CATEGORY // GENERATIVE QUALITY

Multimodal aesthetics ranking for image and video post-training.

Overview Production
CATEGORYGenerative QualityKEY METRIC670K+ units, 100% quality rateKEY METRIC10K units

01 / DATASET OVERVIEW

MULTIMODAL DATA

SCOPE & PURPOSE

The training task.

  • MULTIMODAL RL
  • AESTHETICS RANKING
  • POST-TRAINING

Text-to-image/video foundation-model post-training, plus a multimodal RL aesthetics-ranking project (1 prompt + 7 generated images, ranked across multiple dimensions per group).

02 / DATASET PROFILE & TECHNICAL SIGNALS

SCOPE · SCALE · CONSTRUCTION
01KEY METRIC
670K+ units, 100% quality rate
02KEY METRIC
10K units
03RANKING GROUP
1 prompt + 7 images
Generated images grouped under a single prompt.
04RANKING SCOPE
Multiple dimensions
Aesthetic comparisons within each group.

03 / DATA PRODUCTION & CONSTRUCTION

METHOD & WORKFLOW
  1. 01POST-TRAIN

    Foundation-model quality

    Support text-to-image and text-to-video foundation-model post-training.

  2. 02GROUP

    Prompt-based image sets

    Each ranking group pairs one prompt with seven generated images.

  3. 03RANK

    Multidimensional aesthetics

    Rank each group across multiple dimensions.

  4. 04APPLY

    Multimodal RL

    Produce aesthetics-ranking data for multimodal reinforcement learning.

04 / COVERAGE & CAPABILITY MATRIX

TASKS · METHODS · APPLICATIONS
DIMENSIONCOVERAGE
01Model modalities
Text to imageText to video
02Training stage
Foundation-model post-trainingMultimodal RL
03Comparison unit
1 prompt7 generated images
04Annotation scope
Aesthetics rankingMultiple dimensions
Explore the production method

NEXT STEP