llm-dit-experiments

Multi-pipeline LLM-DiT generation platform. LLM hidden states -> flow-matching DiT -> VAE decode. Single GPU (24GB).

Backend: PyTorch, FastAPI, TOML config. Frontend: React 19, Vite 7, Bun (web/frontend-v2/).

Pipelines

Quick Start

1. Backend

uv sync
cp config.toml.example config.toml   # edit model paths
uv run web/server.py --config config.toml

API on port 7860.

2. Frontend

cd web/frontend-v2
bun install
bun run dev

UI on http://localhost:5175. Vite proxies /api to the backend.

3. CLI (optional)

# Requires server running (step 1)
uv run scripts/gen.py flux2 --prompt "A photo of a cat" --seed 42

4. Batch Generation

Process a directory of images with the same prompt and model. Reads config.toml for server URL and default model.

# Basic -- uses config.toml defaults for server + model
uv run scripts/batch_flux2.py \
  --input-dir /path/to/images \
  --prompt "make this a watercolor painting"

# Override model, match output size to input
uv run scripts/batch_flux2.py \
  --input-dir /path/to/images \
  --output-dir /path/to/outputs \
  --prompt "transform this" \
  --model-name klein-9b-kv-fp8 \
  --match-image-size "0 (First Image)"

Supports resume -- interrupted runs skip already-completed images. Use --no-resume to regenerate all.

API

Experiments

Ablation sweeps and comparison tools in experiments/. See experiments/README.md.

Reference

• Configuration -- TOML config, hardware profiles, HTTPS
• CLI flags -- all command-line options
• API endpoints -- request/response reference
• Quantization -- methods, tradeoffs, backends
• config.toml.example -- annotated example