Custom Nodes cannot auto-resolve XYPlotQueue / XYPlotRender because ComfyLab-Pack publishes those nodes under their display names (XY Plot: Queue (lab)), not their workflow type names. This is an upstream quirk, not a missing dependency.

xy_plot_minimal.json — one extra pack:

| Pack | Manager search | Provides | |---|---|---| | ComfyLab-Pack | ComfyLab Pack | XYPlotQueue, XYPlotRender |

xy_plot_sfw.json — the above plus:

| Pack | Manager search | Provides | |---|---|---| | rgthree-comfy | rgthree | Fast Bypasser, Mute / Bypass Repeater | | ComfyUI-Impact-Pack | Impact Pack | BboxDetectorSEGS, ImpactSEGSToMaskList, MaskListToMaskBatch | | ComfyUI-Impact-Subpack | Impact Subpack | UltralyticsDetectorProvider | | ComfyUI-Inspire-Pack | Inspire Pack | KSampler //Inspire | | ComfyUI-LanPaint | LanPaint | LanPaint_KSampler | | ComfyUI-Easy-Use | Easy Use | easy cleanGpuUsed | | ComfyUI-KJNodes | KJNodes | ImageResizeKJv2, TextPreview |

Both workflows target FLUX.2 Klein 9B — they expect flux-2-klein-base-9b-fp8.safetensors (diffusion model), qwen_3_8b_fp8mixed.safetensors (text encoder), and flux2-vae.safetensors (VAE). Swap in your own model file names in the loader nodes if they differ.

Tips

Lighting is king for Klein 9B — the lighting preset has the most impact on output quality
Use preserve_identity: true when working with character reference images
Connect the Outfit Composer output to the Prompt Builder's outfit input
Connect the Image Edit Composer output to the Prompt Builder's edit_instructions input
Use List Nodes with XY Plot Queue for systematic exploration of poses, scenes, lighting, etc.
Set unload_model: true on the VLM Refiner when running Ollama locally to free VRAM for diffusion

License

MIT