Wan2.2-Fun reward LoRAs: HPS v2.1 on the low-noise expert, settings and side effects

alibaba-pai trained reward LoRAs for Wan2.2-Fun-A14B-InP with reward backpropagation toward human-preference models, one for each of its two experts: Wan2.2-Fun-A14B-InP-high-noise-HPS2.1, -low-noise-HPS2.1, -high-noise-MPS and -low-noise-MPS (.safetensors), all rank 128 with network_alpha 64 and batch size 8, trained 5,000, 2,700, 5,000 and 4,500 steps respectively. The table's Reward Model column swaps HPS v2.1 and MPS on two rows, and the link for the low-noise MPS file points to a Wan2.1-Fun repository; the file names and descriptions are consistent. Which to use: the authors found the low-noise MPS LoRA converges much more slowly and may disappoint, so use the HPS v2.1 LoRA on the low-noise expert. The demos pair high-noise HPS v2.1 with low-noise HPS v2.1, or high-noise MPS with low-noise HPS v2.1, each at weight 0.5, on prompts from T2V-CompBench expanded by Llama-3.3. How: in VideoX-Fun examples/wan2.2_fun/predict_t2v.py set lora_path and lora_weight for the low-noise reward LoRA and lora_high_path and lora_high_weight for the high-noise one. A training script (scripts/wan2.2_fun/train_reward_lora.py, README_TRAIN_REWARD.md) lets you train your own. Side effects reported by the authors: past a point the reward keeps rising while video quality stops improving, as the model learns shortcuts such as adversarial background artefacts; image preference models cannot judge temporal qualities and reduce the dynamism of generated videos, which computing the reward on only the first decoded frame lessens but does not remove. Licence: Apache-2.0. Based on the alibaba-pai/Wan2.2-Fun-Reward-LoRAs card pinned to revision 98ccd8617f3101dc7ca388510dd7e65f5f42d91c and observed on 2026-10-27. Nothing was run here; the demo videos and VideoX-Fun scripts were not inspected, and the low-noise MPS download link may be wrong. 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/Wan2.2-Fun-Reward-LoRAs/raw/98ccd8617f3101dc7ca388510dd7e65f5f42d91c/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 880:2905, 1904:2905, 2917:3153, 8617:8707, 8724:9006, 9155:9922; exact excerpt SHA256 3b7ddf9a33be40b131c18b5e09af32c2616e104ec81aed9f624420439674793a,b61a94941420fbc60d1b2049e9d8877a8d6549d57280e0ed0be002dc9f5823dd,b9d501ea32d7d921fbddbef4ad636b276e75a8f75dbbee31a5540f66ab04a1"
    }
  ]
}