With the rapid advancement of Multi-modal Large Language Models (MLLMs), MLLM-based Image Quality Assessment (IQA) methods have shown promising generalization. However, directly extending these MLLM-based IQA methods to PCQA remains challenging. On the one hand, existing PCQA datasets are limited in scale, which hinders stable and effective instruction tuning of MLLMs. On the other hand, due to large-scale image-text pretraining, MLLMs tend to rely on texture-dominant reasoning and are insufficiently sensitive to geometric structural degradations that are critical for PCQA. To address these gaps, we propose a novel MLLM-based no-reference PCQA framework, termed GT-PCQA, which is built upon two key strategies. First, to enable stable and effective instruction tuning under scarce PCQA supervision, a 2D-3D joint training strategy is proposed. This strategy formulates PCQA as a relative quality comparison problem to unify large-scale IQA datasets with limited PCQA datasets. It incorporates a parameter-efficient Low-Rank Adaptation (LoRA) scheme to support instruction tuning. Second, a geometry-texture decoupling strategy is presented, which integrates a dual-prompt mechanism with an alternating optimization scheme to mitigate the inherent texture-dominant bias of pre-trained MLLMs, while enhancing sensitivity to geometric structural degradations. Extensive experiments demonstrate that GT-PCQA achieves competitive performance and exhibits strong generalization.
Figures & tables
Fig. 1: Architecture of the proposed GT-PCQA. (a) During the training stage, the model alternates between image pairs and multi-view point cloud pairs. Visual inputs are encoded by the vision encoder and abstracted into compact representations, while attribute-specific text prompts ( e.g., texture-aware or geometry-aware) are embedded into textual representations. The aligned multimodal features are then fed into a LoRA-adapted LLM to perform relative quality comparison, enabling stable and effective instruction tuning under heterogeneous 2D–3D supervision. (b) During the inference stage, the trained MLLM is fully frozen, and each test point cloud is evaluated against an anchor set via soft comparison, ultimately predicting the final quality score.
Type
Methods
SROCC ↑
PLCC ↑
KROCC ↑
RMSE ↓
FR
MSE-p2pl [ 9 ]
0.6277
0.5940
0.4825
2.2815
HD-p2pl [ 9 ]
0.6441
0.6874
0.4565
2.1255
PSNR-yuv [ 10 ]
0.7950
0.8170
0.6196
1.3151
GraphSIM [ 11 ]
0.8783
0.8449
0.6947
1.0321
PointSSIM [ 12 ]
0.6867
0.7136
0.4964
1.7001
NR
ResSCNN [ 14 ]
0.8600
0.8100
-
-
TABLE I: Performance comparison on SJTU-PCQA dataset. The best and second-best NR-PCQA results are highlighted in bold and italic , respectively.
Model
SROCC ↑
PLCC ↑
KROCC ↑
RMSE ↓
w/o DP
0.8045
0.8321
0.6887
1.2415
w/o LoRA
0.8031
0.7845
0.5818
1.7629
w/o AO
0.8345
0.8451
0.6921
1.2314
GT-PCQA(Ours)
0.8953
0.8883
0.7382
0.8629
TABLE II: Ablation studies on DP, AO, and LoRA, where “w/o” denotes removing the corresponding component; in particular, “w/o LoRA” indicates full fine-tuning of all LLM parameters.
Rank ( r )
SROCC ↑
PLCC ↑
KROCC ↑
RMSE ↓
32
0.8375
0.8411
0.6515
3.1118
64
0.8562
0.8556
0.6840
2.8394
128
0.8953
0.8883
0.7382
0.8629
256
0.8601
0.8241
0.6638
1.6082
TABLE III: Ablation study on LoRA rank r using the SJTU-PCQA dataset. Bold indicates the best performance.
Methods
SROCC ↑
PLCC ↑
KROCC ↑
RMSE ↓
MM-PCQA [ 16 ]
0.4988
0.4354
0.2929
2.2103
Compare2Score [ 23 ]
0.5157
0.4589
0.3829
2.0103
GT-PCQA(Ours)
0.6708
0.6339
0.4579
1.3672
TABLE IV: Cross-dataset evaluation from SJTU-PCQA to WPC. GT-PCQA is trained on SJTU-PCQA, and no fine-tuning is performed on WPC. Bold indicates the best performance.