[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M72FKQHFWS89BK5P8M9HC37H":3,"skill-same:01M72FKQHFWS89BK5P8M9HC37H":67},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":36,"commitSha":37,"skillDirectory":38,"forksCount":14,"repositoryHtmlUrl":39,"repositoryHomepageUrl":6,"ownerHtmlUrl":40,"ownerBlogUrl":6,"ownerDescription":41,"updatedAt":42,"install":43,"resources":56},"01M72FKQHFWS89BK5P8M9HC37H","blip-2-vision-language",null,"Vision-language pre-training framework bridging frozen image encoders and LLMs. Use when you need image captioning, visual question answering, image-text retrieval, or multimodal chat with state-of-the-art zero-shot performance.","cocyuhao\u002Fai-dev-tooling-research-archive","cocyuhao","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F269083740?v=4","01M72F908ET5R76KP9NXS6Y2GJ","Hao Yu","01M72F909376KCNJVKG8MEPY2Q",0,"MIT","https:\u002F\u002Fgithub.com\u002Fcocyuhao\u002Fai-dev-tooling-research-archive\u002Fblob\u002Fmain\u002Fclaude-code-templates\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fai-research\u002Fmultimodal-blip-2\u002FSKILL.md",[],[],"2026-11-10 18:39:20",false,"---\nname: blip-2-vision-language\ndescription: Vision-language pre-training framework bridging frozen image encoders and LLMs. Use when you need image captioning, visual question answering, image-text retrieval, or multimodal chat with state-of-the-art zero-shot performance.\nversion: 1.0.0\nauthor: Orchestra Research\nlicense: MIT\ntags: [Multimodal, Vision-Language, Image Captioning, VQA, Zero-Shot]\ndependencies: [transformers>=4.30.0, torch>=1.10.0, Pillow]\n---\n\n# BLIP-2: Vision-Language Pre-training\n\nComprehensive guide to using Salesforce's BLIP-2 for vision-language tasks with frozen image encoders and large language models.\n\n## When to use BLIP-2\n\n**Use BLIP-2 when:**\n- Need high-quality image captioning with natural descriptions\n- Building visual question answering (VQA) systems\n- Require zero-shot image-text understanding without task-specific training\n- Want to leverage LLM reasoning for visual tasks\n- Building multimodal conversational AI\n- Need image-text retrieval or matching\n\n**Key features:**\n- **Q-Former architecture**: Lightweight query transformer bridges vision and language\n- **Frozen backbone efficiency**: No need to fine-tune large vision\u002Flanguage models\n- **Multiple LLM backends**: OPT (2.7B, 6.7B) and FlanT5 (XL, XXL)\n- **Zero-shot capabilities**: Strong performance without task-specific training\n- **Efficient training**: Only trains Q-Former (~188M parameters)\n- **State-of-the-art results**: Beats larger models on VQA benchmarks\n\n**Use alternatives instead:**\n- **LLaVA**: For instruction-following multimodal chat\n- **InstructBLIP**: For improved instruction-following (BLIP-2 successor)\n- **GPT-4V\u002FClaude 3**: For production multimodal chat (proprietary)\n- **CLIP**: For simple image-text similarity without generation\n- **Flamingo**: For few-shot visual learning\n\n## Quick start\n\n### Installation\n\n```bash\n# HuggingFace Transformers (recommended)\npip install transformers accelerate torch Pillow\n\n# Or LAVIS library (Salesforce official)\npip install salesforce-lavis\n```\n\n### Basic image captioning\n\n```python\nimport torch\nfrom PIL import Image\nfrom transformers import Blip2Processor, Blip2ForConditionalGeneration\n\n# Load model and processor\nprocessor = Blip2Processor.from_pretrained(\"Salesforce\u002Fblip2-opt-2.7b\")\nmodel = Blip2ForConditionalGeneration.from_pretrained(\n    \"Salesforce\u002Fblip2-opt-2.7b\",\n    torch_dtype=torch.float16,\n    device_map=\"auto\"\n)\n\n# Load image\nimage = Image.open(\"photo.jpg\").convert(\"RGB\")\n\n# Generate caption\ninputs = processor(images=image, return_tensors=\"pt\").to(\"cuda\", torch.float16)\ngenerated_ids = model.generate(**inputs, max_new_tokens=50)\ncaption = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]\nprint(caption)\n```\n\n### Visual question answering\n\n```python\n# Ask a question about the image\nquestion = \"What color is the car in this image?\"\n\ninputs = processor(images=image, text=question, return_tensors=\"pt\").to(\"cuda\", torch.float16)\ngenerated_ids = model.generate(**inputs, max_new_tokens=50)\nanswer = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]\nprint(answer)\n```\n\n### Using LAVIS library\n\n```python\nimport torch\nfrom lavis.models import load_model_and_preprocess\nfrom PIL import Image\n\n# Load model\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\nmodel, vis_processors, txt_processors = load_model_and_preprocess(\n    name=\"blip2_opt\",\n    model_type=\"pretrain_opt2.7b\",\n    is_eval=True,\n    device=device\n)\n\n# Process image\nimage = Image.open(\"photo.jpg\").convert(\"RGB\")\nimage = vis_processors[\"eval\"](image).unsqueeze(0).to(device)\n\n# Caption\ncaption = model.generate({\"image\": image})\nprint(caption)\n\n# VQA\nquestion = txt_processors[\"eval\"](\"What is in this image?