Wildfire Spread Scenarios: Increasing Sample Diversity of Segmentation Diffusion Models with Training-Free Methods

Sebastian Gerard; Josephine Sullivan

cs.CVMar 20, 2026

Wildfire Spread Scenarios: Increasing Sample Diversity of Segmentation Diffusion Models with Training-Free Methods

Authors: Sebastian Gerard, Josephine Sullivan

Organizations: KTH Royal Institute of Technology, Stockholm, Sweden

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

Figures & tables

Figure 1: Diversity-biased sampling: We train a conditional diffusion model to generate different outputs for the same input data. If the goal is to find most, or all, different outputs for the current input, naive sampling can require a large number of samples, due to the redundancy in samples. To reduce this redundancy, we employ methods that bias the sampling towards higher diversity for the same number of samples.

Panel of Figure 2 — Figure 2: MMFire: We use a wildfire spread simulator to generate multiple plausible outcomes based on the current state of the fire. This is done by setting the wind direction to one of eight values across the whole $64\times 64$ image. We impose a highly skewed probability distribution on the eight outcomes during training (see the probabilities above). This represents a difficult situation where naively sampling from the diffusion model is a very slow strategy for finding all modes.

Panel of Figure 3 — Figure 3: Multi-modal binary Cityscapes: The classes road , sidewalk , vegetation , and car are randomly flipped to the positive or negative class, with fixed probabilities, resulting in $2^{4}=16$ separate modes per image.

Figure 4: Applying clustering-based sample-pruning at different sampling steps: Varying after which sampling step the clustering and pruning is performed influences the final performance. For all datasets, there is large gap between evaluating only the cluster centers and evaluating the full set of generated samples. The x-axis represents the index of the sampling step. However, the noise levels at each step do not decrease linearly. See appendix for details on the noise schedule.

Method	Batch size B	Image quality $\uparrow$	Distinct modes $\uparrow$	HM IoU* $\uparrow$	Runtime $\downarrow$
Naive sampling	16	0.956	12.2	0.416	0h41m
	32	0.956	13.3	0.497	1h19m
	64	0.956	14.1	0.586	2h39m
Clustering [B $\rightarrow$ 16]	32	0.953	5.129	0.469	0h46m
	64	0.951	5.680	0.517	0h53m
	128	0.949	5.993	0.552	1h9m

Table 1: Clustering-based sample pruning: Varying initial batch size. We investigate the trade-off between runtime and quality when varying the number of initial samples on Cityscapes. Naive sampling generates a number of samples equal to the batch size, while clustering always generates 16 samples, but starts with a higher number of samples that are clustered.

Method	Image quality $\uparrow$	Distinct modes $\uparrow$	HM IoU* $\uparrow$	Runtime $\downarrow$
	MMFire - 1 batch $\times$ 8 samples
Naive sampling	0.999	3.4	0.638	0h26m
Prob. UNet	0.999	3.1	0.570	0h6m
Clustering [256 $\rightarrow$ 8]	0.999	3.8	0.662	1h50m
PG: batch	0.999	4.0	0.694	0h29m
SPELL: batch	0.999	4.4	0.713	0h25m

Table 2: Single-batch performance: We generate a batch of N samples per input; N is the number of modes of the respective dataset. Methods should produce a diverse set of samples while retaining a high image quality. PG: Particle Guidance. LIDC’s ground truth data does not permit the computation of the image quality and distinct modes metrics.

Method	Image quality $\uparrow$	Distinct modes $\uparrow$	HM IoU* $\uparrow$	Runtime $\downarrow$
	MMFire - 2 batches $\times$ 8 samples
Naive sampling	0.999	4.1	0.705	0h44m
ProbUNet	0.999	3.8	0.637	0h12m
Clustering [256 $\rightarrow$ 16]	0.999	4.4	0.728	2h8m
PG: batch	0.999	4.8	0.751	1h4m
PG: memory bank	0.999	4.6	0.738	1h4m

Table 3: Multi-batch performance: We generate two or four batches of N samples per input; N is the number of modes of the respective dataset. For clustering, we only generate one batch, but increase the number of clusters accordingly. PG: Particle Guidance. LIDC’s ground truth data does not permit the computation of the image quality and distinct modes metrics.

Shield radius $r$	Image quality $\uparrow$	Distinct modes $\uparrow$	HM IoU* $\uparrow$
6.350	0.944	6.6	0.540
9.525	0.934	7.4	0.564
12.700	0.924	7.8	0.561
15.875	0.914	8.1	0.551

Table 4: SPELL: Varying shield radius. Based on diversity statistics on the Cityscapes training set, we perform a coarse search around

r_{0}=12.7

$s_{\min}$	Image quality $\uparrow$	Distinct modes $\uparrow$	HM IoU* $\uparrow$	Runtime $\downarrow$
$\infty$	0.956	4.4	0.416	41m
70	0.938	7.0	0.571	41m
40	0.936	7.3	0.577	40m
20	0.935	7.4	0.571	41m
0	0.934	7.4	0.564	0h40m

Table 5: SPELL: Limit application to high noise. We vary

s_{\min}

, the highest noise level at which SPELL is still applied, to reduce the potentially negative influence during sampling. We begin sampling from pure Gaussian noise with

\sigma_{\max}=80

, the default for the EDM framework.

s_{\min}=\infty

represents never using SPELL.

s_{\min}=0

represents always using SPELL.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	$r_{0}$	r
MMFire	9.525	9.525
Cityscapes	12.700	9.525
LIDC	6.000	6.000

Appendix

Table 6: SPELL: Shield radii used in experiments.

r_{0}

is an initial estimate for a good shield radius determined from dataset diversity statistics in Equation 9 .

r

is the best value we found in a coarse grid search around

r_{0}

. For MMFire and LIDC,

r_{0}

was the best value we found. For Cityscapes, we found in Table 4 that

0.75r_{0}

performs slightly better.

