benchmark-vid-gen: models

# Single self-contained catalogue for this example.
#
# All names lowercase -- required, not just tidy: the submission path
# lowercases before a case-sensitive DB lookup.
#
# Measured on-disk sizes. The full set is roughly 71 GB, dominated by
# text encoders rather than the transformers: Wan2.1-T2V-1.3B is 1.3B
# parameters and 29 GB on disk, because the repository carries UMT5-XXL
# and several weight formats.

#
# add: MiniMaxAI/MiniMax-H3

_templates:

  video: &video
    provider: huggingface
    type: txt2vid
    input: text
    output: video

models:

  - <<: *video
    name: wan-ai/wan2.1-t2v-1.3b-diffusers
    description: >
      29 GB. The smallest credible text-to-video. First exercise of the
      txt2vid loader and the mp4 encode path.

  - <<: *video
    name: thudm/cogvideox-2b
    description: >
      14 GB. Different pipeline class, same loader.

  - <<: *video
    name: lightricks/ltx-video
    description: >
      28 GB. Third pipeline class through the same code path.