Forest fire monitoring in remote sensing imagery has long relied on traditional perception models that primarily focus on detection or segmentation. However, such approaches fall short in understanding complex fire dynamics, including contextual reasoning, fire evolution description, and cross-modal interpretation. With the rise of multi-modal large language models (MLLMs), it becomes possible to move beyond low-level perception toward holistic scene understanding that jointly reasons about semantics, spatial distribution, and descriptive language. To address this gap, we introduce FireMM-IR, a multi-modal large language model tailored for pixel-level scene understanding in remote-sensing forest-fire imagery. FireMM-IR incorporates an infrared-enhanced classification module that fuses infrared and visual modalities, enabling the model to capture fire intensity and hidden ignition areas under dense smoke. Furthermore, we design a mask-generation module guided by language-conditioned segmentation tokens to produce accurate instance masks from natural-language queries. To effectively learn multi-scale fire features, a class-aware memory mechanism is introduced to maintain contextual consistency across diverse fire scenes. We also construct FireMM-Instruct, a unified corpus of 83,000 geometrically aligned RGB–IR pairs with instruction-aligned descriptions, bounding boxes, and pixel-level annotations. Extensive experiments show that FireMM-IR achieves superior performance on pixel-level segmentation and strong results on instruction-driven captioning and reasoning, while maintaining competitive performance on image-level benchmarks. These results indicate that infrared–optical fusion and instruction-aligned learning are key to physically grounded understanding of wildfire scenes.

The closed-loop testing of autonomous driving systems critically depends on large-scale libraries of diverse and realistic 3D vehicle assets, yet current pipelines still rely on labor-intensive modeling or multi-view capture, making efficient construction a key bottleneck. To overcome this bottleneck and enable convenient, cost-effective 3D asset generation, we propose a semantic prior-guided framework for accurate and robust vehicle point cloud reconstruction from casually captured single-view photographs. Our framework is built on a diffusion backbone but is fundamentally driven by two forms of prior knowledge: First, geometric and appearance priors from camera-aware image features, masks, and distance-transform maps are projected onto the evolving point cloud, compensating for the severe information loss in single-view inputs. Second, we introduce distillation-style regulators—pretrained neural networks that encode vehicle type and model semantics; they act as teacher networks that impose high-level constraints on the generated point clouds, transferring rich semantic knowledge and effectively regularizing the learning process. With these priors, our model infers vehicle-specific semantics from limited observations and reconstructs high-quality 3D point cloud assets. On the 3DRealCar++ dataset, our method clearly surpasses state-of-the-art point cloud baselines in both F-score and Chamfer Distance.