Wan2.1-Fun reward LoRAs (HPS v2.1, MPS): weights and reward-hacking limits

Wan2.1-Fun-Reward-LoRAs from Alibaba PAI fine-tune Wan2.1-Fun with reward backpropagation (Clark et al., ICLR 2024; Prabhudesai et al., 2023) so outputs better match human preference models. Four LoRAs, all rank 128 with network_alpha 64 and batch size 8: Wan2.1-Fun-1.3B-InP-HPS2.1 (5,000 steps), -1.3B-InP-MPS (7,500 steps), and HPS v2.1 and MPS versions for Wan2.1-Fun-14B-InP (around 4,500 steps). They plug into the matching base model; the training script is in VideoX-Fun (scripts/wan2.1_fun/train_reward_lora.py). Use: set lora_path and lora_weight in VideoX-Fun's examples/wan2.1_fun/predict_t2v.py. The demos (T2V-CompBench prompts expanded by Llama-3.3) use HPS 0.5 and MPS 0.7 on the 1.3B model and 0.7 for both on the 14B model, so start there rather than at 1.0. Limits from the authors: beyond some point the reward keeps rising while quality does not, because the model learns shortcuts such as background artifacts (adversarial patches); image preference models cannot judge motion or temporal consistency and reduce dynamism, which scoring only the first decoded frame lessens but does not remove. The Wan2.2-Fun reward LoRA record covers the A14B equivalents. Licence: Apache-2.0. Based on the alibaba-pai/Wan2.1-Fun-Reward-LoRAs card pinned to revision 7b22a5e497f239c40cc1e12a76e54f3a1b930ee5 and observed on 2026-10-27. Nothing was run here; the demo videos and the predict script were not inspected, and the 14B step counts are summarised as around 4,500. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/alibaba-pai/Wan2.1-Fun-Reward-LoRAs/raw/7b22a5e497f239c40cc1e12a76e54f3a1b930ee5/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 177:683, 1158:2589, 7640:7728, 12884:12972, 12989:13150, 13321:13919, 13690:13919, 4:133; exact excerpt SHA256 404451bb274d7a2a76664954adb7eb4231aa640328a2138cf5a783306eca4d48,9661e2621cbe17a6ce23d2d2b4529291f2e2745196b59e8a372a307d85be9ed8,1d718692682accaedfb92edf0c155daebb929"
    }
  ]
}