ImageNet 1K Dataset with Stable Diffusion 1.4 Labels

Imagenet 1K Sd 1.4 pairs the classic ImageNet 1K image classification dataset with labels generated by the Stable Diffusion 1.4 model. Created by ChristophSchuhmann and hosted on Hugging Face, this dataset was last updated in January 2023. It provides a bridge between a foundational computer vision benchmark and modern generative AI outputs.

Use Cases

• Training models to predict Stable Diffusion 1.4's internal image representations from ImageNet image pixels.
• Analyzing the correlation between original ImageNet class labels and the text prompts or latent features produced by Stable Diffusion 1.4.
• Benchmarking the performance of multimodal models on tasks linking the ImageNet visual concepts with generative model outputs.
• Fine-tuning image captioning or text-to-image models using aligned ImageNet and Stable Diffusion 1.4 data pairs.

Strengths

• Leverages the well-established ImageNet 1K dataset, which contains over 1 million images across 1000 classes.
• Provides annotations from a specific, widely-known model version (Stable Diffusion 1.4).
• Hosted on a major platform (Hugging Face) with version tracking, last updated in 2023.

Limitations

• The exact size, structure, and specific column names (e.g., 'sd_label', 'latent_vector') of the derived dataset are unknown.
• Potential label noise or bias inherited from the Stable Diffusion 1.4 model's training data and generation process.
• Temporal staleness, as the underlying generative model (SD 1.4) has been superseded by newer versions.

Provenance

Source

Derived from the ImageNet 1K dataset and annotations generated by Stable Diffusion 1.4.

Collection Method

Likely created by processing ImageNet images through the Stable Diffusion 1.4 model to extract labels, captions, or latent features.

Freshness

Last updated on 2023-01-28.