ProductionCrate VFX Dataset 1

Filling the gaps in traditional training data

The ProductionCrate VFX Dataset is uniquely positioned to address the shortfalls of traditional stock-media datasets.

We believe this dataset will fill the gap that generative video AI's require to meet the needs of creative industries, including film production, music videos, games, previsualization and marketing.

Built for the Most Creative AI's

For 10 years, the ProductionCrate VFX dataset has been building content that fulfils the demands of creative professionals. Ranging from explosions, to superpowers, lasers and fire, we have curated content that satisfies the popular creative intentions that traditional media libraries do not possess.

For the film, television and media industry to fully utilize AI, it must cover these creative scenarios that are required by the directors and artists producing media.

Ready to be Fully Licensed

We're proud to have full ownership of our content library, including all stock VFX, motion graphics, music and sound effects.

Each project we've taken on to produce content has maintained a consistent, clearly defined ownership record. Whether the content was produced in-house or from a contracted artists, the content is under our authority.

Documentation can be provided upon request to help verify the ownership of our content.

This clear, simple structure gives us the ability to initiate content deals efficiently, with minimal friction.

Our easy licensing has allowed us to gain trust from the most reputable studios and organizations, with our content being utilized in their productions for film, previsualization, games and music videos.

How was this dataset built?

Over a period of 10+ years, a ProductionCrate has developed a library of specialized stock content that address the needs of filmmakers and artists who required effects that would be otherwise impossible to produce.

The content has been built in-house, as well as from contracted artists, whose produced content is exclusively owned by ourselves. This gives us full control over how the entire VFX library is used, developed and shared.

A variety of techniques were employed, such as practically shooting pyrotechnic elements (muzzle flashes, explosions, fire), as well as professionally built digital simulations for impossible visuals, including portals, superpowers and spaceships.

Due to the nature of our content only requiring a "drag & drop" to be composited onto footage, our data was produced with diverse applicability in mind. Our media has been encoded with transparent backgrounds, making it possible for us to create a unique highly-flexible dataset that meets the specifications of the dataset required.

Demands from the industry, as well as our unique positioning determine the content that we have created. This means the dataset is representative of what artists are likely to attempt creating with any AI's that have been trained with this content.

How big is the dataset?

Due to the transparent nature of our dataset, the quantity of training data highly depends on the format and specifications required from the models utilizing the dataset. We encourage communication to determine what delivery format best suits our client's needs.

Raw RGB+A ProRes Files

10,000 videos

This is the native format that our VFX dataset is distributed in. It includes an alpha channel, which acts as a transparent background. This means only the VFX is exclusively encoded in the video file, without the scene/background that it is within.

These are the largest file formats, ranging between 10MB to 2000MB per video.

Using the ProRes files ensures the best quality, with most files being encoded with 16-bit color precision.

MP4 Files

15,000 - 50,000 videos

These are the classical delivery format for AI video training data, and may better serve the requirements of the model.

With this delivery format, we will composite our dataset onto backgrounds that maximises the efficiency of training. Multiple environments can be used, drastically increasing the amount of possible training data.

Additionally, creative adjustments can be made on a case-by-case basis to even further increase the scale and diversity of the dataset. Coloring, scale, lighting, setting, occlusion and positioning all contribute to this direction.

Includes high-quality human generated training metadata

Our efforts have been invested in understanding the next generation of video based AI training techniques, allowing our training data to be aligned with current training data standards.

We recommend consultation and communication beforehand to identify the required metadata and annotations, as well as the preferred format to be compatible with the model's training system.

We will be happy to process the data to meet the necessary requirements.

A suggestive list of training data can include:

Hard tags (explosion, fire)
Soft tags (blue, energetic, bright)
Human-friendly description (a bright blue magical lightning bolt striking the ground with blue sparks)
Lighting conditions
Camera data when applicable
Timestamps for key events
Position anchors (the source of a laser beam)

Technical support and assistance

Our team is prepared to developing custom solutions for specific challenges, including:

Re-encoding content in a specific format
Providing specific content resolutions
Reformatting training metadata
Creating new training metadata types (eg, Motion Vector)
Creating new specialized datasets on-demand to meet AI-model shortfalls

Samples

You may download a sample of 11 of our VFX data assets from the link below.

ProductionCrate-VFX-Dataset-1-Sample

309 MB

On-Demand Datasets

ProductionCrate content team specializes in creating large quantities of "impossible content" that covers a diverse set of camera angles, lighting conditions and artistic qualities. Our 13 years of experience has given us a powerful pipeline that allows us to create entire batches of training data ready content quickly and at a high-quality, while being cost-efficient.

Our mission for on-demand datasets is to identify weak points in video AI models, and offer the training data to satisfy outlier requirements.