| --- |
| pretty_name: "Overlai.ai | 8K Resolution: 8K vs 1080p Video Training Dataset" |
| license: cc-by-4.0 |
| task_categories: |
| - video-classification |
| - feature-extraction |
| size_categories: |
| - n<1K |
| language: |
| - en |
| tags: |
| - 8k |
| - 4k |
| - 1080p |
| - video-generation |
| - video |
| - spatial-fidelity |
| - resolution |
| - synthetic-resolution |
| - upscaling |
| - video-training-data |
| - benchmark |
| - spatial-provenance |
| - image-quality |
| - world-models |
| - fine-tuning |
| - overlai.ai |
| - 8K-video |
| - 8k-video |
| - 8K-Resolution |
| - 8k-resolution |
| --- |
| |
|  |
|
|
| # 8K RESOLUTION | 8K vs 1080p vs Synthetic 4K |
|
|
| A controlled video dataset comparing native 8K footage, clean 1080p derivatives, and synthetic 4K upscales to evaluate spatial fidelity, resolution trade-offs, and source authenticity in AI video model training. |
|
|
| **8K Resolution** is a controlled video dataset designed to test whether models care where pixels came from. |
|
|
| Each source shot is provided in three matched spatial variants: |
|
|
| - **8K** — original native-resolution source |
| - **1080p** — clean lower-resolution derivative created from the 8K source |
| - **4K-Synthetic** — 4K version upscaled from the 1080p derivative |
|
|
| The dataset is designed to explore both **spatial density** and **spatial contamination** in video training data. |
|
|
| ## Key Features |
|
|
| - 🎥 **Matched Variants**: Every source clip is presented in 8K, 1080p, and synthetic 4K variants, enabling direct A/B/C testing using the same underlying scene. |
| - 🔎 **Spatial Density**: Compare native 8K against clean 1080p to test whether additional real spatial information improves model performance. |
| - 🧪 **Synthetic Resolution Testing**: Compare clean 1080p against synthetic 4K to test whether increasing nominal resolution without recovering native detail changes model behavior. |
| - 🔍 **Spatial Provenance Detection**: The dataset can be used to test whether models can distinguish native high-resolution imagery from upscaled imagery. |
| - 📋 **Detailed Metadata**: Every file includes structured metadata identifying the matched source, delivered resolution, source resolution, spatial provenance, conversion method, and other technical characteristics. |
| - 🎬 **Real-World Visual Detail**: Includes aerial footage, landscapes, people, animals, architecture, water, foliage, macro imagery, static scenes, and camera movement. |
| - 🧠 **World Model Relevance**: Designed for research into visual detail, object boundaries, geometry, fine texture, scene understanding, and physical-world representation. |
|
|
| ## Dataset Details |
|
|
| - **Total source clips:** 50 |
| - **Variants per source:** 3 |
| - **Total videos:** 150 |
| - **Conditions:** 8K, 1080p, 4K-Synthetic |
| - **Native 8K resolution:** 7680x4320 |
| - **1080p resolution:** 1920x1080 |
| - **Synthetic 4K resolution:** 3840x2160 |
| - **Video types:** Aerial, Nature, People, Animals, Architecture, Macro, Motion, Static, Tripod, Dolly |
|
|
| ## Technical Details |
|
|
| - **8K condition:** Native 7680x4320 source footage |
| - **1080p condition:** 1920x1080 derivative created from the native 8K source |
| - **4K-Synthetic condition:** 3840x2160 version created by upscaling the 1080p derivative |
| - **Container:** MP4 |
| - **Spatial structure:** Matched source clips across all three conditions |
| - **8K derivation:** Original source |
| - **1080p derivation:** Downsampled from native 8K |
| - **Synthetic 4K derivation:** Upscaled from the 1080p version |
| - **Spatial provenance:** Explicitly labeled in metadata |
|
|
| ## Primary Research Questions |
|
|
| 8K Resolution is designed to support multiple controlled research questions around spatial density, synthetic resolution, and spatial provenance. |
|
|
| ### 1. Spatial Density |
|
|
| Does native 8K provide useful additional spatial information for video models compared with the same source reduced to 1080p? |
|
|
| **Suggested comparison:** |
|
|
| ```text |
| 8K vs 1080p |
| ``` |
|
|
| This comparison isolates the effect of additional real spatial detail while holding the underlying scene constant. |
|
|
| ### 2. Synthetic Resolution |
|
|
| Does converting clean 1080p footage into synthetic 4K create useful training signal, or simply produce a higher-resolution file without restoring the original captured detail? |
|
|
| **Suggested comparison:** |
|
|
| ```text |
| 1080p vs 4K-Synthetic |
| ``` |
|
|
| This comparison tests whether increasing nominal resolution after spatial information has already been removed changes model training, evaluation, or representation quality. |
|
|
| ### 3. Native vs Upscaled Detection |
|
|
| Can a model reliably distinguish native high-resolution imagery from upscaled imagery? |
|
|
| Potential tasks include: |
|
|
| - native vs synthetic resolution classification |
| - upscaling artifact detection |
| - edge and texture consistency analysis |
| - spatial provenance estimation |
| - resolution authenticity scoring |
| - native-detail confidence estimation |
|
|
| The goal is to explore whether models can identify synthetic or low-confidence spatial provenance before that data enters a training pipeline. |