\")\nanswer = model.generate({\"image\": image, \"prompt\": question})\nprint(answer)\n```\n\n## Core concepts\n\n### Architecture overview\n\n```\nBLIP-2 Architecture:\n┌─────────────────────────────────────────────────────────────┐\n│                        Q-Former                              │\n│  ┌─────────────────────────────────────────────────────┐    │\n│  │     Learned Queries (32 queries × 768 dim)          │    │\n│  └────────────────────────┬────────────────────────────┘    │\n│                           │                                  │\n│  ┌────────────────────────▼────────────────────────────┐    │\n│  │    Cross-Attention with Image Features               │    │\n│  └────────────────────────┬────────────────────────────┘    │\n│                           │                                  │\n│  ┌────────────────────────▼────────────────────────────┐    │\n│  │    Self-Attention Layers (Transformer)               │    │\n│  └────────────────────────┬────────────────────────────┘    │\n└───────────────────────────┼─────────────────────────────────┘\n                            │\n┌───────────────────────────▼─────────────────────────────────┐\n│  Frozen Vision Encoder    │      Frozen LLM                  │\n│  (ViT-G\u002F14 from EVA-CLIP) │      (OPT or FlanT5)            │\n└─────────────────────────────────────────────────────────────┘\n```\n\n### Model variants\n\n| Model | LLM Backend | Size | Use Case |\n|-------|-------------|------|----------|\n| `blip2-opt-2.7b` | OPT-2.7B | ~4GB | General captioning, VQA |\n| `blip2-opt-6.7b` | OPT-6.7B | ~8GB | Better reasoning |\n| `blip2-flan-t5-xl` | FlanT5-XL | ~5GB | Instruction following |\n| `blip2-flan-t5-xxl` | FlanT5-XXL | ~13GB | Best quality |\n\n### Q-Former components\n\n| Component | Description | Parameters |\n|-----------|-------------|------------|\n| Learned queries | Fixed set of learnable embeddings | 32 × 768 |\n| Image transformer | Cross-attention to vision features | ~108M |\n| Text transformer | Self-attention for text | ~108M |\n| Linear projection | Maps to LLM dimension | Varies |\n\n## Advanced usage\n\n### Batch processing\n\n```python\nfrom PIL import Image\nimport torch\n\n# Load multiple images\nimages = [Image.open(f\"image_{i}.jpg\").convert(\"RGB\") for i in range(4)]\nquestions = [\n    \"What is shown in this image?\",\n    \"Describe the scene.\",\n    \"What colors are prominent?\",\n    \"Is there a person in this image?\"\n]\n\n# Process batch\ninputs = processor(\n    images=images,\n    text=questions,\n    return_tensors=\"pt\",\n    padding=True\n).to(\"cuda\", torch.float16)\n\n# Generate\ngenerated_ids = model.generate(**inputs, max_new_tokens=50)\nanswers = processor.batch_decode(generated_ids, skip_special_tokens=True)\n\nfor q, a in zip(questions, answers):\n    print(f\"Q: {q}\\nA: {a}\\n\")\n```\n\n### Controlling generation\n\n```python\n# Control generation parameters\ngenerated_ids = model.generate(\n    **inputs,\n    max_new_tokens=100,\n    min_length=20,\n    num_beams=5,              # Beam search\n    no_repeat_ngram_size=2,   # Avoid repetition\n    top_p=0.9,                # Nucleus sampling\n    temperature=0.7,          # Creativity\n    do_sample=True,           # Enable sampling\n)\n\n# For deterministic output\ngenerated_ids = model.generate(\n    **inputs,\n    max_new_tokens=50,\n    num_beams=5,\n    do_sample=False,\n)\n```\n\n### Memory optimization\n\n```python\n# 8-bit quantization\nfrom transformers import BitsAndBytesConfig\n\nquantization_config = BitsAndBytesConfig(load_in_8bit=True)\n\nmodel = Blip2ForConditionalGeneration.from_pretrained(\n    \"Salesforce\u002Fblip2-opt-6.7b\",\n    quantization_config=quantization_config,\n    device_map=\"auto\"\n)\n\n# 4-bit quantization (more aggressive)\nquantization_config = BitsAndBytesConfig(\n    load_in_4bit=True,\n    bnb_4bit_compute_dtype=torch.float16\n)\n\nmodel = Blip2ForConditionalGeneration.from_pretrained(\n    \"Salesforce\u002Fblip2-flan-t5-xxl\",\n    quantization_config=quantization_config,\n    device_map=\"auto\"\n)\n```\n\n### Image-text matching\n\n```python\n# Using LAVIS for ITM (Image-Text Matching)\nfrom lavis.models import load_model_and_preprocess\n\nmodel, vis_processors, txt_processors = load_model_and_preprocess(\n    name=\"blip2_image_text_matching\",\n    model_type=\"pretrain\",\n    is_eval=True,\n    device=device\n)\n\nimage = vis_processors[\"eval\"](raw_image).unsqueeze(0).to(device)\ntext = txt_processors[\"eval\"](\"a dog sitting on grass\")\n\n# Get matching score\nitm_output = model({\"image\": image, \"text_input\": text}, match_head=\"itm\")\nitm_scores = torch.nn.functional.softmax(itm_output, dim=1)\nprint(f\"Match probability: {itm_scores[:, 1].item():.3f}\")\n```\n\n### Feature extraction\n\n```python\n# Extract image features with Q-Former\nfrom lavis.models import load_model_and_preprocess\n\nmodel, vis_processors, _ = load_model_and_preprocess(\n    name=\"blip2_feature_extractor\",\n    model_type=\"pretrain\",\n    is_eval=True,\n    device=device\n)\n\nimage = vis_processors[\"eval\"](raw_image).unsqueeze(0).to(device)\n\n# Get features\nfeatures = model.extract_features({\"image\": image}, mode=\"image\")\nimage_embeds = features.image_embeds  # Shape: [1, 32, 768]\nimage_features = features.image_embeds_proj  # Projected for matching\n```\n\n## Common workflows\n\n### Workflow 1: Image captioning pipeline\n\n```python\nimport torch\nfrom PIL import Image\nfrom transformers import Blip2Processor, Blip2ForConditionalGeneration\nfrom pathlib import Path\n\nclass ImageCaptioner:\n    def __init__(self, model_name=\"Salesforce\u002Fblip2-opt-2.7b\"):\n        self.processor = Blip2Processor.from_pretrained(model_name)\n        self.model = Blip2ForConditionalGeneration.from_pretrained(\n            model_name,\n            torch_dtype=torch.float16,\n            device_map=\"auto\"\n        )\n\n    def caption(self, image_path: str, prompt: str = None) -> str:\n        image = Image.open(image_path).convert(\"RGB\")\n\n        if prompt:\n            inputs = self.processor(images=image, text=prompt, return_tensors=\"pt\")\n        else:\n            inputs = self.processor(images=image, return_tensors=\"pt\")\n\n        inputs = inputs.to(\"cuda\", torch.float16)\n\n        generated_ids = self.model.generate(\n            **inputs,\n            max_new_tokens=50,\n            num_beams=5\n        )\n\n        return self.processor.decode(generated_ids[0], skip_special_tokens=True)\n\n    def caption_batch(self, image_paths: list, prompt: str = None) -> list:\n        images = [Image.open(p).convert(\"RGB\") for p in image_paths]\n\n        if prompt:\n            inputs = self.processor(\n                images=images,\n                text=[prompt] * len(images),\n                return_tensors=\"pt\",\n                padding=True\n            )\n        else:\n            inputs = self.processor(images=images, return_tensors=\"pt\", padding=True)\n\n        inputs = inputs.to(\"cuda\", torch.float16)\n\n        generated_ids = self.model.generate(**inputs, max_new_tokens=50)\n        return self.processor.batch_decode(generated_ids, skip_special_tokens=True)\n\n# Usage\ncaptioner = ImageCaptioner()\n\n# Single image\ncaption = captioner.caption(\"photo.jpg\")\nprint(f\"Caption: {caption}\")\n\n# With prompt for style\ncaption = captioner.caption(\"photo.jpg\", \"a detailed description of\")\nprint(f\"Detailed: {caption}\")\n\n# Batch processing\ncaptions = captioner.caption_batch([\"img1.jpg\", \"img2.jpg\", \"img3.jpg\"])\nfor i, cap in enumerate(captions):\n    print(f\"Image {i+1}: {cap}\")\n```\n\n### Workflow 2: Visual Q&A system\n\n```python\nclass VisualQA:\n    def __init__(self, model_name=\"Salesforce\u002Fblip2-flan-t5-xl\"):\n        self.processor = Blip2Processor.from_pretrained(model_name)\n        self.model = Blip2ForConditionalGeneration.from_pretrained(\n            model_name,\n            torch_dtype=torch.float16,\n            device_map=\"auto\"\n        )\n        self.current_image = None\n        self.current_inputs = None\n\n    def set_image(self, image_path: str):\n        \"\"\"Load image for multiple questions.\"\"\"\n        self.current_image = Image.open(image_path).convert(\"RGB\")\n\n    def ask(self, question: str) -> str:\n        \"\"\"Ask a question about the current image.\"\"\"\n        if self.current_image is None:\n            raise ValueError(\"No image set. Call set_image() first.\")\n\n        # Format question for FlanT5\n        prompt = f\"Question: {question} Answer:\"\n\n        inputs = self.processor(\n            images=self.current_image,\n            text=prompt,\n            return_tensors=\"pt\"\n        ).to(\"cuda\", torch.float16)\n\n        generated_ids = self.model.generate(\n            **inputs,\n            max_new_tokens=50,\n            num_beams=5\n        )\n\n        return self.processor.decode(generated_ids[0], skip_special_tokens=True)\n\n    def ask_multiple(self, questions: list) -> dict:\n        \"\"\"Ask multiple questions about current image.\"\"\"\n        return {q: self.ask(q) for q in questions}\n\n# Usage\nvqa = VisualQA()\nvqa.set_image(\"scene.jpg\")\n\n# Ask questions\nprint(vqa.ask(\"What objects are in this image?\"))\nprint(vqa.ask(\"What is the weather like?\"))\nprint(vqa.ask(\"How many people are there?\"))\n\n# Batch questions\nresults = vqa.ask_multiple([\n    \"What is the main subject?\",\n    \"What colors are dominant?\",\n    \"Is this indoors or outdoors?\"\n])\n```\n\n### Workflow 3: Image search\u002Fretrieval\n\n```python\nimport torch\nimport numpy as np\nfrom PIL import Image\nfrom lavis.models import load_model_and_preprocess\n\nclass ImageSearchEngine:\n    def __init__(self):\n        self.device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n        self.model, self.vis_processors, self.txt_processors = load_model_and_preprocess(\n            name=\"blip2_feature_extractor\",\n            model_type=\"pretrain\",\n            is_eval=True,\n            device=self.device\n        )\n        self.image_features = []\n        self.image_paths = []\n\n    def index_images(self, image_paths: list):\n        \"\"\"Build index from images.\"\"\"\n        self.image_paths = image_paths\n\n        for path in image_paths:\n            image = Image.open(path).convert(\"RGB\")\n            image = self.vis_processors[\"eval\"](image).unsqueeze(0).to(self.device)\n\n            with torch.no_grad():\n                features = self.model.extract_features({\"image\": image}, mode=\"image\")\n                # Use projected features for matching\n                self.image_features.append(\n                    features.image_embeds_proj.mean(dim=1).cpu().numpy()\n                )\n\n        self.image_features = np.vstack(self.image_features)\n\n    def search(self, query: str, top_k: int = 5) -> list:\n        \"\"\"Search images by text query.