img2prompt is a local Python tool that transforms images into detailed text descriptions (prompts) using the BLIP image captioning model. Its goal is to evaluate the impact of different decoding configurations to generate more diverse, detailed, and context-rich prompts—especially useful for downstream applications like image-to-image generation or semantic analysis.

This project explores image captioning via Vision-Language Transformers, specifically the BLIP (Bootstrapped Language Image Pretraining) architecture. BLIP employs a multimodal transformer model that has been pre-trained for tasks like image-text matching and conditional generation. We test caption generation across varying decoding strategies, evaluating:

• Minimum sequence length
• Beam search with penalties for repetition
• Score and token analysis per inference step

This allows us to quantify caption diversity, semantic variance, and token confidence evolution across different configurations.

• Python 3
• Transformers (Hugging Face)
• PyTorch with MPS backend (for macOS acceleration)
• Pillow (PIL)
• Standard libraries: json, platform, os, time

• Loads the Salesforce/blip-image-captioning-large model and processor.
• Uses torch.float16 on macOS/Apple Sillicon GPUs via MPS when available.
• Reads and processes a local image (test.jpg).

• Performs a first pass with default settings to generate a base caption for later comparison.

Three decoding configurations are tested:

[
  {"min_length": 10},
  {"min_length": 30},
  {"min_length": 50}
]

Each configuration applies:

• beam search with 7 beams
• length_penalty = 1.2
• repetition_penalty = 1.2
• no_repeat_ngram_size = 3

For each configuration, the system records:

• Final caption
• Generation score
• Difference vs base caption (in word set)
• Token-level scores per step (top 5 candidates per position)
• Full JSON reports saved in ./resultados/

Example output structure:

{
  "caption_base": "...",
  "results": [
    {
      "caption": "...",
      "score": 12.33,
      "delta_vs_base": 9,
      "tokens": [...],
      "inference_time": 3.5
    }
  ],
  "step_scores": [
    {
      "step": 1,
      "top_tokens": [["a", 2.31], ["the", 2.21], ...]
    }
  ]
}

• BLIP generates high-quality captions out-of-the-box, but customizing decoding strategies can unlock more nuanced or artistic outputs.
• This tool is suitable for developers aiming to extract detailed prompts from images to use in generative art or caption-based indexing.
• Local execution ensures full control over inference, performance profiling, and offline experimentation.