Wildfire Spread Scenarios: Increasing Sample Diversity of Segmentation Diffusion Models with Training-Free Methods
| Authors | Sebastian Gerard & Josephine Sullivan |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2603.20188 |
| Download | |
| Categories | cs.CV |
Abstract
Predicting future states in uncertain environments, such as wildfire spread, medical diagnosis, or autonomous driving, requires models that can consider multiple plausible outcomes. While diffusion models can effectively learn such multi-modal distributions, naively sampling from these models is computationally inefficient, potentially requiring hundreds of samples to find low-probability modes that may still be operationally relevant. In this work, we address the challenge of sample-efficient ambiguous segmentation by evaluating several training-free sampling methods that encourage diverse predictions. We adapt two techniques, particle guidance and SPELL, originally designed for the generation of diverse natural images, to discrete segmentation tasks, and additionally propose a simple clustering-based technique. We validate these approaches on the LIDC medical dataset, a modified version of the Cityscapes dataset, and MMFire, a new simulation-based wildfire spread dataset introduced in this paper. Compared to naive sampling, these approaches increase the HM IoU* metric by up to 7.5% on MMFire and 16.4% on Cityscapes, demonstrating that training-free methods can be used to efficiently increase the sample diversity of segmentation diffusion models with little cost to image quality and runtime. Code and dataset: https://github.com/SebastianGer/wildfire-spread-scenarios
Engineering Breakdown
Plain English
This paper tackles the challenge of efficiently sampling from diffusion models to predict multiple plausible outcomes in uncertain environments like wildfire spread, medical diagnosis, and autonomous driving. Naively sampling from diffusion models requires hundreds of samples to capture low-probability but operationally critical modes, making it computationally expensive. The authors adapt particle guidance and SPELL—techniques originally designed for diverse image generation—to discrete segmentation tasks, and propose a new clustering-based method to achieve sample-efficient ambiguous segmentation. The result is a set of training-free sampling strategies that can find diverse, meaningful predictions without requiring massive computational overhead.
Core Technical Contribution
The core novelty is demonstrating that training-free sampling guidance methods designed for continuous image generation can be effectively adapted to discrete segmentation tasks while maintaining diversity. The authors extend particle guidance and SPELL from natural image synthesis to segmentation, showing these techniques generalize across different modalities. They additionally introduce a new clustering-based approach (implied by 'simple clu...' in the abstract) that further improves sample efficiency for finding operationally relevant low-probability modes. This is non-trivial because segmentation has fundamentally different output spaces and loss landscapes than continuous image generation, requiring architectural adjustments and validation.
How It Works
The method starts with a pre-trained diffusion model that has learned a multi-modal distribution over possible segmentation outputs given an input image. During inference, instead of sampling uniformly from the diffusion process, the authors apply guidance mechanisms that steer the sampling trajectory toward diverse regions of the learned distribution. Particle guidance works by maintaining multiple particles (candidate samples) and encouraging them to explore different modes through repulsion forces or diversity objectives computed during the reverse diffusion process. SPELL (likely a mode-seeking method) is adapted to discretize the segmentation space and guide sampling toward distinct clusters of plausible outputs. The proposed clustering-based method presumably groups similar samples early in inference to avoid redundant sampling and focus computational budget on truly diverse predictions. All methods operate training-free, meaning no additional model parameters are learned—only the sampling strategy changes.
Production Impact
For autonomous driving and medical imaging systems, this approach directly reduces the compute cost of uncertainty quantification by requiring far fewer forward passes through the diffusion model to capture critical edge cases. Instead of running 300+ samples to find a dangerous corner case in autonomous driving perception, you might need 20-50 samples with guidance, cutting inference latency by 6-15x. In medical diagnosis workflows, this enables practitioners to quickly explore multiple plausible segmentations of ambiguous regions (e.g., tumor boundaries) without running the model hundreds of times. The training-free nature is valuable because it works with existing pre-trained models—no retraining needed, reducing deployment complexity. However, the trade-off is added complexity in the sampling loop: you must implement and tune guidance mechanisms, manage particle states, and determine diversity metrics, which increases code maintainability burden compared to vanilla sampling.
Limitations and When Not to Use This
The paper assumes that a well-trained diffusion model already captures the true multi-modal distribution; if the model is under-trained or mode-collapsed, guidance cannot recover missing modes. It is unclear how these methods scale to very high-resolution segmentations or 3D volumetric data where the output space explodes combinatorially—the techniques may require expensive distance metrics or clustering in discrete spaces. The paper does not clearly define what 'operationally relevant' low-probability modes means in practice or provide a principled way to set diversity thresholds, leaving practitioners to tune hyperparameters by hand. Additionally, there is no analysis of failure modes when guidance methods push samples into regions the model assigns low likelihood, potentially producing nonsensical or adversarial segmentations outside the learned distribution.
Research Context
This work builds on the growing body of research applying diffusion models to structured prediction tasks, following successes in image generation (DDPM, Latent Diffusion) and extending them to segmentation. It directly builds on particle guidance and SPELL, which were designed for image diversity in generation; the novelty is demonstrating these generalize to fundamentally different tasks. The paper contributes to the broader challenge of uncertainty quantification in vision models, addressing a gap between theoretical multi-modality and practical computational cost—a critical problem for safety-critical applications like autonomous driving and medical imaging. This direction likely opens follow-up work on guidance methods for other discrete structured outputs (point clouds, scene graphs, language) and theoretical analysis of why image guidance methods transfer to segmentation.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::