\"\"\"\n        # Get text features\n        text = self.txt_processors[\"eval\"](query)\n        text_input = {\"text_input\": [text]}\n\n        with torch.no_grad():\n            text_features = self.model.extract_features(text_input, mode=\"text\")\n            text_embeds = text_features.text_embeds_proj[:, 0].cpu().numpy()\n\n        # Compute similarities\n        similarities = np.dot(self.image_features, text_embeds.T).squeeze()\n        top_indices = np.argsort(similarities)[::-1][:top_k]\n\n        return [(self.image_paths[i], similarities[i]) for i in top_indices]\n\n# Usage\nengine = ImageSearchEngine()\nengine.index_images([\"img1.jpg\", \"img2.jpg\", \"img3.jpg\", ...])\n\n# Search\nresults = engine.search(\"a sunset over the ocean\", top_k=5)\nfor path, score in results:\n    print(f\"{path}: {score:.3f}\")\n```\n\n## Output format\n\n### Generation output\n\n```python\n# Direct generation returns token IDs\ngenerated_ids = model.generate(**inputs, max_new_tokens=50)\n# Shape: [batch_size, sequence_length]\n\n# Decode to text\ntext = processor.batch_decode(generated_ids, skip_special_tokens=True)\n# Returns: list of strings\n```\n\n### Feature extraction output\n\n```python\n# Q-Former outputs\nfeatures = model.extract_features({\"image\": image}, mode=\"image\")\n\nfeatures.image_embeds          # [B, 32, 768] - Q-Former outputs\nfeatures.image_embeds_proj     # [B, 32, 256] - Projected for matching\nfeatures.text_embeds          # [B, seq_len, 768] - Text features\nfeatures.text_embeds_proj     # [B, 256] - Projected text (CLS)\n```\n\n## Performance optimization\n\n### GPU memory requirements\n\n| Model | FP16 VRAM | INT8 VRAM | INT4 VRAM |\n|-------|-----------|-----------|-----------|\n| blip2-opt-2.7b | ~8GB | ~5GB | ~3GB |\n| blip2-opt-6.7b | ~16GB | ~9GB | ~5GB |\n| blip2-flan-t5-xl | ~10GB | ~6GB | ~4GB |\n| blip2-flan-t5-xxl | ~26GB | ~14GB | ~8GB |\n\n### Speed optimization\n\n```python\n# Use Flash Attention if available\nmodel = Blip2ForConditionalGeneration.from_pretrained(\n    \"Salesforce\u002Fblip2-opt-2.7b\",\n    torch_dtype=torch.float16,\n    attn_implementation=\"flash_attention_2\",  # Requires flash-attn\n    device_map=\"auto\"\n)\n\n# Compile model (PyTorch 2.0+)\nmodel = torch.compile(model)\n\n# Use smaller images (if quality allows)\nprocessor = Blip2Processor.from_pretrained(\"Salesforce\u002Fblip2-opt-2.7b\")\n# Default is 224x224, which is optimal\n```\n\n## Common issues\n\n| Issue | Solution |\n|-------|----------|\n| CUDA OOM | Use INT8\u002FINT4 quantization, smaller model |\n| Slow generation | Use greedy decoding, reduce max_new_tokens |\n| Poor captions | Try FlanT5 variant, use prompts |\n| Hallucinations | Lower temperature, use beam search |\n| Wrong answers | Rephrase question, provide context |\n\n## References\n\n- **[Advanced Usage](references\u002Fadvanced-usage.md)** - Fine-tuning, integration, deployment\n- **[Troubleshooting](references\u002Ftroubleshooting.md)** - Common issues and solutions\n\n## Resources\n\n- **Paper**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2301.12597\n- **GitHub (LAVIS)**: https:\u002F\u002Fgithub.com\u002Fsalesforce\u002FLAVIS\n- **HuggingFace**: https:\u002F\u002Fhuggingface.co\u002FSalesforce\u002Fblip2-opt-2.7b\n- **Demo**: https:\u002F\u002Fhuggingface.co\u002Fspaces\u002FSalesforce\u002FBLIP2\n- **InstructBLIP**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.06500 (successor)\n","# BLIP-2: Vision-Language Pre-training\n\nComprehensive guide to using Salesforce's BLIP-2 for vision-language tasks with frozen image encoders and large language models.\n\n## When to use BLIP-2\n\n**Use BLIP-2 when:**\n- Need high-quality image captioning with natural descriptions\n- Building visual question answering (VQA) systems\n- Require zero-shot image-text understanding without task-specific training\n- Want to leverage LLM reasoning for visual tasks\n- Building multimodal conversational AI\n- Need image-text retrieval or matching\n\n**Key features:**\n- **Q-Former architecture**: Lightweight query transformer bridges vision and language\n- **Frozen backbone efficiency**: No need to fine-tune large vision\u002Flanguage models\n- **Multiple LLM backends**: OPT (2.7B, 6.7B) and FlanT5 (XL, XXL)\n- **Zero-shot capabilities**: Strong performance without task-specific training\n- **Efficient training**: Only trains Q-Former (~188M parameters)\n- **State-of-the-art results**: Beats larger models on VQA benchmarks\n\n**Use alternatives instead:**\n- **LLaVA**: For instruction-following multimodal chat\n- **InstructBLIP**: For improved instruction-following (BLIP-2 successor)\n- **GPT-4V\u002FClaude 3**: For production multimodal chat (proprietary)\n- **CLIP**: For simple image-text similarity without generation\n- **Flamingo**: For few-shot visual learning\n\n## Quick start\n\n### Installation\n\n```bash\n# HuggingFace Transformers (recommended)\npip install transformers accelerate torch Pillow\n\n# Or LAVIS library (Salesforce official)\npip install salesforce-lavis\n```\n\n### Basic image captioning\n\n```python\nimport torch\nfrom PIL import Image\nfrom transformers import Blip2Processor, Blip2ForConditionalGeneration\n\n# Load model and processor\nprocessor = Blip2Processor.from_pretrained(\"Salesforce\u002Fblip2-opt-2.7b\")\nmodel = Blip2ForConditionalGeneration.from_pretrained(\n    \"Salesforce\u002Fblip2-opt-2.7b\",\n    torch_dtype=torch.float16,\n    device_map=\"auto\"\n)\n\n# Load image\nimage = Image.open(\"photo.jpg\").convert(\"RGB\")\n\n# Generate caption\ninputs = processor(images=image, return_tensors=\"pt\").to(\"cuda\", torch.float16)\ngenerated_ids = model.generate(**inputs, max_new_tokens=50)\ncaption = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]\nprint(caption)\n```\n\n### Visual question answering\n\n```python\n# Ask a question about the image\nquestion = \"What color is the car in this image?