Feedback-Driven Image Pipeline

Local Streamlit UI for the pipeline in the reference slide:

User prompt → Qwen3-4B enhance → Bonsai 4B generate → vision critic score → Qwen3-4B fix

Run

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
ollama serve
ollama pull qwen3:4b
streamlit run app.py

The critic needs a vision-capable Ollama model. The Bonsai field expects a local HTTP image-generation service that accepts {"prompt":"..."} and returns an image, {"image":"<base64>"}, or {"url":"..."}. If no endpoint is configured, upload a source image to exercise the enhancement and scoring loop.

The sidebar's Pull slide models button downloads the exact models represented in the slide:

• qwen3:4b — via Ollama
• prism-ml/Bonsai-4B-mlx-1bit — via Hugging Face MLX
• mlx-community/VisualQuality-R1-7B-bf16 — via Hugging Face MLX

Note: MLX weights and Ollama model files are separate formats; the downloaded MLX folders are not automatically importable into Ollama. Bonsai and VisualQuality-R1 need an MLX-compatible serving/inference adapter to replace the current Bonsai HTTP endpoint and Ollama vision critic at runtime.

Latency Optimizations

• Prepare pipeline models checks /api/tags, skips models already installed, and pulls missing models concurrently
• Ollama requests reuse a cached HTTP session, warm both models in parallel, and set keep_alive: -1, so Qwen and the vision critic remain loaded between pipeline runs
• Vision inputs are resized to a maximum 1280px edge and JPEG-compressed before upload
• Prompt-only calls use small output limits and low temperature
• The input form prevents a full Streamlit rerun on every keystroke or file selection

Tip for low-memory systems: Change keep_alive in app.py from -1 to a duration such as "10m" so Ollama can unload idle models.