--- license: cc-by-nc-4.0 tags: - nsfw - CivitAI - text-to-image - prompts - metadata extra_gated_fields: organization: type: text label: "Your Organization/Affiliation (Optional)" research_purpose: type: text label: "Please briefly describe your intended use of this dataset and how it aligns with research purposes or community enhancement measures (e.g., training NSFW classifiers, analyzing user behavior)." --- # Image-Metadata Mini Dataset (Civ-40K sampled from Civ-40M) This dataset is licensed under CC BY-NC 4.0, which allows for non-commercial use with proper attribution. While the license governs legal use, we also ask that this dataset **not** be used to: - Train generative models that produce NSFW content. The dataset is intended for **research purposes and community enhancement measures**, such as training NSFW classifiers or analyzing user behavior. Users are encouraged to consider the ethical implications of using this data or else use a different dataset. --- ## Column Preview | Column Name | Description | Example Value | |---------------------|-----------------------------------------------------------------------------|--------------------------------------------------| | `id` | Unique identifier for each image | `14198` | | `createdAt` | Timestamp of when the image was created/uploaded | `2022-11-16 23:48:47.429000+00:00` | | `url` | Direct URL to the image on CivitAI | `https://image.civitai.com/...` | | `positivePrompt` | The prompt input describing what to generate (positive conditioning) | `Maltese Falcon in the style of Kate Greenway...`| | | `negativePrompt` | Negative prompt used to suppress unwanted features | `blurry, bad hands, low quality` | | `nsfw` | Boolean indicating whether the image is NSFW | `False` | | `browsingLevel` | Numerical level of content filtering (1 = SFW, higher = more NSFW) | `1` | | `statsSummary` | Engagement stats (e.g., likes/views, unclear exact meaning here) | `4` | | `usernameHash` | Hash representing the user (anonymized) | `f55e9ac512ce8de1` | | `Model` | Name or ID of the model used (often missing) | `NaN` | | `cfgScale` | Classifier-free guidance scale (higher = more prompt adherence) | `6.5` | | `sampler` | Sampling algorithm used to generate the image | `Euler a` | | `Size` | Resolution of the image in pixels | `512x512` | | `seed` | Random seed used to generate the image | `2465122345` | | `VAE` | Info about the Variational Autoencoder used (always missing here) | `NaN` | | `generationSystem` | Origin system or version name of the generator | `CivitAI` | | `resourceIDs` | List of model or resource IDs used in generation (CivitAI-ID or Autohash) | `[a9263745, a898923]` | --- ## Dataset Description This dataset provides detailed metadata for AI-generated images. Each record includes the prompts used to generate the image, user interactions like likes, and metadata such as the model used and specific generation parameters. The dataset is particularly useful for researchers and developers working on AI image generation, prompt engineering, and user interaction analysis. The dataset includes both positive and negative prompts, indicating desired and undesired features in the generated images. Additionally, it captures NSFW content indicators and browsing level requirements, providing a comprehensive view of the content's accessibility and safety. **Examples of Metadata Fields**: - **Size**: The dimensions of the generated image (e.g., 832x1216). - **Seed**: The seed value used for image generation. - **Steps**: Number of steps used in the image generation process. - **VAE**: Variational Autoencoder details used in the generation process. --- Here’s your LaTeX table content cleanly converted into **Markdown** with a readable layout and formatted notes: --- ### Image-Metadata Mini Dataset (Civ-40K sampled from Civ-40M) | Year | Month | Total Images | NSFW True | NSFW False | Days | Daily Avg | NSFW Ratio | |------|-------|---------------|------------|-------------|------|-------------|--------------| | 2023 | 1 | 26 | 9 | 17 | 31 | 0.84 | 0.35 | | 2023 | 2 | 67 | 32 | 35 | 28 | 2.39 | 0.48 | | 2023 | 3 | 133 | 67 | 66 | 31 | 4.29 | 0.50 | | 2023 | 4 | 164 | 88 | 76 | 30 | 5.47 | 0.54 | | 2023 | 5 | 230 | 119 | 111 | 31 | 7.42 | 0.52 | | 2023 | 6 | 251 | 119 | 132 | 30 | 8.37 | 0.47 | | 2023 | 7 | 325 | 163 | 162 | 31 | 10.48 | 0.50 | | 2023 | 8 | 334 | 150 | 184 | 31 | 10.77 | 0.45 | | 2023 | 9 | 339 | 193 | 146 | 30 | 11.30 | 0.57 | | 2023 | 10 | 365 | 191 | 174 | 31 | 11.77 | 0.52 | | 2023 | 11 | 576 | 346 | 230 | 30 | 19.20 | 0.60 | | 2023 | 12 | 696 | 365 | 331 | 31 | 22.45 | 0.52 | | 2024 | 1 | 782 | 459 | 323 | 31 | 25.23 | 0.59 | | 2024 | 2 | 896 | 514 | 382 | 29 | 30.90 | 0.57 | | 2024 | 3 | 1360 | 895 | 465 | 31 | 43.87 | 0.66 | | 2024 | 4 | 1856 | 1338 | 518 | 30 | 61.87 | 0.72 | | 2024 | 5 | 2412 | 1869 | 543 | 31 | 77.81 | 0.77 | | 2024 | 6 | 3006 | 2417 | 589 | 30 | 100.20 | 0.80 | | 2024 | 7 | 3505 | 2835 | 670 | 31 | 113.06 | 0.81 | | 2024 | 8 | 3958 | 3102 | 856 | 31 | 127.68 | 0.78 | | 2024 | 9 | 4157 | 3301 | 856 | 30 | 138.57 | 0.79 | | 2024 | 10 | 4559 | 3716 | 843 | 31 | 147.06 | 0.82 | | 2024 | 11 | 5088 | 4110 | 978 | 30 | 169.60 | 0.81 | | 2024 | 12 | 5533 | 4405 | 1128 | 31 | 178.48 | 0.80 | --- #### Summary: - **N Images:** 40,636 - **Total NSFW Images:** 30,754,835 - **Dataset Reduction:** The original dataset was reduced to **0.1%** by sampling every 1 in 1,000 images to ensure a representative distribution. #### Usage: This reduced dataset was used in **Section 6.2**, where **MiVOLO** age and gender estimation was applied to estimate the demographics of subjects depicted in imagery posted to CivitAI between **November 2022 and December 2024**. - **NSFW Classification:** - **NSFW-true**: Browsing levels 2, 4, 8, 16 - **NSFW-false**: Browsing level 1 - **NSFW Ratio**: Calculated as `NSFW-true / Total Images` per month. ---