Library / SDK
salesforce/LAVIS avatar
salesforce/LAVIS

LAVIS froze at v1.0.2 with transformers pinned to 4.33.2

LAVIS - A One-stop Library for Language-Vision Intelligence

11,254 stars1,107 forksJupyter NotebookBSD-3-Clause

At a glance

What is it?
LAVIS gave multimodal researchers one Python interface over captioning, VQA, retrieval and 20+ datasets, backed by 30+ pretrained weights from ALBEF, BLIP, ALPRO and CLIP. The repository is archived, its release line stops at v1.0.2 from March 2023, and its exact dependency pins make it something to read and fork rather than adopt.
Who is it for?
LAVIS suits researchers who need to read or extend one specific BLIP-2 or InstructBLIP code path and can dedicate an environment to its pinned stack. It does not suit teams standardising on a current transformers release, or shops that need a maintained security contact.
Can I use it commercially?
Yes. BSD-3-Clause is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
No. The owners have archived the repository on GitHub, so it is read-only and no longer receives changes.
What is it written in?
Mainly Jupyter Notebook, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • salesforce/LAVIS11,254 ★ · -9
Star history
Repository2026-09-112026-11-15
salesforce/LAVIS11.3k11.3k

-9 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

The repository is archived while the release line stops at v1.0.2

GitHub reports the repository as archived, and the last push landed on 2026-09-18. That date sits more than three years after v1.0.2 was published on 2023-03-06, which came three days after v1.0.1 and a month after v1.0.0. So the last push and the newest tag are far apart. setup.py still declares version="1.0.1", one behind the newest tag, which means a build from source produces 1.0.1 metadata regardless of which commit you check out.

The root of the tree carries a file named CONTRIBUTING-ARCHIVED.md next to CODE_OF_CONDUCT.md and SECURITY.md. That filename is the clearest statement of the project's state available to a reader, more informative than any changelog entry. An archived repository receives no further maintainer cut, so the pin set described below is the pin set you get permanently unless you fork it. Plan on owning every dependency fix yourself,.

Six model families landed between October 2022 and November 2023

The release notes walk a dense timeline. PNP-VQA arrived in October 2022 as a modular zero-shot VQA framework requiring no PLM training. Img2LLM-VQA followed in December 2022, a CVPR 2023 paper, as a plug-and-play module that puts off-the-shelf LLMs to work on visual question answering. BLIP-2 landed in January 2023 in the same month the package appeared on PyPI, then InstructBLIP in May 2023, BLIP-Diffusion in July 2023, and X-InstructBLIP in November 2023.

What that sequence buys is breadth inside one interface: 10+ tasks spanning retrieval, captioning, visual question answering and multimodal classification, 20+ datasets including COCO, Flickr, Nocaps, Conceptual Commons and SBU, and 30+ pretrained weights covering ALBEF, BLIP, ALPRO and CLIP. The directory layout backs that claim, with a project page per model under projects/ and fourteen worked notebooks under examples/, among them blip2_instructed_generation.ipynb and clip_zero_shot_classification.ipynb. The newest release noted in the README is from November 2023, more than two years ago.

requirements.txt pins transformers and timm to exact versions

Dependencies are read straight out of requirements.txt by setup.py, so every line in that file becomes an install requirement. The middle of the file is where the constraint bites:

bash
scikit-image
sentencepiece
spacy
streamlit
timm==0.4.12
torch>=1.10.0
torchvision
tqdm
transformers==4.33.2
webdataset
wheel

transformers==4.33.2 and timm==0.4.12 are exact, diffusers<=0.16.0 is a ceiling, and torch>=1.10.0 is only a floor. The pins cost you co-installation. Any project sharing the environment that needs a newer transformers or timm cannot resolve alongside LAVIS, and you have to decide which one gets a separate environment.

Two details deserve more attention than they usually get. The same file pins fairscale==0.4.4, opencv-python-headless==4.5.5.64 and open3d==0.13.0, and it lists pre-commit and wheel, which are development tools, as install requirements.

Only the lavis package ships, so train.py and evaluate.py stay behind

The packaging is deliberately narrow. These consecutive lines decide exactly what an install gives you:

python
    packages=find_namespace_packages(include="lavis.*"),
    install_requires=fetch_requirements("requirements.txt"),
    python_requires=">=3.7.0",
    include_package_data=True,
    dependency_links=DEPENDENCY_LINKS,
    zip_safe=False,
)

The include filter restricts discovery to the lavis namespace. The repository root holds train.py, evaluate.py, run_scripts/, app/, dataset_card/, docs/ and tests/, and none of those sit inside that namespace, so an installed copy of salesforce-lavis provides the library and its weights interface but not the training or evaluation entry points. To run them you need a source checkout of the whole tree.

That split has a direct consequence for automation. A pipeline intending to finetune a BLIP-2 checkpoint cannot build its image from the installed package alone. It has to clone the repository, install in editable mode, and accept that the clone it depends on is archived and will never receive a fix. Budget for that arrangement rather than discovering it during a release.

