Vision-LLM prompt generation for ComfyUI — analyze images, build style libraries, and generate prompts that match any aesthetic.
Prompt808 learns content and styles from images. Drop in reference photos or artwork, and a vision model extracts the elements that define each image — lighting, composition, palette, mood, and more. Those elements become a growing library that Prompt808 draws from to generate prompts tailored to any visual style.
- Learns from images — A vision model analyzes each upload and extracts structured scene elements into a growing library
- Multiple prompt styles — Photo-Architectural, Photo-Boudoir, Photo-Cinematic, Photo-Documentary, Photo-Fashion, Photo-Fine Art, Photo-Portrait, Photo-Street, plus Native (any medium) and an Any option for random selection
- Optional LLM-powered composition — Natural-language prompts with seven enrichment levels from faithful reproduction to full creative freedom
- Multi-library generation — Combine elements from multiple libraries in a single generation using the Select Libraries node, or select "All" for every library at once
- Library management — Maintain separate libraries for different projects or genres, each fully isolated. Export and import libraries as
.p808files for backup or sharing - ComfyUI native — Sidebar panel for interactive use, plus a workflow node for automated generation
- Medium-aware — Automatically detects whether an image is a photograph, painting, illustration, or 3D render, and adjusts extraction accordingly
- ComfyUI (latest recommended)
- Python 3.10+
- CUDA GPU (8+ GB VRAM for analysis, 1+ GB for generation)
Install as a ComfyUI custom node:
cd ComfyUI/custom_nodes
git clone <repo-url> Prompt808
pip install -r Prompt808/requirements.txtRestart ComfyUI after installation.
Click the camera icon in ComfyUI's sidebar to open the Prompt808 panel. On first launch, click "+ Create Your First Library" and give it a name (e.g. "Portraits", "Landscapes"). The first library is automatically activated.
Navigate to the Analyze tab. Add images by:
- Drag and drop from a file manager
- Click to browse
- Drag images from a browser window
- Paste from clipboard (Ctrl+V)
Supported formats: JPG, PNG, WebP, BMP, TIFF, HEIC. Batch upload is supported — each image is processed sequentially with live progress.
The Library tab shows all extracted elements. Filter by category, edit descriptions and tags, or delete elements.
The Photos tab displays thumbnails of all analyzed images. Click a photo to see its extracted elements, or delete a photo to remove it with all associated data.
Add a Generate Prompt node to your ComfyUI workflow. All generation settings are exposed as node inputs — library, style, mood, archetype, archetype influence, LLM model, enrichment, temperature, and more. Each archetype also has an "All Elements" option that biases toward every element in that archetype's categories.
Use the library dropdown and action buttons at the top of the sidebar to:
- Switch between libraries (each has its own elements, archetypes, style profiles, and dedup caches)
- Create new libraries with the + button
- Rename or delete libraries
- Export a library as a
.p808file for backup or sharing - Import one or more
.p808files to add libraries from others
Libraries are fully isolated — the same image can exist in different libraries with different extracted elements.
To generate prompts that draw from multiple libraries at once:
- Quick method — Select "All" in the Generate Prompt node's library dropdown to use every library.
- Fine-grained control — Add a Select Libraries node. Each slot has a library dropdown and an on/off toggle. Click "+ Add Library" to add more slots. Connect its
librariesoutput to the Generate Prompt node — the library dropdown auto-hides when connected.
When multiple libraries are selected, their elements, archetypes, and style profiles are merged into a virtual combined pool for generation.
Open Settings > Prompt808 > General in ComfyUI to configure:
- NSFW — Enable mature content in prompt types and moods
- Balance Libraries (default: on) — When generating from multiple libraries, each library contributes equally regardless of size. A small 30-element library gets the same representation as a 3,000-element library. Turn off to let larger libraries dominate naturally. Has no effect with a single library
Changes take effect immediately — no restart required.
| Model | VRAM | Notes |
|---|---|---|
| Qwen3-VL-8B-Instruct | ~12 GB | Default, good quality |
| Qwen3-VL-8B-Instruct-abliterated | ~12 GB | Abliterated variant for unrestricted extraction |
| Qwen3-VL-8B-Instruct-abliterated-v2 | ~12 GB | Alternative abliterated variant |
| Qwen3-VL-8B-Instruct-FP8 | ~7.5 GB | Pre-quantized, good for 8 GB cards |
| Qwen3-VL-8B-Thinking | ~12 GB | Chain-of-thought reasoning |
| Qwen3-VL-8B-Thinking-FP8 | ~7.5 GB | Thinking model, pre-quantized |
| Qwen3-VL-32B-Instruct | ~28 GB | Higher quality extraction |
| Qwen3-VL-32B-Instruct-FP8 | ~24 GB | Pre-quantized 32B |
| Qwen3-VL-32B-Thinking | ~28 GB | Best quality, chain-of-thought |
| Qwen3-VL-32B-Thinking-FP8 | ~24 GB | Thinking 32B, pre-quantized |
Vision models with non-standard tensor dimensions automatically fall back from FP8 to FP16 if block quantization is incompatible.
| Model | VRAM (FP16) | VRAM (4-bit) | Notes |
|---|---|---|---|
| Qwen3-0.6B | 1.5 GB | 0.7 GB | Default, lightweight |
| Qwen3-1.7B | 3.5 GB | 1.3 GB | |
| Qwen3-4B | 8.5 GB | 2.8 GB | |
| Qwen3-8B | 17 GB | 5.5 GB | |
| Qwen3-8B-Abliterated | 17 GB | 5.5 GB | Abliterated for unrestricted output |
| Qwen3-8B-Abliterated-v2 | 17 GB | 5.5 GB | Alternative abliterated variant |
| Qwen3-8B-Abliterated-FP8 | ~9 GB | — | Native FP8, abliterated for unrestricted output |
| Qwen3-32B | 64 GB | 18 GB | |
| Qwen2.5-3B-Instruct | 6 GB | 2 GB | |
| Qwen2.5-7B-Instruct | 15 GB | 5 GB | |
| Qwen2.5-14B-Instruct | 28 GB | 8.5 GB | |
| Qwen2.5-32B-Instruct | 64 GB | 18 GB |
FP8 variants of Qwen3 models (0.6B through 32B) are also available. Models are downloaded automatically from Hugging Face on first use.
| Level | Fidelity | Creativity | Description |
|---|---|---|---|
| Any | — | — | Picks a random enrichment level per generation |
| Baseline | High | Low | Vivid photographic phrases, preserves original meaning |
| Vivid | High | Low-Med | Adds sensory and textural detail |
| Expressive | Medium | Medium | Evocative mood and atmosphere |
| Poetic | Low-Med | High | Metaphor, art references, heightened language |
| Lyrical | Low | High | Original phrases inspired by tags only |
| Freeform | Medium | High | Photoshoot director's instructions, grounded in scene context |
See PIPELINE.md for architecture, API reference, and technical details.
Prompt808 is free and open source. If you find it useful, you can support its development:
Coming soon... grab a pre-made library from prompt808.com — ready-to-use element collections across a range of styles and genres, so you can start generating right away.