---
title: 'BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models'
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
license_note: 'paper is CC BY 4.0 (verified on arXiv abs page 2026-11-12; icon by-4.0.png, http://creativecommons.org/licenses/by/4.0/); note is our own summary'
source: https://iinventedlight.com/library/notes/2301-12597-blip-2.md
source_license: 'CC BY 4.0'
attribution: 'Summary by iinventedlight (AI-assisted). Source: Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi, BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models (CC BY 4.0).'
author: 'iinventedlight (AI-assisted)'
date: 2026-11-12
status: published
approved_by: "Cole Withington, 2026-11-12 14:44 PT, Librarian chat widget t51s316: \"Yes, publish all 100 second-list arXiv notes\""
checked_on: 2026-11-12
as_of: 2026-11-12
confidence: medium
untested: 'Not reproduced. Abs page and HTML key sections read 2026-11-12. Named code URLs checked for HTTP 200; code not run.'
category: arxiv
topic_area: 'ai-research: multimodal'
tags: arxiv, multimodal, vision-language
arxiv_id: 2301.12597
alias: 2301.12597-blip-2
sources:
  - id: S1
    title: 'BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models'
    publisher: arXiv
    authors: 'Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi'
    url: https://arxiv.org/abs/2301.12597
    version_read: 'v3 (15 Jun 2023)'
    accessed: 2026-11-12
    license: 'CC BY 4.0'
---
# BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models

## Summary
BLIP-2 bootstraps vision-language pre-training from frozen pre-trained image encoders and frozen large language models. It bridges the modality gap with a lightweight Querying Transformer, pre-trained in two stages. The authors report state-of-the-art results with fewer trainable parameters, including outperforming Flamingo80B by 8.7% on zero-shot VQAv2 with 54x fewer trainable parameters. [S1]

## Claims
- id: C1
  claim: 'BLIP-2 bootstraps vision-language pre-training from frozen image encoders and frozen language models.'
  source: S1
  evidence: "bootstraps vision-language pre-training from off-the-shelf frozen pre-trained image encoders and frozen large language models"
- id: C2
  claim: 'A lightweight Querying Transformer is pre-trained in two stages.'
  source: S1
  evidence: "a lightweight Querying Transformer, which is pre-trained in two stages"
- id: C3
  claim: 'The authors report outperforming Flamingo80B by 8.7% on zero-shot VQAv2 with 54x fewer trainable parameters.'
  source: S1
  evidence: "outperforms Flamingo80B by 8.7% on zero-shot VQAv2 with 54x fewer trainable parameters"

## Verification
- Code available: Paper: https://github.com/salesforce/LAVIS/tree/main/projects/blip2, which returned HTTP 200 on 2026-11-12. Not run.
- Reproduced: not reproduced (2026-11-12). Code not run.
- Related work: Live notes [2103-00020-clip](https://iinventedlight.com/library/notes/2103-00020-clip/) and [2304-08485-visual-instruction-tuning](https://iinventedlight.com/library/notes/2304-08485-visual-instruction-tuning/).
- Limits: The 8.7% and 54x figures are the authors' reported comparison with Flamingo80B on zero-shot VQAv2 (v3, 15 Jun 2023).
- License: CC BY 4.0 on the abs page (2026-11-12; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 30 Jan 2023 (v1); last revised 15 Jun 2023 (v3). Sources: [S1] https://arxiv.org/abs/2301.12597, v3 (15 Jun 2023). checked_on: 2026-11-12.

## When to apply
Use when asking how a frozen image encoder was connected to a frozen language model with a Querying Transformer.

## Facts
- id: F1
  claim: 'BLIP-2 connects a frozen image encoder to a frozen language model with a lightweight Querying Transformer.'
  confidence: medium
  date: 2026-11-12
  sources: [S1]
  license: CC BY 4.0
- id: F2
  claim: 'The Querying Transformer is pre-trained in two stages, first for representation learning and then for generative learning.'
  confidence: medium
  date: 2026-11-12
  sources: [S1]
  license: CC BY 4.0
- id: F3
  claim: 'The authors report an 8.7% gain over Flamingo80B on zero-shot VQAv2 with 54x fewer trainable parameters.'
  confidence: medium
  date: 2026-11-12
  sources: [S1]
  license: CC BY 4.0

## Sources
- [S1] arXiv 2301.12597, https://arxiv.org/abs/2301.12597. License: CC BY 4.0. v3 (15 Jun 2023). Accessed 2026-11-12.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.

---

Agents: [/agents.md](/agents.md) | [/.well-known/agent-library.json](/.well-known/agent-library.json).
