promptkit Github 

Prompts for image models

Give it an image.
Get the prompt.

A vision model reads the picture and writes the prompt that would generate it. One per image model you run: FLUX, SDXL, Pony, or whatever else you add.

Run it yourself Read the source 
Promptkit detail view: a photograph beside its Flux.1 Dev and Stable Diffusion XL prompts
The detail view. One image, one prompt per model.
Self-hostedYour hardware, your bucket, your LLM endpoint
NeedsDocker and any OpenAI-compatible vision model
TakesUploads, pasted URLs, or a watched Discord channel

§ How it works

  1. 01

    Ingest

    Upload a file, paste a URL, or let the Discord bridge pick up whatever gets posted in the watched channel.

  2. 02

    Analyse

    A worker pulls the job off the queue and asks a vision model what is actually in the frame.

  3. 03

    Write prompts

    Every image model you have enabled gets its own prompt, written in the format that one expects.

§ The core loop

One image in.
N prompts out.

FLUX takes prose. SDXL takes comma-separated tags. Pony takes score tags and a booru accent.

You get a prompt for each image model you have switched on, each stored on its own, so rewriting the FLUX one leaves SDXL untouched. Each image model carries its own formatting instructions, and those are yours to edit in settings, along with the wording wrapped around all of them.

  • The original output is kept. Your edit is stored next to it.
  • Edits save as you type. Favourited prompts collect on their own page.
Promptkit settings on a phone, editing the Stable Diffusion XL label and its prompt instructions
Editing what one model asks for.

§ Library

Everything you've ever fed it.

Filter by status or favourite, or search across filenames, notes, scene descriptions and the prompt text itself.

Promptkit library grid of ingested images, each marked done with three prompts
The library, with filters and search.

§ In the box

Your own model list
Image models live in settings, not in the code. Give one a name and the instructions for how its prompts should be written, or switch it off without deleting what it made.
Backfill
Add an image model later and generate its prompts for the images you already have. Or regenerate a single one you didn't like.
Notes that steer the output
Add a note to an image and it goes into the instructions the vision model gets. Set it when the image arrives or edit it later.
Discord ingest
A bot watches one channel and sends image attachments straight in. Caption it note: … to set the note from Discord.
Prompts without a picture
Describe what you want in words and get a prompt back for each image model you pick. Sampling is yours to tune, conversations are kept, and any reply can be saved to your favourites.

§ Image models

Three to start with.

Each one has its own instructions, so the same picture comes back written differently for every model on the list. Add or remove them in settings.

§ Chat

Promptkit chat on a phone: Stable Diffusion XL and Pony Diffusion replies, with the model picker below
Chat. One reply per selected model.

No picture?
Just ask.

Describe the thing you want instead of feeding it a reference. One reply streams per image model you have ticked, each written to that model's own instructions. Keep a reply and it joins your favourites.

Chat runs on its own system prompt, separate from the extraction one and editable in the same place.

§ Run it

One compose file and an endpoint.

Postgres, Redis and a MinIO bucket all come up with compose. The only thing you bring is an OpenAI-compatible vision model. A local llama.cpp server works, so does a hosted provider.

cp .env.example .env    # fill in LLM_*, change the two secrets
docker compose up -d --build

Then open localhost:3000 and sign in as admin / changeme. Change that first.

Get the repo
Promptkit dashboard with status counts, recent uploads and the processing queue
The dashboard once it is running.

§ Stack

What it's built with.

One Next.js app, two Node services, and the packages they share.

Built with

  • languageTypeScript, strict, shared config across the monorepo
  • webNext.js 15 and React 19, App Router and server actions
  • uiTailwind 4, Radix primitives, virtualised lists
  • dataPrisma, with migrations and a seed
  • jobsBullMQ, with server-sent events to the browser
  • authbetter-auth, roles checked on every server action
  • botdiscord.js, running as its own process
  • toolingpnpm workspaces, Turborepo, zod for config

Runs on

  • nodeNode 20
  • databasePostgreSQL 16
  • queueRedis 7
  • storageMinIO, or any S3-compatible bucket
  • visionAny OpenAI-compatible endpoint, local or hosted
  • imagesDocker, multi-stage builds, non-root
  • composeOne file for local, one for deployment
  • ciGitea Actions, type-check and lint before build