Hugging Face's logo

Dataseeds
/
BLIP2-opt-2.7b-DSD-FineTune

Image-to-Text
Transformers
English
vision-language
multimodal
blip2
opt
fine-tuned
photography
image-captioning
scene-analysis
Eval Results
Model card Files Community
1

BLIP2-OPT-2.7B Fine-tuned on DataSeeds.AI Dataset

Code: https://github.com/DataSeeds-ai/DSD-finetune-blip-llava

This model is a fine-tuned version of Salesforce/blip2-opt-2.7b specialized for photography scene analysis and technical description generation. The model was fine-tuned on the DataSeeds.AI Sample Dataset (DSD) to enhance its capabilities in generating detailed photographic descriptions with focus on composition, lighting, and technical aspects.

The model was presented in the paper Peer-Ranked Precision: Creating a Foundational Dataset for Fine-Tuning Vision Models from DataSeeds' Annotated Imagery.

Model Description

  • Base Model: BLIP2-OPT-2.7B
  • Vision Encoder: EVA-CLIP ViT-g/14
  • Language Model: OPT-2.7B (2.7 billion parameters)
  • Architecture: BLIP-2 with Q-Former bridging vision and language
  • Fine-tuning Approach: Full model fine-tuning (ViT unfrozen)
  • Task: Photography scene analysis and detailed image captioning
  • Precision: Mixed Precision (AMP enabled)
  • Image Resolution: 364Γ—364 pixels

Training Details

Dataset

The model was fine-tuned on the DataSeeds.AI Sample Dataset (DSD), containing 10,610 curated photography images with comprehensive annotations:

  • Training Set: 9,549 image-text pairs (90%)
  • Validation Set: 1,061 image-text pairs (10%)
  • Content Focus: Technical scene descriptions, compositional analysis, lighting conditions
  • Annotation Quality: Professional photography standards with 15+ word descriptions

Training Configuration

Parameter Value
Learning Rate 1e-5
Optimizer AdamW
LR Schedule Linear warmup + Cosine decay
Warmup Steps 17
Weight Decay 0.01
Batch Size 8
Gradient Accumulation 1
Training Epochs 10
Mixed Precision AMP enabled
Hardware Single NVIDIA A100 80GB
Vision Encoder Unfrozen (trainable)
Gradient Checkpointing Enabled

Generation Configuration

Parameter Value
Max Length 100 tokens
Min Length 8 tokens
Num Beams 5
Beam Search Enabled
Task Captioning

Checkpoint Selection

  • Selection Metric: Aggregate validation score (CIDEr + BLEU-4)
  • Best Epoch: Epoch 1 (aggregate score: 0.2626)
  • Training Loss: Decreased from 2.780 (epoch 1) to 1.692 (epoch 10)
  • Final Model: checkpoint_best.pth (selected based on validation performance)

Performance

Quantitative Results

The fine-tuned model shows significant improvements in lexical overlap metrics, with notable trade-offs in semantic understanding:

Metric Base Model Fine-tuned Absolute Ξ” Relative Ξ”
BLEU-4 0.001 0.047 +0.046 +4600%*
ROUGE-L 0.126 0.242 +0.116 +92.06%
BERTScore F1 0.0545 -0.0537 -0.1082 -198.53%
CLIPScore 0.2854 0.2583 -0.0271 -9.49%

*Note: The extreme BLEU-4 improvement is due to the very low baseline (0.001), making relative improvements appear dramatic.

Key Observations

Strengths:

  • Enhanced Lexical Matching: Substantial improvements in BLEU-4 and ROUGE-L indicate better n-gram alignment with reference descriptions
  • Photography Terminology: Model learned to incorporate photographic vocabulary and technical terms
  • Compositional Awareness: Improved description of camera angles, lighting conditions, and visual elements

Trade-offs:

  • Semantic Coherence: Negative BERTScore suggests divergence from reference semantic patterns
  • Visual-Text Alignment: Moderate decrease in CLIPScore indicates reduced image-text semantic alignment
  • Repetitive Patterns: Tendency toward formulaic descriptions with some repetition

Performance Characteristics

The fine-tuning results reveal a model that has specialized in lexical pattern matching for photography descriptions but with trade-offs in semantic understanding. This suggests the model is particularly suited for applications requiring technical photography terminology rather than general-purpose image captioning.

