blip-2-vision-language

BLIP-2：视觉-语言预训练

使用 Salesforce 的 BLIP-2 进行视觉-语言任务的综合指南，采用冻结的图像编码器和大模型。

何时使用 BLIP-2

在以下情况下使用 BLIP-2：

• 需要带有自然描述的高质量图像描述生成
• 构建视觉问答（VQA）系统
• 需要零样本图像-文本理解，无需特定任务训练
• 希望利用大模型推理能力处理视觉任务
• 构建多模态对话式 AI
• 需要图像-文本检索或匹配

核心特性：

• Q-Former 架构：轻量级查询 Transformer 连接视觉和语言
• 冻结骨干网络效率：无需微调大型视觉/语言模型
• 多种大模型后端：OPT（2.7B、6.7B）和 FlanT5（XL、XXL）
• 零样本能力：无需特定任务训练即可获得强劲性能
• 高效训练：仅训练 Q-Former（约 188M 参数）
• 最先进的结果：在 VQA 基准测试上超越更大的模型

改用替代方案的情况：

• LLaVA：用于指令跟随的多模态对话
• InstructBLIP：用于改进的指令跟随（BLIP-2 后继者）
• GPT-4V/Claude 3：用于生产环境的多模态对话（专有）
• CLIP：用于简单的图像-文本相似度计算，无需生成
• Flamingo：用于少样本视觉学习

快速开始

安装

# HuggingFace Transformers (recommended)
pip install transformers accelerate torch Pillow

# Or LAVIS library (Salesforce official)
pip install salesforce-lavis

基础图像描述生成

import torch
from PIL import Image
from transformers import Blip2Processor, Blip2ForConditionalGeneration

# Load model and processor
processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
model = Blip2ForConditionalGeneration.from_pretrained(
    "Salesforce/blip2-opt-2.7b",
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load image
image = Image.open("photo.jpg").convert("RGB")

# Generate caption
inputs = processor(images=image, return_tensors="pt").to("cuda", torch.float16)
generated_ids = model.generate(**inputs, max_new_tokens=50)
caption = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(caption)

视觉问答

# Ask a question about the image
question = "What color is the car in this image?"

inputs = processor(images=image, text=question, return_tensors="pt").to("cuda", torch.float16)
generated_ids = model.generate(**inputs, max_new_tokens=50)
answer = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(answer)

使用 LAVIS 库

import torch
from lavis.models import load_model_and_preprocess
from PIL import Image

# Load model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model, vis_processors, txt_processors = load_model_and_preprocess(
    name="blip2_opt",
    model_type="pretrain_opt2.7b",
    is_eval=True,
    device=device
)

# Process image
image = Image.open("photo.jpg").convert("RGB")
image = vis_processors["eval"](image).unsqueeze(0).to(device)

# Caption
caption = model.generate({"image": image})
print(caption)

# VQA
question = txt_processors["eval"]("What is in this image?")
answer = model.generate({"image": image, "prompt": question})
print(answer)

核心概念

架构概览

BLIP-2 Architecture:
┌─────────────────────────────────────────────────────────────┐
│                        Q-Former                              │
│  ┌─────────────────────────────────────────────────────┐    │
│  │     Learned Queries (32 queries × 768 dim)          │    │
│  └────────────────────────┬────────────────────────────┘    │
│                           │                                  │
│  ┌────────────────────────▼────────────────────────────┐    │
│  │    Cross-Attention with Image Features               │    │
│  └────────────────────────┬────────────────────────────┘    │
│                           │                                  │
│  ┌────────────────────────▼────────────────────────────┐    │
│  │    Self-Attention Layers (Transformer)               │    │
│  └────────────────────────┬────────────────────────────┘    │
└───────────────────────────┼─────────────────────────────────┘
                            │
┌───────────────────────────▼─────────────────────────────────┐
│  Frozen Vision Encoder    │      Frozen LLM                  │
│  (ViT-G/14 from EVA-CLIP) │      (OPT or FlanT5)            │
└─────────────────────────────────────────────────────────────┘

