metadata
license: apache-2.0
tags:
- video-generation
- reward-model
- videoalign
- adversarial-distillation
AFD VideoReward Discriminators
Trainable scalar VideoReward discriminators initialized from
KlingTeam/VideoReward and adapted with prompt-paired Bradley-Terry loss.
Variants
kling25-turbo-pro/: teacher videos from KlingAI 2.5 Turbo Pro.seedance-1.5-pro/: teacher videos from Seedance 1.5 Pro.openvid-1m/: 200 deterministic OpenVid-1M samples (seed 42, at least 81 frames, 3–8 seconds).
Each variant contains discriminator.pt (r=64 LoRA + rm_head), the exact
warmup YAML, and machine-readable metadata. The Self-Forcing generator is not
included.