\"\n\ninputs = processor(images=image, text=question, return_tensors=\"pt\").to(\"cuda\", torch.float16)\ngenerated_ids = model.generate(**inputs, max_new_tokens=50)\nanswer = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]\nprint(answer)\n```\n\n### Using LAVIS library\n\n```python\nimport torch\nfrom lavis.models import load_model_and_preprocess\nfrom PIL import Image\n\n# Load model\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\nmodel, vis_processors, txt_processors = load_model_and_preprocess(\n    name=\"blip2_opt\",\n    model_type=\"pretrain_opt2.7b\",\n    is_eval=True,\n    device=device\n)\n\n# Process image\nimage = Image.open(\"photo.jpg\").convert(\"RGB\")\nimage = vis_processors[\"eval\"](image).unsqueeze(0).to(device)\n\n# Caption\ncaption = model.generate({\"image\": image})\nprint(caption)\n\n# VQA\nquestion = txt_processors[\"eval\"](\"What is in this image?\")\nanswer = model.generate({\"image\": image, \"prompt\": question})\nprint(answer)\n```\n\n## Core concepts\n\n### Architecture overview\n\n```\nBLIP-2 Architecture:\n┌─────────────────────────────────────────────────────────────┐\n│                        Q-Former                              │\n│  ┌─────────────────────────────────────────────────────┐    │\n│  │     Learned Queries (32 queries × 768 dim)          │    │\n│  └────────────────────────┬────────────────────────────┘    │\n│                           │                                  │\n│  ┌────────────────────────▼────────────────────────────┐    │\n│  │    Cross-Attention with Image Features               │    │\n│  └────────────────────────┬────────────────────────────┘    │\n│                           │                                  │\n│  ┌────────────────────────▼────────────────────────────┐    │\n│  │    Self-Attention Layers (Transformer)               │    │\n│  └────────────────────────┬────────────────────────────┘    │\n└───────────────────────────┼─────────────────────────────────┘\n                            │\n┌───────────────────────────▼─────────────────────────────────┐\n│  Frozen Vision Encoder    │      Frozen LLM                  │\n│  (ViT-G\u002F14 from EVA-CLIP) │      (OPT or FlanT5)            │\n└─────────────────────────────────────────────────────────────┘\n```\n\n### Model variants\n\n| Model | LLM Backend | Size | Use Case |\n|-------|-------------|------|----------|\n| `blip2-opt-2.7b` | OPT-2.7B | ~4GB | General captioning, VQA |\n| `blip2-opt-6.7b` | OPT-6.7B | ~8GB | Better reasoning |\n| `blip2-flan-t5-xl` | FlanT5-XL | ~5GB | Instruction following |\n| `blip2-flan-t5-xxl` | FlanT5-XXL | ~13GB | Best quality |\n\n### Q-Former components\n\n| Component | Description | Parameters |\n|-----------|-------------|------------|\n| Learned queries | Fixed set of learnable embeddings | 32 × 768 |\n| Image transformer | Cross-attention to vision features | ~108M |\n| Text transformer | Self-attention for text | ~108M |\n| Linear projection | Maps to LLM dimension | Varies |\n\n## Advanced usage\n\n### Batch processing\n\n```python\nfrom PIL import Image\nimport torch\n\n# Load multiple images\nimages = [Image.open(f\"image_{i}.jpg\").convert(\"RGB\") for i in range(4)]\nquestions = [\n    \"What is shown in this image?\",\n    \"Describe the scene.\",\n    \"What colors are prominent?\",\n    \"Is there a person in this image?\"\n]\n\n# Process batch\ninputs = processor(\n    images=images,\n    text=questions,\n    return_tensors=\"pt\",\n    padding=True\n).to(\"cuda\", torch.float16)\n\n# Generate\ngenerated_ids = model.generate(**inputs, max_new_tokens=50)\nanswers = processor.batch_decode(generated_ids, skip_special_tokens=True)\n\nfor q, a in zip(questions, answers):\n    print(f\"Q: {q}\\nA: {a}\\n\")\n```\n\n### Controlling generation\n\n```python\n# Control generation parameters\ngenerated_ids = model.generate(\n    **inputs,\n    max_new_tokens=100,\n    min_length=20,\n    num_beams=5,              # Beam search\n    no_repeat_ngram_size=2,   # Avoid repetition\n    top_p=0.9,                # Nucleus sampling\n    temperature=0.7,          # Creativity\n    do_sample=True,           # Enable sampling\n)\n\n# For deterministic output\ngenerated_ids = model.generate(\n    **inputs,\n    max_new_tokens=50,\n    num_beams=5,\n    do_sample=False,\n)\n```\n\n### Memory optimization\n\n```python\n# 8-bit quantization\nfrom transformers import BitsAndBytesConfig\n\nquantization_config = BitsAndBytesConfig(load_in_8bit=True)\n\nmodel = Blip2ForConditionalGeneration.from_pretrained(\n    \"Salesforce\u002Fblip2-opt-6.7b\",\n    quantization_config=quantization_config,\n    device_map=\"auto\"\n)\n\n# 4-bit quantization (more aggressive)\nquantization_config = BitsAndBytesConfig(\n    load_in_4bit=True,\n    bnb_4bit_compute_dtype=torch.float16\n)\n\nmodel = Blip2ForConditionalGeneration.from_pretrained(\n    \"Salesforce\u002Fblip2-flan-t5-xxl\",\n    quantization_config=quantization_config,\n    device_map=\"auto\"\n)\n```\n\n### Image-text matching\n\n```python\n# Using LAVIS for ITM (Image-Text Matching)\nfrom lavis.models import load_model_and_preprocess\n\nmodel, vis_processors, txt_processors = load_model_and_preprocess(\n    name=\"blip2_image_text_matching\",\n    model_type=\"pretrain\",\n    is_eval=True,\n    device=device\n)\n\nimage = vis_processors[\"eval\"](raw_image).unsqueeze(0).to(device)\ntext = txt_processors[\"eval\"](\"a dog sitting on grass\")\n\n# Get matching score\nitm_output = model({\"image\": image, \"text_input\": text}, match_head=\"itm\")\nitm_scores = torch.nn.functional.softmax(itm_output, dim=1)\nprint(f\"Match probability: {itm_scores[:, 1].item():.3f}\")\n```\n\n### Feature extraction\n\n```python\n# Extract image features with Q-Former\nfrom lavis.models import load_model_and_preprocess\n\nmodel, vis_processors, _ = load_model_and_preprocess(\n    name=\"blip2_feature_extractor\",\n    model_type=\"pretrain\",\n    is_eval=True,\n    device=device\n)\n\nimage = vis_processors[\"eval\"](raw_image).unsqueeze(0).to(device)\n\n# Get features\nfeatures = model.extract_features({\"image\": image}, mode=\"image\")\nimage_embeds = features.image_embeds  # Shape: [1, 32, 768]\nimage_features = features.image_embeds_proj  # Projected for matching\n```\n\n## Common workflows\n\n### Workflow 1: Image captioning pipeline\n\n```python\nimport torch\nfrom PIL import Image\nfrom transformers import Blip2Processor, Blip2ForConditionalGeneration\nfrom pathlib import Path\n\nclass ImageCaptioner:\n    def __init__(self, model_name=\"Salesforce\u002Fblip2-opt-2.7b\"):\n        self.processor = Blip2Processor.from_pretrained(model_name)\n        self.model = Blip2ForConditionalGeneration.from_pretrained(\n            model_name,\n            torch_dtype=torch.float16,\n            device_map=\"auto\"\n        )\n\n    def caption(self, image_path: str, prompt: str = None) -> str:\n        image = Image.open(image_path).convert(\"RGB\")\n\n        if prompt:\n            inputs = self.processor(images=image, text=prompt, return_tensors=\"pt\")\n        else:\n            inputs = self.processor(images=image, return_tensors=\"pt\")\n\n        inputs = inputs.to(\"cuda\", torch.float16)\n\n        generated_ids = self.model.generate(\n            **inputs,\n            max_new_tokens=50,\n            num_beams=5\n        )\n\n        return self.processor.decode(generated_ids[0], skip_special_tokens=True)\n\n    def caption_batch(self, image_paths: list, prompt: str = None) -> list:\n        images = [Image.open(p).convert(\"RGB\") for p in image_paths]\n\n        if prompt:\n            inputs = self.processor(\n                images=images,\n                text=[prompt] * len(images),\n                return_tensors=\"pt\",\n                padding=True\n            )\n        else:\n            inputs = self.processor(images=images, return_tensors=\"pt\", padding=True)\n\n        inputs = inputs.to(\"cuda\", torch.float16)\n\n        generated_ids = self.model.generate(**inputs, max_new_tokens=50)\n        return self.processor.batch_decode(generated_ids, skip_special_tokens=True)\n\n# Usage\ncaptioner = ImageCaptioner()\n\n# Single image\ncaption = captioner.caption(\"photo.jpg\")\nprint(f\"Caption: {caption}\")\n\n# With prompt for style\ncaption = captioner.caption(\"photo.jpg\", \"a detailed description of\")\nprint(f\"Detailed: {caption}\")\n\n# Batch processing\ncaptions = captioner.caption_batch([\"img1.jpg\", \"img2.jpg\", \"img3.jpg\"])\nfor i, cap in enumerate(captions):\n    print(f\"Image {i+1}: {cap}\")\n```\n\n### Workflow 2: Visual Q&A system\n\n```python\nclass VisualQA:\n    def __init__(self, model_name=\"Salesforce\u002Fblip2-flan-t5-xl\"):\n        self.processor = Blip2Processor.from_pretrained(model_name)\n        self.model = Blip2ForConditionalGeneration.from_pretrained(\n            model_name,\n            torch_dtype=torch.float16,\n            device_map=\"auto\"\n        )\n        self.current_image = None\n        self.current_inputs = None\n\n    def set_image(self, image_path: str):\n        \"\"\"Load image for multiple questions.\"\"\"\n        self.current_image = Image.open(image_path).convert(\"RGB\")\n\n    def ask(self, question: str) -> str:\n        \"\"\"Ask a question about the current image.\"\"\"\n        if self.current_image is None:\n            raise ValueError(\"No image set. Call set_image() first.\")\n\n        # Format question for FlanT5\n        prompt = f\"Question: {question} Answer:\"\n\n        inputs = self.processor(\n            images=self.current_image,\n            text=prompt,\n            return_tensors=\"pt\"\n        ).to(\"cuda\", torch.float16)\n\n        generated_ids = self.model.generate(\n            **inputs,\n            max_new_tokens=50,\n            num_beams=5\n        )\n\n        return self.processor.decode(generated_ids[0], skip_special_tokens=True)\n\n    def ask_multiple(self, questions: list) -> dict:\n        \"\"\"Ask multiple questions about current image.\"\"\"\n        return {q: self.ask(q) for q in questions}\n\n# Usage\nvqa = VisualQA()\nvqa.set_image(\"scene.jpg\")\n\n# Ask questions\nprint(vqa.ask(\"What objects are in this image?