模型变体

模型	大模型后端	大小	使用场景
blip2-opt-2.7b	OPT-2.7B	~4GB	通用图像描述生成、VQA
blip2-opt-6.7b	OPT-6.7B	~8GB	更好的推理能力
blip2-flan-t5-xl	FlanT5-XL	~5GB	指令跟随
blip2-flan-t5-xxl	FlanT5-XXL	~13GB	最佳质量

Q-Former 组件

组件	描述	参数
学习查询	固定的可学习嵌入集合	32 × 768
图像 Transformer	对视觉特征的交叉注意力	~108M
文本 Transformer	文本的自注意力	~108M
线性投影	映射到大模型维度	不定

高级用法

批处理

from PIL import Image
import torch

# Load multiple images
images = [Image.open(f"image_{i}.jpg").convert("RGB") for i in range(4)]
questions = [
    "What is shown in this image?",
    "Describe the scene.",
    "What colors are prominent?",
    "Is there a person in this image?"
]

# Process batch
inputs = processor(
    images=images,
    text=questions,
    return_tensors="pt",
    padding=True
).to("cuda", torch.float16)

# Generate
generated_ids = model.generate(**inputs, max_new_tokens=50)
answers = processor.batch_decode(generated_ids, skip_special_tokens=True)

for q, a in zip(questions, answers):
    print(f"Q: {q}\nA: {a}\n")

控制生成

# Control generation parameters
generated_ids = model.generate(
    **inputs,
    max_new_tokens=100,
    min_length=20,
    num_beams=5,              # Beam search
    no_repeat_ngram_size=2,   # Avoid repetition
    top_p=0.9,                # Nucleus sampling
    temperature=0.7,          # Creativity
    do_sample=True,           # Enable sampling
)

# For deterministic output
generated_ids = model.generate(
    **inputs,
    max_new_tokens=50,
    num_beams=5,
    do_sample=False,
)

内存优化

# 8-bit quantization
from transformers import BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(load_in_8bit=True)

model = Blip2ForConditionalGeneration.from_pretrained(
    "Salesforce/blip2-opt-6.7b",
    quantization_config=quantization_config,
    device_map="auto"
)

# 4-bit quantization (more aggressive)
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16
)

model = Blip2ForConditionalGeneration.from_pretrained(
    "Salesforce/blip2-flan-t5-xxl",
    quantization_config=quantization_config,
    device_map="auto"
)

图像-文本匹配

# Using LAVIS for ITM (Image-Text Matching)
from lavis.models import load_model_and_preprocess

model, vis_processors, txt_processors = load_model_and_preprocess(
    name="blip2_image_text_matching",
    model_type="pretrain",
    is_eval=True,
    device=device
)

image = vis_processors["eval"](raw_image).unsqueeze(0).to(device)
text = txt_processors["eval"]("a dog sitting on grass")

# Get matching score
itm_output = model({"image": image, "text_input": text}, match_head="itm")
itm_scores = torch.nn.functional.softmax(itm_output, dim=1)
print(f"Match probability: {itm_scores[:, 1].item():.3f}")

特征提取

# Extract image features with Q-Former
from lavis.models import load_model_and_preprocess

model, vis_processors, _ = load_model_and_preprocess(
    name="blip2_feature_extractor",
    model_type="pretrain",
    is_eval=True,
    device=device
)

image = vis_processors["eval"](raw_image).unsqueeze(0).to(device)

# Get features
features = model.extract_features({"image": image}, mode="image")
image_embeds = features.image_embeds  # Shape: [1, 32, 768]
image_features = features.image_embeds_proj  # Projected for matching

