OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
Image & Video AI

Lisa

Lisa is a Python-based, large-language-model segmentation assistant that takes image questions or natural-language instructions and produces segmentation masks, with project examples also showing explanatory text responses.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Lisa — OpenViST project profile in Image & Video AI.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use Lisa when a visual task calls for a segmentation mask selected through a natural-language question or contextual description, and when local model inference or experimentation is relevant.

Who it is for

  • Developers exploring language-guided image segmentation and multimodal reasoning.
  • Researchers or practitioners who want to run the documented inference workflow with a LISA model.
  • Teams investigating training or adapting a segmentation assistant using the listed datasets.

Documented uses

  • Ask a question about an image and request a mask for the relevant person, object, or region.
  • Describe a target using a property or contextual relationship rather than only an object label.
  • Run the supplied inference script with a prompt and image path to generate a masked-image result.
  • Prepare a separate training workflow using the dataset categories listed in the README.
Before you choose

Requirements & limits

The repository is public and lists an Apache-2.0 license. The project page links to online demo deployments and documents local inference; the README does not specify account or pricing requirements.

  • The documented local inference route uses CUDA and provides GPU memory estimates for the 13B model.
  • The README offers 16-bit, 8-bit, and 4-bit inference options, with reduced generation quality noted for lower-bit loading.
  • Inference and local deployment are separate documented workflows, using chat.py and app.py respectively.
  • Training requires preparing several types of datasets, and the README says to follow the original datasets' rules.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Run local inference with a LISA 13B model

Prepare the documented environment and start the chat-based inference script with a LISA-13B-llama2-v1 model.

Before you start

  • The project repository and its requirements.
  • Python environment with CUDA and a GPU.
  • Flash Attention installed as described in the README.
  • A supported LISA-13B-llama2-v1 model version.

Steps

  1. Install dependencies with the command below.

    Documented command
    pip install -r requirements.txt

  2. Install Flash Attention with the command below.

    Documented command
    pip install flash-attn --no-build-isolation

  3. Start chat.py with a listed LISA-13B-llama2-v1 model version and choose the documented precision or quantized loading option.

  4. Enter a text prompt and the image path when prompted.

Expected outcome

The inference workflow produces a response to the prompt and can save a masked-image result.

Guide limitations

  • The README estimates 30 GB of VRAM for 16-bit inference, 16 GB for 8-bit, and 9 GB for 4-bit with the 13B model; lower-bit loading may reduce generation quality.
  • The documented chat script does not support the v0 models without checking out the specified legacy repository revision.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
13 October 2026
Original website title
JIA-Lab-research/LISA · Official homepage title · retained in its source language
How we review sourcesSuggest a correction