\"))\nprint(vqa.ask(\"What is the weather like?\"))\nprint(vqa.ask(\"How many people are there?\"))\n\n# Batch questions\nresults = vqa.ask_multiple([\n    \"What is the main subject?\",\n    \"What colors are dominant?\",\n    \"Is this indoors or outdoors?\"\n])\n```\n\n### Workflow 3: Image search\u002Fretrieval\n\n```python\nimport torch\nimport numpy as np\nfrom PIL import Image\nfrom lavis.models import load_model_and_preprocess\n\nclass ImageSearchEngine:\n    def __init__(self):\n        self.device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n        self.model, self.vis_processors, self.txt_processors = load_model_and_preprocess(\n            name=\"blip2_feature_extractor\",\n            model_type=\"pretrain\",\n            is_eval=True,\n            device=self.device\n        )\n        self.image_features = []\n        self.image_paths = []\n\n    def index_images(self, image_paths: list):\n        \"\"\"Build index from images.\"\"\"\n        self.image_paths = image_paths\n\n        for path in image_paths:\n            image = Image.open(path).convert(\"RGB\")\n            image = self.vis_processors[\"eval\"](image).unsqueeze(0).to(self.device)\n\n            with torch.no_grad():\n                features = self.model.extract_features({\"image\": image}, mode=\"image\")\n                # Use projected features for matching\n                self.image_features.append(\n                    features.image_embeds_proj.mean(dim=1).cpu().numpy()\n                )\n\n        self.image_features = np.vstack(self.image_features)\n\n    def search(self, query: str, top_k: int = 5) -> list:\n        \"\"\"Search images by text query.\"\"\"\n        # Get text features\n        text = self.txt_processors[\"eval\"](query)\n        text_input = {\"text_input\": [text]}\n\n        with torch.no_grad():\n            text_features = self.model.extract_features(text_input, mode=\"text\")\n            text_embeds = text_features.text_embeds_proj[:, 0].cpu().numpy()\n\n        # Compute similarities\n        similarities = np.dot(self.image_features, text_embeds.T).squeeze()\n        top_indices = np.argsort(similarities)[::-1][:top_k]\n\n        return [(self.image_paths[i], similarities[i]) for i in top_indices]\n\n# Usage\nengine = ImageSearchEngine()\nengine.index_images([\"img1.jpg\", \"img2.jpg\", \"img3.jpg\", ...])\n\n# Search\nresults = engine.search(\"a sunset over the ocean\", top_k=5)\nfor path, score in results:\n    print(f\"{path}: {score:.3f}\")\n```\n\n## Output format\n\n### Generation output\n\n```python\n# Direct generation returns token IDs\ngenerated_ids = model.generate(**inputs, max_new_tokens=50)\n# Shape: [batch_size, sequence_length]\n\n# Decode to text\ntext = processor.batch_decode(generated_ids, skip_special_tokens=True)\n# Returns: list of strings\n```\n\n### Feature extraction output\n\n```python\n# Q-Former outputs\nfeatures = model.extract_features({\"image\": image}, mode=\"image\")\n\nfeatures.image_embeds          # [B, 32, 768] - Q-Former outputs\nfeatures.image_embeds_proj     # [B, 32, 256] - Projected for matching\nfeatures.text_embeds          # [B, seq_len, 768] - Text features\nfeatures.text_embeds_proj     # [B, 256] - Projected text (CLS)\n```\n\n## Performance optimization\n\n### GPU memory requirements\n\n| Model | FP16 VRAM | INT8 VRAM | INT4 VRAM |\n|-------|-----------|-----------|-----------|\n| blip2-opt-2.7b | ~8GB | ~5GB | ~3GB |\n| blip2-opt-6.7b | ~16GB | ~9GB | ~5GB |\n| blip2-flan-t5-xl | ~10GB | ~6GB | ~4GB |\n| blip2-flan-t5-xxl | ~26GB | ~14GB | ~8GB |\n\n### Speed optimization\n\n```python\n# Use Flash Attention if available\nmodel = Blip2ForConditionalGeneration.from_pretrained(\n    \"Salesforce\u002Fblip2-opt-2.7b\",\n    torch_dtype=torch.float16,\n    attn_implementation=\"flash_attention_2\",  # Requires flash-attn\n    device_map=\"auto\"\n)\n\n# Compile model (PyTorch 2.0+)\nmodel = torch.compile(model)\n\n# Use smaller images (if quality allows)\nprocessor = Blip2Processor.from_pretrained(\"Salesforce\u002Fblip2-opt-2.7b\")\n# Default is 224x224, which is optimal\n```\n\n## Common issues\n\n| Issue | Solution |\n|-------|----------|\n| CUDA OOM | Use INT8\u002FINT4 quantization, smaller model |\n| Slow generation | Use greedy decoding, reduce max_new_tokens |\n| Poor captions | Try FlanT5 variant, use prompts |\n| Hallucinations | Lower temperature, use beam search |\n| Wrong answers | Rephrase question, provide context |\n\n## References\n\n- **[Advanced Usage](references\u002Fadvanced-usage.md)** - Fine-tuning, integration, deployment\n- **[Troubleshooting](references\u002Ftroubleshooting.md)** - Common issues and solutions\n\n## Resources\n\n- **Paper**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2301.12597\n- **GitHub (LAVIS)**: https:\u002F\u002Fgithub.com\u002Fsalesforce\u002FLAVIS\n- **HuggingFace**: https:\u002F\u002Fhuggingface.co\u002FSalesforce\u002Fblip2-opt-2.7b\n- **Demo**: https:\u002F\u002Fhuggingface.co\u002Fspaces\u002FSalesforce\u002FBLIP2\n- **InstructBLIP**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.06500 (successor)\n",{"name":5,"tags":24,"author":30,"license":15,"version":31,"description":7,"dependencies":32},[25,26,27,28,29],"Multimodal","Vision-Language","Image Captioning","VQA","Zero-Shot","Orchestra