常见工作流

工作流 1：图像描述生成流水线

import torch
from PIL import Image
from transformers import Blip2Processor, Blip2ForConditionalGeneration
from pathlib import Path

class ImageCaptioner:
    def __init__(self, model_name="Salesforce/blip2-opt-2.7b"):
        self.processor = Blip2Processor.from_pretrained(model_name)
        self.model = Blip2ForConditionalGeneration.from_pretrained(
            model_name,
            torch_dtype=torch.float16,
            device_map="auto"
        )

    def caption(self, image_path: str, prompt: str = None) -> str:
        image = Image.open(image_path).convert("RGB")

        if prompt:
            inputs = self.processor(images=image, text=prompt, return_tensors="pt")
        else:
            inputs = self.processor(images=image, return_tensors="pt")

        inputs = inputs.to("cuda", torch.float16)

        generated_ids = self.model.generate(
            **inputs,
            max_new_tokens=50,
            num_beams=5
        )

        return self.processor.decode(generated_ids[0], skip_special_tokens=True)

    def caption_batch(self, image_paths: list, prompt: str = None) -> list:
        images = [Image.open(p).convert("RGB") for p in image_paths]

        if prompt:
            inputs = self.processor(
                images=images,
                text=[prompt] * len(images),
                return_tensors="pt",
                padding=True
            )
        else:
            inputs = self.processor(images=images, return_tensors="pt", padding=True)

        inputs = inputs.to("cuda", torch.float16)

        generated_ids = self.model.generate(**inputs, max_new_tokens=50)
        return self.processor.batch_decode(generated_ids, skip_special_tokens=True)

# Usage
captioner = ImageCaptioner()

# Single image
caption = captioner.caption("photo.jpg")
print(f"Caption: {caption}")

# With prompt for style
caption = captioner.caption("photo.jpg", "a detailed description of")
print(f"Detailed: {caption}")

# Batch processing
captions = captioner.caption_batch(["img1.jpg", "img2.jpg", "img3.jpg"])
for i, cap in enumerate(captions):
    print(f"Image {i+1}: {cap}")

工作流 2：视觉问答系统

class VisualQA:
    def __init__(self, model_name="Salesforce/blip2-flan-t5-xl"):
        self.processor = Blip2Processor.from_pretrained(model_name)
        self.model = Blip2ForConditionalGeneration.from_pretrained(
            model_name,
            torch_dtype=torch.float16,
            device_map="auto"
        )
        self.current_image = None
        self.current_inputs = None

    def set_image(self, image_path: str):
        """Load image for multiple questions."""
        self.current_image = Image.open(image_path).convert("RGB")

    def ask(self, question: str) -> str:
        """Ask a question about the current image."""
        if self.current_image is None:
            raise ValueError("No image set. Call set_image() first.")

        # Format question for FlanT5
        prompt = f"Question: {question} Answer:"

        inputs = self.processor(
            images=self.current_image,
            text=prompt,
            return_tensors="pt"
        ).to("cuda", torch.float16)

        generated_ids = self.model.generate(
            **inputs,
            max_new_tokens=50,
            num_beams=5
        )

        return self.processor.decode(generated_ids[0], skip_special_tokens=True)

    def ask_multiple(self, questions: list) -> dict:
        """Ask multiple questions about current image."""
        return {q: self.ask(q) for q in questions}

# Usage
vqa = VisualQA()
vqa.set_image("scene.jpg")

# Ask questions
print(vqa.ask("What objects are in this image?"))
print(vqa.ask("What is the weather like?"))
print(vqa.ask("How many people are there?"))

