inteintegrity2026-09-29

RK1820 Vision Story Studio

Upload an image and turn it into a postcard, short story, scene description, or creative prompts with a local Qwen2.5-VL-3B model on reComputer RK3576 and RK1820.

Image UnderstandingVLMCreative AIGitHub

Vision Story Studio turns an image into creative text using a local vision-language model. On a reComputer RK3576 with an RK1820 accelerator, upload a picture, choose a writing direction, and watch the response stream into the browser. The project's model is Qwen2.5-VL-3B_prune; image and prompt processing stay on the device while its local model service is running.

Project source code and README

Image upload and local VLM response on RK1820

The clip shows a related image-understanding demo using RK1820 and Qwen2.5-VL-3B_prune: an image is uploaded and the response appears progressively. Vision Story Studio uses the same hardware and model combination for its creative writing presets.

What you can create

Upload a JPG, PNG, or WEBP image and select one of the built-in directions:

  • a postcard inspired by the image;
  • a short story;
  • a visual scene description; or
  • a set of creative prompts.

You can edit the prompt before pressing Create Story. The answer streams from the local model service to the browser.

Hardware and model files

  • reComputer RK3576 with an RK1820 module installed and recognized.
  • A compatible RKNN3/RKLLM runtime with rkllm3-server available on the board.
  • Python 3 with venv for the FastAPI web interface; its packages are listed in the project's requirements.txt.
  • A separately obtained Qwen2.5-VL-3B_prune model package. Model binaries are not included in the standalone project archive.

The model launcher expects these files under MODEL_DIR:

text
llm_Qwen2.5-VL-3B_prune.rknn
llm_Qwen2.5-VL-3B_prune.weight
vision_Qwen2.5-VL-3B_prune.rknn
vision_Qwen2.5-VL-3B_prune.weight
Qwen2.5-VL-3B_prune.tokenizer.gguf
Qwen2.5-VL-3B_prune.embed.bin

Check that the model package and the board's runtime are compatible before starting the service.

Run the studio

Clone the repository and enter the project directory:

bash
git clone https://github.com/Seeed-Projects/Lab-Prooject-RK.git
cd Lab-Prooject-RK/projects/rk1820-vlm-creative
chmod +x *.sh

In the first terminal, point to the prepared model directory and start the VLM service:

bash
MODEL_DIR=/path/to/Qwen2.5-VL-3B_prune ./start_vlm_server.sh

Keep that terminal running. In a second terminal, enter the same project directory and start the web interface:

bash
./start_ui.sh

Open http://<RK3576-IP>:7861 from a browser on the same network. The UI connects to the model service at http://127.0.0.1:8080 by default. On its first run, start_ui.sh creates a virtual environment and installs the Python dependencies, so package access is required unless they are already cached.

Notes

The project's code and launch scripts are in Seeed-Projects/Lab-Prooject-RK. Keep the web interface on a trusted local network or add access controls before exposing it more widely. The model package and compatible board runtime must be prepared separately.