NVIDIA-NeMo / NVIDIA-NeMo/Automodel
Support resolution and duration bucketing for diffusion video datasets
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 963
- Forks
- 318
- Avg merge
- 3d 20h
- Merged PRs (30d)
- 143
Description
Problem
The current diffusion video dataset groups samples by aspect ratio and resolution. It does not group them by video duration or frame count.
Real training datasets may contain videos with different durations, such as clips ranging from 2 to 10 seconds. These variable-length videos cannot be efficiently batched based only on resolution.
Requested behavior
Support jointly bucketing video samples by:
- Resolution
- Duration or frame count
This would allow variable-duration videos to be used in the same training dataset while batching videos with compatible spatial and temporal shapes together.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
No file or test is named. Locate the diffusion video dataset implementation and its existing aspect-ratio/resolution bucketing path, then inspect related tests if available. Done means variable-duration videos are jointly bucketed by resolution and duration or frame count for compatible batching.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- data, machine-learning
- Issue type
- Feature
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Quiet
- Clarity
- Mostly clear
- Newbie friendliness
- 45/100