# Batch questions
results = vqa.ask_multiple([
    "What is the main subject?",
    "What colors are dominant?",
    "Is this indoors or outdoors?"
])

工作流 3：图像搜索/检索

import torch
import numpy as np
from PIL import Image
from lavis.models import load_model_and_preprocess

class ImageSearchEngine:
    def __init__(self):
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.model, self.vis_processors, self.txt_processors = load_model_and_preprocess(
            name="blip2_feature_extractor",
            model_type="pretrain",
            is_eval=True,
            device=self.device
        )
        self.image_features = []
        self.image_paths = []

    def index_images(self, image_paths: list):
        """Build index from images."""
        self.image_paths = image_paths

        for path in image_paths:
            image = Image.open(path).convert("RGB")
            image = self.vis_processors["eval"](image).unsqueeze(0).to(self.device)

            with torch.no_grad():
                features = self.model.extract_features({"image": image}, mode="image")
                # Use projected features for matching
                self.image_features.append(
                    features.image_embeds_proj.mean(dim=1).cpu().numpy()
                )

        self.image_features = np.vstack(self.image_features)

    def search(self, query: str, top_k: int = 5) -> list:
        """Search images by text query."""
        # Get text features
        text = self.txt_processors["eval"](query)
        text_input = {"text_input": [text]}

        with torch.no_grad():
            text_features = self.model.extract_features(text_input, mode="text")
            text_embeds = text_features.text_embeds_proj[:, 0].cpu().numpy()

        # Compute similarities
        similarities = np.dot(self.image_features, text_embeds.T).squeeze()
        top_indices = np.argsort(similarities)[::-1][:top_k]

        return [(self.image_paths[i], similarities[i]) for i in top_indices]

# Usage
engine = ImageSearchEngine()
engine.index_images(["img1.jpg", "img2.jpg", "img3.jpg", ...])

# Search
results = engine.search("a sunset over the ocean", top_k=5)
for path, score in results:
    print(f"{path}: {score:.3f}")

输出格式

生成输出

# Direct generation returns token IDs
generated_ids = model.generate(**inputs, max_new_tokens=50)
# Shape: [batch_size, sequence_length]

# Decode to text
text = processor.batch_decode(generated_ids, skip_special_tokens=True)
# Returns: list of strings

特征提取输出

# Q-Former outputs
features = model.extract_features({"image": image}, mode="image")

features.image_embeds          # [B, 32, 768] - Q-Former outputs
features.image_embeds_proj     # [B, 32, 256] - Projected for matching
features.text_embeds          # [B, seq_len, 768] - Text features
features.text_embeds_proj     # [B, 256] - Projected text (CLS)

性能优化

GPU 内存需求

模型	FP16 VRAM	INT8 VRAM	INT4 VRAM
blip2-opt-2.7b	~8GB	~5GB	~3GB
blip2-opt-6.7b	~16GB	~9GB	~5GB
blip2-flan-t5-xl	~10GB	~6GB	~4GB
blip2-flan-t5-xxl	~26GB	~14GB	~8GB

速度优化

# Use Flash Attention if available
model = Blip2ForConditionalGeneration.from_pretrained(
    "Salesforce/blip2-opt-2.7b",
    torch_dtype=torch.float16,
    attn_implementation="flash_attention_2",  # Requires flash-attn
    device_map="auto"
)

# Compile model (PyTorch 2.0+)
model = torch.compile(model)

# Use smaller images (if quality allows)
processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
# Default is 224x224, which is optimal

常见问题

问题	解决方案
CUDA 内存不足	使用 INT8/INT4 量化，更小的模型
生成缓慢	使用贪心解码，减少 max_new_tokens
描述质量差	尝试 FlanT5 变体，使用提示词
幻觉	降低温度，使用束搜索
答案错误	重新表述问题，提供上下文

参考资料

• 高级用法 - 微调、集成、部署
• 故障排除 - 常见问题与解决方案

资源

• 论文：https://arxiv.org/abs/2301.12597
• GitHub（LAVIS）：https://github.com/salesforce/LAVIS
• HuggingFace：https://huggingface.co/Salesforce/blip2-opt-2.7b
• 演示：https://huggingface.co/spaces/Salesforce/BLIP2
• InstructBLIP：https://arxiv.org/abs/2305.06500（后继者）