Open_Notebook

Open_Notebook is a SenseNova-powered workspace for full-canvas visual generation from user requests and uploaded sources. It is inspired by NotebookLM-style source-grounded workflows, but the default output mode relies on SenseNova's strength in complete image composition and text layout.

What It Builds

• Auto: the LLM classifies the user's intent and dispatches to the right workflow.
• PPT: the LLM plans a deck, SenseNova generates one full 16:9 image per slide, and Open_Notebook combines those images into both PPTX and PDF.
• Poster: sources are summarized into one high-density full-canvas poster prompt, then SenseNova generates a complete large poster image.
• Research Figure: creates complete architecture diagrams, mechanism diagrams, principle diagrams, graphical abstracts, or 3D-style research illustrations.
• Image Edit: analyzes the user's edit request and reference sources, then drives an image-editing prompt path.

The project stores sessions, uploads, jobs, prompts, plans, events, and artifacts persistently in SQLite plus a local artifact directory.

Repository Policy

This repository does not include SenseNova model weights, generated showcase images, private API keys, or user uploads. Keep large weights such as .safetensors outside git.

The repository code is MIT licensed. External SenseNova source checkouts, SenseNova model weights, user uploads, generated outputs, and provider API accounts keep their own ownership and license terms.

See NOTICE, OWNERSHIP_AND_LICENSE.md, and THIRD_PARTY_NOTICES.md before publishing or redistributing a fork.

Quick Start

cd Open_Notebook
python3.11 -m venv .venv311
source .venv311/bin/activate
pip install -U pip -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install -e ".[dev,local-u1]" -i https://pypi.tuna.tsinghua.edu.cn/simple
cp .env.example .env

Start the Web UI with Python only:

python start.py

Open http://127.0.0.1:8017.

Runtime does not require Node.js when frontend/dist is present. Node.js is only needed if you want to modify and rebuild the frontend.

For real SenseNova image generation, set:

SN_IMAGE_GEN_API_KEY="..."
SN_IMAGE_GEN_BASE_URL="https://token.sensenova.cn/v1"
SN_IMAGE_GEN_MODEL="sensenova-u1-fast"
SN_CHAT_API_KEY="..."
SN_CHAT_BASE_URL="https://token.sensenova.cn/v1"
SN_CHAT_MODEL="sensenova-6.7-flash-lite"

For a no-key local smoke test:

python start.py --fake-image

Frontend development, optional:

cd frontend
npm install
npm run dev

Build

cd frontend
npm run build
cd ..
python start.py

If frontend/dist exists, FastAPI serves it from /.

LLM Configuration

The backend reimplements the Clouds_Coder-style LLM profile model:

• profiles, model_profiles, llm_profiles, or llms
• flat keys like openai_url, openai_model, openai_key
• providers: openai_compat, siliconflow, vllm, lmstudio, anthropic, glm, kimi, openrouter, custom_http, ollama
• active/default profile selection
• media/capability declarations

You can import or export this JSON from the GUI.

Local SenseNova-U1

Local U1 inference is optional. For GitHub use, keep weights outside git and point to them through .env. If models/Full contains a complete local U1 checkpoint, python start.py automatically selects the local U1 image backend.

The Web UI includes a SenseNova model panel with this download link:

https://huggingface.co/sensenova/SenseNova-U1-8B-MoT

Clicking download stores the model under the gitignored relative directory models/Full.

The SenseNova-U1 source checkout is kept outside this repository. The default relative path is ../SenseNova-U1-main/src. If that directory is missing, Open_Notebook can clone https://github.com/OpenSenseNova/SenseNova-U1 into ../SenseNova-U1-main on demand when checking model status or using local U1.

Relative paths are resolved from the repository root:

OPEN_NOTEBOOK_IMAGE_BACKEND=local_u1
SENSENOVA_U1_MODEL_PATH=models/Full
SENSENOVA_U1_SOURCE_ROOT=../SenseNova-U1-main/src
SENSENOVA_U1_DEVICE=mps

You can also use absolute paths locally, but do not commit them. The external SenseNova-U1 checkout and model weights are not covered by this repository's MIT license; follow the upstream license and model terms. For local U1 inference, follow the upstream SenseNova-U1 dependency guidance; Python 3.11 is the recommended model runtime even though the Web UI itself can run on Python 3.10+.

Open_Notebook uses the official SenseNova-U1 transformers T2I defaults for local generation:

SENSENOVA_U1_DTYPE=bfloat16
SENSENOVA_U1_NUM_STEPS=50
SENSENOVA_U1_CFG_SCALE=4.0
SENSENOVA_U1_CFG_NORM=none
SENSENOVA_U1_TIMESTEP_SHIFT=3.0
SENSENOVA_U1_CFG_INTERVAL=0.0,1.0

The upstream local runtime is validated around Python >=3.11,<3.12, torch==2.8.0, and transformers==4.57.1. CUDA is the recommended backend. macOS MPS is treated as experimental; if it produces flat, NaN, or patch-noise images, Open_Notebook rejects the result instead of exporting it into a deck.

Output Layout

Runtime files are under data/ by default:

data/
  open_notebook.sqlite3
  projects/<session_id>/
    sources/
    jobs/<job_id>/
      sources_digest.json
      plan.json
      prompts/
      images/
      exports/

data/ is ignored by git.

Development Checks

pytest
ruff check backend tests