Liidar Image Workflow Studio

Local AMD-first model studio for dataset prep, character analysis, generation, and LoRA training.
ComfyUI-oriented workflow management, prompt recipes, local training helpers, runtime checks, and output review.

What It Is • Runtime • Start • Verification • Hygiene • License

What This Project Is

• A local workflow manager for AI image generation.
• A React/Vite frontend plus FastAPI backend.
• A ComfyUI-oriented job and workflow helper.
• A place to manage profile schemas, prompt recipes, workflow templates, training metadata, and output review.
• A source-code project, not a bundled model, dataset, or media pack.

What This Project Is Not

• It is not a model-weight repository.
• It is not a dataset repository.
• It is not a generated-image pack.
• It is not a service for impersonating real people or bypassing consent requirements.

Current State

Liidar currently includes:

• React/Vite frontend at http://127.0.0.1:5274.
• FastAPI backend at http://127.0.0.1:8000.
• Runtime and live system monitoring.
• Profile creation and local reference analysis.
• Single-brief local prompt enhancement.
• ComfyUI generation preview, preflight, queueing, and output tracking.
• Dataset preparation and local training workflow helpers.
• Training run launch, polling, cancellation, and log tail support.
• Frontend and backend test coverage for the current workflow.

Optional Anthropic API Feature

Everything above runs locally. The backend also has one optional cloud feature, off by default: dataset-prep crop scanning with Anthropic's Claude API.

• It is used only when you save an Anthropic API key (POST /api/settings/anthropic-key) and run dataset prep with scan_mode: "claude" (or the older use_ai: true). The frontend uses local scanning.
• When it is used, Liidar uploads each scanned image to Anthropic (https://api.anthropic.com/v1/messages), downscaled to at most 384 x 384 pixels as a JPEG, for up to ai_max_images images per run (default 100). The key test (POST /api/settings/anthropic-key/test) sends a short text-only request. Anthropic's terms and data policies apply to what you upload, so only use it for images you have the right and consent to share.
• The API key is stored in plain text in config/secrets/anthropic.json inside the project folder (the folder is git-ignored). Anyone who can read that folder can use the key. Remove it with DELETE /api/settings/anthropic-key or by deleting the file.

Runtime Requirements

This repo does not include ComfyUI, checkpoints, adapters, model files, generated outputs, datasets, logs, or secrets.

For full generation and training tests, the local machine should provide:

• ComfyUI/ in the project root.
• A working ComfyUI Python environment.
• Compatible checkpoint files under the local ComfyUI model folders.
• Optional trained adapter files under the local ComfyUI adapter folders or registered absolute paths.
• GPU drivers and the correct PyTorch backend for the machine.

On machines without ComfyUI or a compatible GPU runtime, the app can still be developed and UI-tested, but generation preflight will report missing runtime or model readiness.

Getting Started

Requirements: Windows with PowerShell, Git, Python 3.12, uv, and Node.js with npm. The backend dependencies are locked in app/backend/uv.lock.

git clone https://github.com/gkaragioul/Liidar_Image_Workflow_Studio.git
cd Liidar_Image_Workflow_Studio\app\backend
uv sync --extra test
cd ..\frontend
npm install
cd ..\..

For generation, run .\tools\setup\install-comfyui.ps1, then install a GPU-enabled PyTorch build, the ComfyUI requirements, and your chosen model files as the script's closing notes describe. For LoRA training, .\tools\setup\install-sd-scripts.ps1 installs kohya-ss sd-scripts with an AMD RDNA4 ROCm PyTorch build. .\tools\setup\check-system.ps1 reports the OS, CPU, RAM, and GPU it finds. Then start the app as below and open http://127.0.0.1:5274.

Start The App

Use the project launcher:

.\Start-Liidar.ps1

The launcher starts:

• ComfyUI on http://127.0.0.1:8188
• Backend on http://127.0.0.1:8000
• Frontend on http://127.0.0.1:5274

If ComfyUI is not installed, start the backend and frontend manually for development:

cd app\backend
.\.venv\Scripts\python.exe -m uvicorn local_model_studio.main:app --reload --host 127.0.0.1 --port 8000

cd app\frontend
npm run dev -- --host 127.0.0.1 --port 5274 --strictPort

Verification

Backend:

cd app\backend
.\.venv\Scripts\python.exe -m pytest

Frontend:

cd app\frontend
npm test
npm run build

Repository Hygiene

Public releases should not include:

• Generated images or output media.
• Training datasets or prepared local crops.
• Model weights, checkpoints, adapters, or other binary model artifacts.
• Private prompt sets, personal reference files, logs, credentials, or cloud endpoints.
• Temporary local worktree/cache folders.

The repository is intended to contain source code, tests, configuration templates, and documentation only.

Disclaimer

Liidar is provided as is, without warranty of any kind, under the MIT License. Use it at your own risk. It reads the image folders you choose, writes and removes data in its own profile, dataset, and training folders, and runs long GPU workloads; the setup scripts clone and install third-party software. You are responsible for the images you use and generate, including consent from anyone they depict, and for the licenses of the models you install.

License

MIT. See LICENSE.