Windows is the only platform branch, and PyPI is the only stated route in

The January 2023 note records LAVIS as available on PyPI, and setup.py gives the distribution name as salesforce-lavis.

One platform difference is encoded directly in setup.py:

python
DEPENDENCY_LINKS = []
if platform.system() == "Windows":
    DEPENDENCY_LINKS.append("https://download.pytorch.org/whl/torch_stable.html")


def fetch_requirements(filename):
    with open(filename) as f:
        return [ln.strip() for ln in f.read().split("\n")]

The PyTorch wheel index is surfaced only when the platform is Windows.. On Linux and macOS the torch>=1.10.0 requirement resolves from whichever index pip is configured against, so GPU selection becomes your decision and your problem. Combined with a floor of >=3.7.0 in python_requires, that leaves the interpreter and accelerator pair determined in practice by which pinned wheels still build.

The zero-shot VQAv2 figures are citations, not something the tree reproduces

The headline numbers arrive as references to published results rather than as something you can execute. The January 2023 BLIP-2 note claims a zero-shot VQAv2 score of 65.0 against Flamingo's 56.3, and 121.6 CIDEr on NoCaps against a previous best of 113.2. The December 2022 Img2LLM-VQA note claims 61.9 against that same 56.3 baseline while requiring no end-to-end training, and the October 2022 PNP-VQA note describes zero-shot VQA with no PLM training at all.

The approach difference behind those comparisons is real and worth stating plainly, since Flamingo is the common reference point. Flamingo is a trained model family, whereas BLIP-2 harvests pretrained vision models and large language models such as OPT and FlanT5 into one framework, and Img2LLM-VQA and PNP-VQA target off-the-shelf LLMs with zero training. That is the architectural claim. The scores themselves are figures quoted in the README's release notes. The benchmark link points at opensource.salesforce.com rather than at a configuration file in the tree, even though evaluate.py and run_scripts/ are both present.

Two dead ends: non-image modalities and checkpoint retrieval

Two things are worth checking early.

The first is modality. X-InstructBLIP is described as a cross-modality framework built atop frozen LLMs that integrates image, video, audio and 3D, and the README points to a project page under projects/xinstructblip for it, the same way PNP-VQA, Img2LLM-VQA and BLIP-2 are each organised.

The second is model retrieval. Check how the pretrained weights are fetched before attempting a container build or an air-gapped install. The repository still carries a SECURITY.md, and on an archived repository there is no maintainer left to receive a report, which deserves weight before you place any of these models in front of untrusted input.

Editorial conclusion

LAVIS suits researchers who need to read or extend one specific BLIP-2 or InstructBLIP code path and can dedicate an environment to its pinned stack. It does not suit teams standardising on a current transformers release, or shops that need a maintained security contact. Before adopting, verify that transformers 4.33.2 and timm 0.4.12 install on your Python and CUDA combination, check how the 30+ pretrained checkpoints are retrieved, and read setup.py to confirm that only the lavis package ships before you plan a CI build.

Frequently asked questions

How do you install LAVIS?

The project records LAVIS as available on PyPI, and setup.py names the distribution salesforce-lavis with a python_requires floor of >=3.7.0.

Is the LAVIS repository still maintained?

No. GitHub reports the repository as archived, and the contributing guide is named CONTRIBUTING-ARCHIVED.md. The last push was on 2026-09-18, while the newest tagged release is v1.0.2 from 2023-03-06 and setup.py still declares version 1.0.1.

Which models and tasks does LAVIS include?

The release notes cover BLIP-2, InstructBLIP, BLIP-Diffusion, X-InstructBLIP, Img2LLM-VQA and PNP-VQA, and the weight set includes ALBEF, BLIP, ALPRO and CLIP. The interface covers 10+ tasks such as retrieval, captioning, visual question answering and multimodal classification, across 20+ datasets including COCO, Flickr, Nocaps, Conceptual Commons and SBU.

What dependency versions does LAVIS require?

setup.py feeds requirements.txt straight into install_requires. That file pins transformers to 4.33.2 and timm to 0.4.12 exactly, caps diffusers at <=0.16.0, fixes fairscale at 0.4.4 and opencv-python-headless at 4.5.5.64, and requires torch>=1.10.0.

Can LAVIS be installed alongside other PyTorch projects?

Rarely in the same environment. Because requirements.txt becomes install_requires unchanged, the exact pins on transformers==4.33.2 and timm==0.4.12 must be satisfied, so any project needing a newer release of either library needs a separate environment or a fork.

Does LAVIS reproduce the zero-shot VQAv2 scores it quotes?

The 65.0 and 61.9 zero-shot VQAv2 figures and the 121.6 NoCaps CIDEr score are stated in the README's release notes as results from the linked papers. evaluate.py and run_scripts/ exist in the tree.

Official sources

  1. Issues
  2. License: BSD-3-Clause
  3. README
  4. Releases
  5. salesforce/LAVIS on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/salesforce-lavis.svg)](https://hysenlabs.com/projects/salesforce-lavis)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/salesforce-lavis.svg)](https://hysenlabs.com/projects/salesforce-lavis)

Was this useful?