Research","1.0.0",[33,34,35],"transformers>=4.30.0","torch>=1.10.0","Pillow","main","f4a24bd1ab047411d4ae57ee9a96701a7706e779","claude-code-templates\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fai-research\u002Fmultimodal-blip-2","https:\u002F\u002Fgithub.com\u002Fcocyuhao\u002Fai-dev-tooling-research-archive","https:\u002F\u002Fgithub.com\u002Fcocyuhao","Beijing Wuzi University","2026-05-12 10:57:02",{"source":8,"skillName":5,"commands":44,"prompt":54,"zipAvailable":55},[45,48,51],{"tool":46,"command":47},"npx","npx skills add cocyuhao\u002Fai-dev-tooling-research-archive --skill blip-2-vision-language",{"tool":49,"command":50},"bunx","bunx skills add cocyuhao\u002Fai-dev-tooling-research-archive --skill blip-2-vision-language",{"tool":52,"command":53},"pnpm","pnpm dlx skills add cocyuhao\u002Fai-dev-tooling-research-archive --skill blip-2-vision-language","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fcocyuhao\u002Fai-dev-tooling-research-archive\u002Fblob\u002Fmain\u002Fclaude-code-templates\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fai-research\u002Fmultimodal-blip-2\u002FSKILL.md",true,[57,63],{"path":58,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":61,"sourceHtmlUrl":62},"references\u002Fadvanced-usage.md","REFERENCE","参考文档",18889,"https:\u002F\u002Fgithub.com\u002Fcocyuhao\u002Fai-dev-tooling-research-archive\u002Fblob\u002Fmain\u002Fclaude-code-templates\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fai-research\u002Fmultimodal-blip-2\u002Freferences\u002Fadvanced-usage.md",{"path":64,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":65,"sourceHtmlUrl":66},"references\u002Ftroubleshooting.md",11981,"https:\u002F\u002Fgithub.com\u002Fcocyuhao\u002Fai-dev-tooling-research-archive\u002Fblob\u002Fmain\u002Fclaude-code-templates\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fai-research\u002Fmultimodal-blip-2\u002Freferences\u002Ftroubleshooting.md",[68,76,84,91,98,105],{"skillNo":69,"name":70,"nameZh":6,"description":71,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":72,"categories":73,"keywords":74,"publishedAt":75},"01M72JQ593HE3SRJD3446CCKA8","web-access","所有联网操作必须通过此 skill 处理，包括：搜索、网页抓取、登录后操作、网络交互等。 触发场景：用户要求搜索信息、查看网页内容、访问需要登录的网站、操作网页界面、抓取社交媒体内容（小红书、微博、推特等）、读取动态渲染页面、以及任何需要真实浏览器环境的网络任务。","https:\u002F\u002Fgithub.com\u002Fcocyuhao\u002Fai-dev-tooling-research-archive\u002Fblob\u002Fmain\u002Fweb-access\u002FSKILL.md",[],[],"2026-11-10 22:39:42",{"skillNo":77,"name":78,"nameZh":6,"description":79,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":80,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":75},"01M72JKXJYFT2RBJ2HP5ZRA853","last30days","Multi-query social research across Reddit, X, YouTube, TikTok, Instagram, Hacker News, Polymarket, and the web. Intelligent query planning with Gemini\u002FOpenAI fallback.","https:\u002F\u002Fgithub.com\u002Fcocyuhao\u002Fai-dev-tooling-research-archive\u002Fblob\u002Fmain\u002Flast30days-skill\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fcocyuhao\u002Fai-dev-tooling-research-archive\u002Fblob\u002Fmain\u002Flast30days-skill\u002FSKILL.md",[],[],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":75},"01M72JKMRQSD0ZY66KQ863HVY5","writing-plans","Use when you have a spec or requirements for a multi-step task. Creates comprehensive implementation plans with bite-sized tasks, exact file paths, and complete code examples.","https:\u002F\u002Fgithub.com\u002Fcocyuhao\u002Fai-dev-tooling-research-archive\u002Fblob\u002Fmain\u002Fhermes-agent\u002Fskills\u002Fsoftware-development\u002Fwriting-plans\u002FSKILL.md",[],[],{"skillNo":92,"name":93,"nameZh":6,"description":94,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":95,"categories":96,"keywords":97,"publishedAt":75},"01M72JKHCV8DT6VWSD4KS5YWT9","test-driven-development","Use when implementing any feature or bugfix, before writing implementation code. Enforces RED-GREEN-REFACTOR cycle with test-first approach.","https:\u002F\u002Fgithub.com\u002Fcocyuhao\u002Fai-dev-tooling-research-archive\u002Fblob\u002Fmain\u002Fhermes-agent\u002Fskills\u002Fsoftware-development\u002Ftest-driven-development\u002FSKILL.md",[],[],{"skillNo":99,"name":100,"nameZh":6,"description":101,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":102,"categories":103,"keywords":104,"publishedAt":75},"01M72JKD7VX7EQM38R704F3KT1","systematic-debugging","Use when encountering any bug, test failure, or unexpected behavior. 4-phase root cause investigation — NO fixes without understanding the problem first.","https:\u002F\u002Fgithub.com\u002Fcocyuhao\u002Fai-dev-tooling-research-archive\u002Fblob\u002Fmain\u002Fhermes-agent\u002Fskills\u002Fsoftware-development\u002Fsystematic-debugging\u002FSKILL.md",[],[],{"skillNo":106,"name":107,"nameZh":6,"description":108,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":109,"categories":110,"keywords":111,"publishedAt":75},"01M72JKA9XXB92FS8JHPCYW8CM","subagent-driven-development","Use when executing implementation plans with independent tasks. Dispatches fresh delegate_task per task with two-stage review (spec compliance then code quality).","https:\u002F\u002Fgithub.com\u002Fcocyuhao\u002Fai-dev-tooling-research-archive\u002Fblob\u002Fmain\u002Fhermes-agent\u002Fskills\u002Fsoftware-development\u002Fsubagent-driven-development\u002FSKILL.md",[],[]]