A preprint reports higher automated scores for a method that groups mixed video-training categories by compatible learning signals, with the clearest gap on CogVideoX-2B. On Wan2.1, grouped training scored 48.18, 1.11 points above the base model and 1.19 above joint LoRA. On CogVideoX-2B, it scored 40.99 versus 27.95 for joint training, a 13.04-point difference. The Wan2.1 gain was less settled: its 95% bootstrap interval over the base model ran from -0.08 to +2.26.
A different way to sort training data
The study asks whether heterogeneous reward-defined categories should be grouped according to their compatibility before they are combined for adapter training. It uses self-generated videos from Wan2.1-T2V-1.3B-Diffusers, divided across 17 non-diversity dimensions from VBench2.0, with approximately 560 videos per dimension.
The G3-LoRA method probes the gradient associated with each reward-weighted category, removes the shared global direction, and clusters the remaining compatibility pattern. It then trains one LoRA expert for each group, merges those experts, and can add on-policy expert distillation along the student's sampling trajectories. In ordinary terms, it tries to keep categories with compatible training updates in the same group.
The raw and residual measurements told different stories. Before the shared direction was removed, all 136 Wan2.1 category pairs had positive raw-gradient similarity. After removal, 75 pairs, or 55%, were negative. The selected five-group partition had a mean within-group similarity of 0.473, compared with 0.345 for four groups. Its separation was 0.633 versus 0.559, and it had no negative within-group pair; the four-group partition left three.
Resampled probes reproduced the Wan2.1 compatibility matrix with a Pearson correlation of 0.847. Residual similarity was associated with held-out-loss changes, with a Spearman correlation of 0.256 and a permutation p-value of 0.003. The main adapter comparison used a matched quick VBench2.0 protocol with 25 prompts per dimension and two generated videos per prompt, applied to the base, joint, partition-control, grouped and OPD adapters.
The first model gave a mixed picture
On Wan2.1, the grouped score of 48.18 was not a clean sweep. It was 4.52 points below the base model on material and 4.00 below it on human interaction. Camera motion scored 12.00 for grouped LoRA, compared with 20.00 for joint LoRA. The mean score therefore showed differences between dimensions.
The uncertainty around that average is important. The gain over base was positive in 96.6% of paired bootstrap resamples, but the 95% interval ran from -0.08 to +2.26, crossing zero. The paper describes the result as a trend rather than a significant difference at the 5% level.
A stronger repeat and a separate check
The CogVideoX-2B comparison showed a wider separation. The base, joint and grouped scores were 39.87, 27.95 and 40.99, respectively. The grouped-versus-joint difference was 13.04 points, with a 95% bootstrap interval from 9.08 to 16.92.
An independent VideoScore-v1.1 evaluation of matched Wan2.1 videos pointed in the same direction. Grouped LoRA's five-aspect mean was 3.080, versus 2.914 for the base model, a difference of 0.166. The paired 95% interval ran from 0.140 to 0.192, and the reported probability that the difference was positive was 1.000.
The study also tested on-policy expert distillation as a way to narrow the gap between the merged adapter and a diagnostic upper bound. The mean rose from 48.18 to 49.91, an increase of 1.73 points, with a paired 95% bootstrap interval from 0.42 to 3.05. That closed about 60% of the gap to the 51.07 upper bound and improved 12 of 17 dimensions over base.
The evidence remains narrow
The result has narrow boundaries. It covers two open text-to-video backbones, each no larger than 2 billion parameters, and a taxonomy derived from VBench2.0. The paper reports no human study, so it offers no direct human-preference or human-video-quality evidence. The authors say they plan to release the implementation, configurations, partitions, prompt manifests and raw evaluator outputs.
Paper data and sources
Original title: G$^3$-LoRA: Organizing Reward-Weighted Video Data with Gradient-Guided Grouped LoRA
Authors: Jia Song, Wenhow Li, Lichen Bai et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-28
DOI: Not available
Original paper · Full text