MultiImageClient

A C# console app that composes chains of image-generation and prompt-rewriting steps across many APIs. You build up a prompt through randomization, stylization and (optional) Claude rewrites, then fan it out to one or more image generators at once.

Example of one chain:

"a cat"
   ─[randomizer step]─▶ "a cat, weather conditions: sleet, location: intergalactic crossroads"
   ─[Claude rewrite: put this in the voice of an insane medieval hierophant]─▶ "<insane prompt>"
   ─[BFL/Flux image generation]─▶ <image>

What it does

Two workflows are exposed in Program.cs at startup:

1. BatchWorkflow — take a list of prompts, optionally transform them (randomize, stylize, Claude-rewrite, add permanent suffix/prefix), and fan each one out to every configured image generator.
2. RoundTripWorkflow (image → describe → re-generate) — take an image, caption it with a describer, then generate a new image from that caption on every configured generator.

Supported providers

Everything below is wired up in MultiImageClient/ImageGenerators/. Any provider is optional — if you don't want to use it, just don't put it in the generator list in Workflows/GeneratorGroups.cs and you won't need its API key.

Provider	File	Uses setting
Black Forest Labs — Flux v1.1 / v1.1 Ultra	BFLGenerator.cs	BFLApiKey
Ideogram — v2, v2-turbo, v2a, v2a-turbo	IdeogramGenerator.cs	IdeogramApiKey
Ideogram — v3	IdeogramV3Generator.cs	IdeogramApiKey
OpenAI — DALL·E 3	Dalle3Generator.cs	OpenAIApiKey
OpenAI — GPT-Image-1 / GPT-Image-1 Mini	GptImageOneGenerator.cs	OpenAIApiKey
Recraft (many styles & substyles)	RecraftGenerator.cs	RecraftApiKey
Google Gemini image ("NanoBanana")	GoogleGenerator.cs	GoogleGeminiApiKey
Google Imagen 4 (Vertex AI)	GoogleImagen4Generator.cs	GoogleCloudProjectId + GoogleCloudLocation + GoogleServiceAccountKeyPath (+ GoogleGeminiApiKey)

Prompt-rewriting / transformation steps (see MultiImageClient/promptTransformation/):

• ClaudeRewriteStep — uses AnthropicApiKey.
• RandomizerStep, StylizerStep, ManualModificationStep — no API key required.

Describers (for RoundTripWorkflow, see MultiImageClient/Describers/):

• Claude / OpenAI / Gemini image-to-text.
• Local InternVL (via do_flask_intern.py — a small Flask wrapper around OpenGVLab/InternVL3-1B-Pretrained) and local Qwen. These are optional and only used if you wire them into a run.

Requirements

• .NET 9 SDK. Projects target net9.0 (and net9.0-windows for the main app, because it uses Windows Forms for compositing). Check with dotnet --list-sdks; install with winget install Microsoft.DotNet.SDK.9 if missing.
• Windows (the main project is net9.0-windows + UseWindowsForms). The three API-client projects are plain net9.0 and portable, but MultiImageClient.csproj itself is Windows-only.
• Visual Studio 2022 (17.9+) works, or just dotnet CLI.
• For the Imagen 4 path: a Google Cloud project with Vertex AI enabled and a service-account JSON key.
• For the local InternVL describer: Python 3.10+, PyTorch with CUDA, transformers, flask, pillow, torchvision.
• For the (experimental) Django gallery under djangoManager/: Python + django mysqlclient requests ipdb aiohttp, plus MySQL.

Build & run

# one-time
dotnet restore MultiImageClient.sln

# build
dotnet build MultiImageClient.sln

# run the main app (from repo root)
dotnet run --project MultiImageClient/MultiImageClient.csproj

At startup the program asks whether to run BatchWorkflow (1) or RoundTripWorkflow (2). Prompts come from prompts.txt next to the project (MultiImageClient/prompts.txt) — or wherever LoadPromptsFrom points — and from whatever you wire into Program.cs.

