open-gigaai/giga-models

GigaModels – a toolbox of ready‑to‑run AI pipelines

What it is – GigaModels is an open‑source Python package that bundles a large collection of pre‑trained vision, diffusion and multimodal models behind a unified pipeline API. The repo ships code for loading, running inference and (for a few models) training, all from a single load_pipeline call.

Key ideas

• Unified interface – each model type (e.g., depth estimation, object detection, diffusion video generation) is wrapped in a *Pipeline class that follows the same call signature (pipeline(input)).
• Broad coverage – the repository lists more than 30 pipelines covering:
  • Vision‑Language‑Action (VLA) models such as GigaBrain‑0, Pi0, Pi0.5
  • Diffusion models for video prediction/transfer (GigaWorld‑0, Cosmos‑Predict2.5, Cosmos‑Transfer2.5)
  • Classic vision tasks: depth (Depth Anything, DPT), detection (Grounding DINO), edge detection (Canny, HED, Lineart, etc.), frame interpolation (Film), image restoration (PromptIR), keypoint detection (OpenPose, RTMPose), optical flow (UniMatch), segmentation (Grounded SAM 2, Segment Anything, UperNet), shot‑boundary detection (TransNetV2).
• Both inference and training – for the VLA and diffusion families the README links to separate inference scripts and training configs, enabling researchers to fine‑tune the models.
• Modular layout – the source tree is organized under giga_models/pipelines/<task>/<model>/pipeline_*.py and a projects/ folder that contains example scripts and config files.

Installation

conda create -n giga_models python=3.11.10
conda activate giga_models
git clone https://github.com/open-gigaai/giga-models.git
cd giga-models
pip install -e .

The package is installed in editable mode so you can edit pipelines locally.

Quick‑start example

from PIL import Image
from giga_models import load_pipeline

img = Image.open('my_photo.jpg')
# Object detection with Grounding DINO
pipe = load_pipeline('detection/grounding_dino/swint_ogc')
boxes, labels, scores = pipe(img, ['person', 'car'])

# Depth estimation with Depth Anything V2 Large
from giga_models import DepthAnythingPipeline
pipe = DepthAnythingPipeline('depth-anything/Depth-Anything-V2-Large-hf').to('cuda')
depth = pipe(img)

The same pattern works for any listed pipeline.

How to contribute – The project follows an Apache‑2.0 license and welcomes pull requests. A CONTRIBUTING.md file explains the coding style, testing, and how to add new pipelines.

Citation – If you use GigaModels in research, cite the repository as shown in the README.

---

Bottom line – GigaModels is a genuine, production‑ready collection of multimodal and perceptual AI models, aimed at developers who want a plug‑and‑play way to experiment with many state‑of‑the‑art vision and generative pipelines without handling each model’s boilerplate separately.