ComfyUI-SenseNova-SI

ComfyUI custom node wrapper for SenseNova-SI VLM inference.

Features

• Loads SenseNova-SI Qwen and InternVL models from a local source checkout.
• Caches loaded model handles so repeated ComfyUI runs do not rebuild the model unless requested.
• Accepts batched ComfyUI IMAGE tensors and converts them to the format expected by each model type.
• Provides an image-list helper node for multi-image prompts.

Layout

ComfyUI/custom_nodes/ComfyUI-SenseNova-SI/
  __init__.py
  nodes.py
  utils.py
  docs/
    single.png
    multi.png
  example_workflows/
    sensenova-si.json
  deps/
    SenseNova-SI/

This custom node expects a source checkout of SenseNova-SI. By default it looks for:

1. repo_path from the loader node
2. SENSENOVA_SI_REPO environment variable
3. deps/SenseNova-SI/
4. The parent directory of this custom node, which makes local development inside the upstream repo work

Requirements

• Python 3.10 or newer.
• A working ComfyUI installation.
• A local checkout of SenseNova-SI.
• A GPU environment that can run the selected SenseNova-SI model.

Install

1. Clone ComfyUI:

   git clone https://github.com/comfyanonymous/ComfyUI.git
   cd ComfyUI

2. Clone this custom node:

   cd custom_nodes
   git clone https://github.com/GACLove/ComfyUI-SenseNova-SI.git

3. Clone or link SenseNova-SI into deps/:

   cd ComfyUI-SenseNova-SI
   git clone https://github.com/OpenSenseNova/SenseNova-SI.git deps/SenseNova-SI

4. Install dependencies:

   pip install -r requirements.txt

5. Install the upstream SenseNova-SI dependencies required by the model you plan to run. Follow the upstream README for model-specific setup and checkpoints.

6. Start ComfyUI:

   cd ../../..  # back to ComfyUI root
   python main.py --listen 0.0.0.0 --port 8000 --cuda-device 0

Example Workflow

An example workflow is available at example_workflows/sensenova-si.json.

To use it, copy or import the workflow in ComfyUI, update the loader node's model_path and repo_path if needed, and connect your input image nodes.

Single-image workflow:

Multi-image workflow:

Nodes

SenseNova SI Loader

Loads a SenseNova-SI model once and caches it by:

• repo path
• model path
• dtype
• device map
• generation config path

Use force_reload when you need to rebuild the model object.

Inputs:

Input	Description
model_path	Local path or Hugging Face model ID passed to the upstream SenseNova-SI model class.
model_type	auto, qwen, or internvl. Auto-detection uses the model path.
repo_path	Optional local path to the SenseNova-SI source checkout.
dtype	Qwen dtype option: auto, bfloat16, float16, or float32.
device_map	Qwen device map passed through to the upstream loader.
generation_config_path	Optional absolute path or path relative to the SenseNova-SI repo root.
force_reload	Rebuilds the cached model handle when enabled.

SenseNova SI Generate

Runs repeated inference against the loaded model handle.

• IMAGE inputs are converted from ComfyUI tensors [B, H, W, C] into PIL.Image
• if the prompt does not contain <image>, the node prepends one token per image
• if the prompt already contains <image>, the count must match the image batch size
• extra_generation_kwargs_json overrides the exposed generation settings

SenseNova Image List

Aggregates up to 10 individual ComfyUI IMAGE inputs into one batched IMAGE output. Use this when a prompt should reference multiple images.

Notes

• Supports Qwen and InternVL model types (auto-detected from model path).
• The upstream SenseNova-SI repo is used as source, not installed as a Python package.
• InternVL generation receives temporary image files because the upstream InternVL wrapper expects file paths.
• This repository is MIT licensed. Check upstream SenseNova-SI model, code, and checkpoint licenses before redistributing models or generated artifacts.

Development

Run the same checks used by CI:

ruff check .
ruff format --check .

Format local changes with:

ruff format .

License

This project is licensed under the MIT License.