To customize a run, edit:

• MultiImageClient/Program.cs — top-level wiring.
• MultiImageClient/Workflows/GeneratorGroups.cs — the myGenerators list controls which image generators (and which parameter combinations) are actually hit for each prompt.
• MultiImageClient/Workflows/BatchWorkflow.cs or RountripWorkflow.cs — the workflow logic itself.

Configuration

Copy MultiImageClient/settings - Fill this in and rename it.json to MultiImageClient/settings.json and fill in only the keys for the services you want to use. settings.json is already .gitignored.

Fields validated at startup (required):

• LogFilePath
• ImageDownloadBaseFolder
• GoogleCloudLocation, GoogleCloudProjectId, GoogleServiceAccountKeyPath — note: these are required by Settings.Validate() even if you don't use Imagen 4. If you aren't using Imagen, just point them at placeholder values.

Optional (only needed for the matching generators):

• IdeogramApiKey — https://ideogram.ai/manage-api
• OpenAIApiKey — https://platform.openai.com/api-keys
• BFLApiKey — https://api.bfl.ai/
• RecraftApiKey — https://www.recraft.ai/
• GoogleGeminiApiKey — https://ai.google.dev/gemini-api/docs/api-key
• GoogleCloudApiKey — Vertex API key (alternative to the service-account path)
• AnthropicApiKey — https://console.anthropic.com/settings/keys

Cosmetic flags: SaveJsonLog, EnableLogging, AnnotationSide.

Project layout

MultiImageClient/            main console app (net9.0-windows)
  Program.cs                 entry point
  Workflows/                 BatchWorkflow, RoundTripWorkflow, GeneratorGroups
  ImageGenerators/           one file per provider (BFL, Ideogram, Dalle3, Recraft, GPT-Image-1, Gemini, Imagen 4)
  Describers/                image-to-text (Claude, OpenAI, Gemini, local InternVL, local Qwen)
  promptTransformation/      Claude rewrite, randomizer, stylizer, manual edits
  promptGenerators/          prompt sources (from file, from code, scenes-from-story, etc.)
  Utils/                     ImageCombiner, ImageSaving, TextFormatting
  Enums/, Interfaces/, Implementation/, TextLLMs/

ImageGenerationClasses/      shared types + Settings.cs (net9.0 library)
BFLApi/                      BFL/Flux API client (net9.0)
IdeogramAPI/                 Ideogram API client (net9.0)
RecraftAPI/                  Recraft API client (net9.0)

do_flask_intern.py           optional local InternVL3-1B Flask server
save_b64.py                  helper to decode base64 payloads from provider responses
djangoManager/               experimental Django gallery (MySQL-backed, not actively developed)
IdeogramHistoryExtractor/    older scratch folder

Gallery

If you haven't tried BFL/Flux, you should — better composition than DALL·E 3, 2× the pixels, not quite as strong as Ideogram on in-image text but generally much cleaner images.

• BFL gallery: https://photos.app.goo.gl/baJNz9SWX1fq1tT77
• Ideogram gallery: https://photos.app.goo.gl/QJn5xPUNEg1uuNdaA

Design goals (aspirational)

• Easily compose, order, and manage combinations of image-to-text and text-to-image steps across endpoints.
• Given one interesting prompt, see side-by-side what every current generator would do with it, with annotations, as a distributable composite image.
• Per-provider "ability maps" — known strengths, known censorship patterns — so requests that will be blocked aren't sent in the first place.
• Viewing & browsing history with tag jumping and "re-run this exact prompt on every API" buttons.
• Persistent image-level metadata (prompt, settings, seed) baked into the file or a sidecar, so images shared out never lose their provenance.

Todos / not done yet

• Django gallery is still an experiment; no active work.
• More / cheaper / less-censored prompt rewriters beyond Claude.
• Move run configuration out of Program.cs into JSON so the app is distributable as a bare exe.