|
|
| ### 4. Metadata vs Spatial Reality |
|
|
| How reliably can nominal resolution metadata be used as a proxy for actual captured spatial information? |
|
|
| A file may report a resolution of 3840x2160 or 7680x4320 while containing spatial information derived from a substantially lower-resolution source. |
|
|
| The research question is: |
|
|
| **Does a file labeled 4K actually contain native 4K spatial information?** |
|
|
| Resolution metadata describes the dimensions of the delivered file, but not necessarily the amount of original detail captured by the camera. |
|
|
| ### 5. Detail Sensitivity |
|
|
| Are some scene types more sensitive to spatial degradation or synthetic upscaling than others? |
|
|
| Possible categories include: |
|
|
| - foliage |
| - hair |
| - skin texture |
| - architecture |
| - text and signage |
| - distant objects |
| - rocks and terrain |
| - water |
| - fine edges |
| - macro detail |
| - aerial landscapes |
|
|
| This may help identify which visual domains benefit most from native high-resolution source data. |
|
|
| ### 6. Downstream Model Effects |
|
|
| Does native spatial fidelity affect downstream model behavior differently depending on task? |
|
|
| Potential areas include: |
|
|
| - video generation |
| - image generation |
| - super-resolution |
| - object recognition |
| - scene understanding |
| - segmentation |
| - visual representation learning |
| - world models |
| - physical reasoning |
| - fine-tuning |
|
|
| A useful result may be that native high-resolution information matters substantially for some tasks while having little measurable effect on others. |
|
|
| ### 7. Scraped Dataset Filtering |
|
|
| Can these matched examples help train or evaluate systems that identify low-confidence spatial provenance in large scraped video datasets? |
|
|
| Scraped video pipelines often have access to the delivered resolution and file metadata, but not necessarily the original capture resolution or processing history. |
|
|
| A file labeled 4K may have been: |
|
|
| - captured natively in 4K |
| - downscaled from a higher-resolution source |
| - upscaled from 1080p |
| - transcoded multiple times |
| - processed through an unknown enhancement pipeline |
|
|
| This dataset can be used to explore whether spatial provenance can become a measurable filtering signal before video enters a training corpus. |
|
|
| ### 8. Provenance Confidence |
|
|
| Can models assign a confidence score to whether a video's declared resolution reflects genuine source detail? |
|
|
| Rather than treating spatial provenance as a simple native/synthetic label, researchers may be able to estimate confidence that the observed spatial structure is consistent with genuine high-resolution capture. |
|
|
| This could support future dataset-quality pipelines where spatial provenance is treated as a first-class quality signal. |
|
|
| ## Core Thesis |
|
|
| Declared resolution is not the same thing as captured spatial information. |
|
|
| A video can be delivered as 4K without containing native 4K detail. |
|
|
| A video can be delivered as 8K without containing native 8K detail. |
|
|
| This dataset is designed to help test the difference between: |
|
|
| ```text |
| file dimensions |
| ``` |
|
|
| and: |
|
|
| ```text |
| source-authentic spatial information |
| ``` |
|
|
| The central question is: |
|
|
| **Does the model care whether the pixels were actually captured?** |
|
|
| ## Metadata |
|
|
| The included `metadata.csv` contains structured fields describing each file and matched source. |
|
|
| Fields include: |
|
|
| - `filename` |
| - `file_name` |
| - `source_id` |
| - `variant` |
| - `spatial_provenance` |
| - `source_resolution` |
| - `conversion_method` |
| - `resolution` |
| - `width` |
| - `height` |
| - `aspect_ratio` |
| - `frame_rate` |
| - `duration_sec` |
| - `file_size_mb` |
| - `slow_motion` |
| - `capture_framerate` |
| - `capture_framerate_source` |
| - `camera_platform` |
| - `camera_movement` |
| - `shot_type` |
| - `dataset` |
|
|
| Each `source_id` corresponds to a matched triplet containing: |
|
|
| ```text |
| 8K |
| 1080p |
| 4K-Synthetic |
| ``` |
|
|
| This allows direct comparison across spatial conditions using the same underlying scene. |
|
|
| ## Suggested Uses |
|
|
| This dataset may be useful for: |
|
|
| - evaluating whether native 8K improves model representations |
| - comparing native high-resolution footage with lower-resolution derivatives |
| - detecting upscaled or synthetic resolution |
| - evaluating spatial provenance classifiers |
| - studying spatial artifacts introduced by upscaling |
| - benchmarking resolution-aware training pipelines |
| - filtering scraped video datasets |
| - fine-tuning video and multimodal models |
| - evaluating world-model sensitivity to image detail |
| - studying whether nominal resolution accurately reflects source information |
|
|
| ## What is Overlai.ai? |
|
|
| Overlai.ai specializes in high-fidelity video datasets for AI training, including 8K, 60p, HDR, and other source-authentic media designed to improve photo, video, and world models. |
|
|
| Contact us for the complete dataset list or access to larger commercial datasets. |
|
|
| ## Contact |
|
|
| *luke@overlai.app* |