Usage

Installation 

pip install transformers torch pillow

Basic Usage 

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

# Load model and processor
processor = Blip2Processor.from_pretrained("Dataseeds/BLIP2-opt-2.7b-DSD-FineTune")
model = Blip2ForConditionalGeneration.from_pretrained(
    "Dataseeds/BLIP2-opt-2.7b-DSD-FineTune",
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load and process image
image = Image.open("your_image.jpg")
inputs = processor(image, return_tensors="pt").to(model.device, torch.float16)

# Generate caption
generated_ids = model.generate(
    **inputs,
    max_length=100,
    min_length=8,
    num_beams=5,
    do_sample=False
)

caption = processor.decode(generated_ids[0], skip_special_tokens=True)
print(f"Generated caption: {caption}")

Model Architecture

The model maintains the BLIP-2 architecture with the following components:

Core Architecture

  • Vision Encoder: EVA-CLIP ViT-g/14 (unfrozen during fine-tuning)
  • Q-Former: 32-layer transformer bridging vision and language modalities
  • Language Model: OPT-2.7B (2.7 billion parameters)
  • Architecture: BLIP-2 with Q-Former bridging vision and language
  • Bootstrapping: Two-stage pre-training methodology preserved

Technical Specifications

  • Vision Resolution: 364Γ—364 pixels
  • Vision Patch Size: 14Γ—14
  • Q-Former Queries: 32 learnable queries
  • Language Model Layers: 32
  • Total Parameters: ~2.7B (language model) + vision components
  • Precision: Mixed precision (FP16/FP32)

Fine-tuning Approach

  • Vision Encoder: Trainable (freeze_vit: False)
  • Q-Former: Trainable
  • Language Model: Trainable
  • Full Model: End-to-end fine-tuning enabled
  • Gradient Checkpointing: Memory optimization enabled

Training Data & Methodology

Dataset Characteristics

The DataSeeds.AI Sample Dataset (DSD) focuses on:

  • Technical Photography: Camera settings, composition analysis
  • Lighting Descriptions: Ambient, directional, studio lighting analysis
  • Subject Matter: Diverse photographic subjects and styles
  • Annotation Style: Technical scene descriptions (20-30 words typical)

Data Processing

  • Image Preprocessing: blip2_image_train (364Γ—364 resolution)
  • Text Processing: blip_caption processor
  • Evaluation: blip_image_eval for consistent validation
  • Format: Input-output pairs with scene analysis prompts

Recommended Use Cases

  • βœ… Photography scene analysis and technical descriptions
  • βœ… Camera composition and lighting analysis
  • βœ… Product photography captioning
  • βœ… Photography education and training applications
  • ⚠️ General-purpose image captioning (may show repetitive patterns)
  • ❌ Non-photographic content analysis (not optimized)

Citation

If you use this model in your research or applications, please cite:

@article{abdoli2025peerranked,
    title={Peer-Ranked Precision: Creating a Foundational Dataset for Fine-Tuning Vision Models from GuruShots' Annotated Imagery}, 
    author={Sajjad Abdoli and Freeman Lewin and Gediminas Vasiliauskas and Fabian Schonholz},
    journal={arXiv preprint arXiv:2506.05673},
    year={2025},
}

@misc{blip2-opt-dsd-finetune-2024,
  title={BLIP2-OPT-2.7B Fine-tuned on DataSeeds.AI Dataset for Photography Analysis},
  author={Dataseeds},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/Dataseeds/BLIP2-opt-2.7b-DSD-FineTune},
  note={Fine-tuned model for photography scene analysis and technical description}
}

@inproceedings{li2023blip2,
  title={BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models},
  author={Li, Junnan and Li, Dongxu and Xiong, Caiming and Hoi, Steven},
  booktitle={International Conference on Machine Learning},
  pages={19730--19742},
  year={2023},
  organization={PMLR}
}

@article{zhang2022opt,
  title={OPT: Open Pre-trained Transformer Language Models},
  author={Zhang, Susan and Roller, Stephen and Goyal, Naman and Artetxe, Mikel and Chen, Moya and Chen, Shuohui and Dewan, Christopher and Diab, Mona and Li, Xian and Lin, Xi Victoria and others},
  journal={arXiv preprint arXiv:2205.01068},
  year={2022}
}

License

This model is released under the MIT license, consistent with the base BLIP2 model licensing terms.

Acknowledgments

  • Base Model: Salesforce Research for BLIP2 architecture and pre-training
  • Language Model: Meta AI for OPT-2.7B foundation
  • Vision Encoder: OpenAI for EVA-CLIP vision components
  • Dataset: GuruShots photography community for source imagery
  • Framework: Hugging Face Transformers for model infrastructure

Training Artifacts

This repository includes comprehensive training artifacts:

  • Checkpoints: All epoch checkpoints (0-9) plus best performing checkpoint
  • Evaluation Results: Validation metrics for each epoch
  • Training Logs: Complete training and evaluation logs
  • Configuration: Original training configuration and hyperparameters

For questions, issues, or collaboration opportunities, please visit the model repository or contact the DataSeeds.AI team.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
Image-to-Text
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Dataseeds/BLIP2-opt-2.7b-DSD-FineTune

Base model

Salesforce/blip2-opt-2.7b
Finetuned
(12)
this model

Dataset used to train Dataseeds/BLIP2-opt-2.7b-DSD-FineTune

Evaluation results

View on Papers With Code