Guidance steps	$\alpha$	Image quality $\uparrow$	Distinct modes $\uparrow$	HM IoU* $\uparrow$
None	0	0.956	4.4	0.416
First	1	0.953	5.2	0.477
First	2.5	0.950	5.8	0.518
First	5	0.946	6.1	0.550
First	10	0.938	6.5	0.575
First	25	0.915	7.0	0.596

Appendix

Table 7: Particle guidance: Varying guidance strength

\alpha

. We investigate the trade-off between image quality and diversity on Cityscapes. Applying guidance on the first step only saves several backward passes, but reaches almost exactly the same performance.

Figure 5: Method comparison - MMFire. We compare the different diversity-encouraging methods on the MMFire dataset, sampled from the same starting noise and conditioning. The order of generated samples is determined via Hungarian matching, such that the samples are positioned below the closest ground truth. This example is cherry-picked for visualization. Non-cherry-picked examples often show the same number of correct samples across most methods, or can have duplicate ground truths. The low difference in visual appearance makes sense when we consider that the best method, SPELL, only performs 7.5% better than naive sampling, and that all methods use the same underlying diffusion model.

Figure 6: Method comparison - Cityscapes. We compare the different diversity-encouraging methods in the Cityscapes dataset, sampled from the same starting noise and conditioning. The order of generated samples is determined via Hungarian matching, such that the samples are positioned below the closest ground truth. This example is cherry-picked for visualization to ensure that the different ground truth modes are easily enough visually distinguishable.

Distance function	Image quality $\uparrow$	Distinct modes $\uparrow$	HM IoU* $\uparrow$	Runtime $\downarrow$
Chamfer	0.948	6.1	0.571	1h47m
L2	0.952	5.9	0.560	1h33m

Appendix

Table 8: Clustering - comparing distance functions: We use our clustering-based approach to prune from 256 initial samples to 16 samples on Cityscapes, using either chamfer or L2 distance. L2 distance performs slightly worse, so we use chamfer distance for all experiments.

Attenuation $\gamma$	Image quality $\uparrow$	Distinct modes $\uparrow$	HM IoU* $\uparrow$
0.00	0.956	4.4	0.416
0.10	0.983	1.5	0.142
0.25	0.995	1.1	0.101
0.50	0.999	1.0	0.085
1.00	1.000	1.0	0.078

Appendix

Table 9: CADS: We impose a noise schedule on the conditioning information, attenuated by a factor

\gamma

. We see that any noise added to the conditioning information greatly hurts performance.

Diffusion Sampling

Diffusion sampling studies how to turn a trained diffusion model into high-quality samples efficiently and reliably.

See all 264 papers

Generating Satellite Imagery Data for Wildfire Detection through Mask-Conditioned Generative AI
Apr 2, 2026Valeria Martin, K. Brent Venable, Derek MorganSynthetic Data Augmentation Conditional Image Generation
Department of Intelligent Systems and Robotics, University of West Florida, Pensacola, FL, USA · IHMC, Pensacola, FL, USA
Deterministic Mode Proposals: An Efficient Alternative to Generative Sampling for Ambiguous Segmentation
Mar 20, 2026Sebastian Gerard, Josephine SullivanSegmentation Uncertainty Estimation Set Prediction
KTH Royal Institute of Technology, Stockholm, Sweden
Diffuse2Seg: Diffusion Models Can Segment Anything Without Supervision
Sep 6, 2026Christoph Hümmer, Joachim Sicking, Fabian Hüger +1Unsupervised Image Segmentation Instance Segmentation
Institute of Mathematics, Technical University of Berlin · CARIAD SE, Volkswagen Group

Apr 2, 2026J/K move · Enter open · S save

Generating Satellite Imagery Data for Wildfire Detection through Mask-Conditioned Generative AI

Valeria Martin, K. Brent Venable, Derek Morgan

Department of Intelligent Systems and Robotics, University of West Florida, Pensacola, FL, USA · IHMC, Pensacola, FL, USA

Synthetic Data Augmentation Conditional Image Generation Diffusion-Based Image Generation Wildfire Detection

Open paper AskPDF arXiv

The scarcity of labeled satellite imagery remains a fundamental bottleneck for deep-learning (DL)-based wildfire monitoring systems. This paper investigates whether a diffusion-based foundation model for Earth Observation (EO), EarthSynth, can synthesize realistic post-wildfire Sentinel-2 RGB imagery conditioned on existing burn masks, without task-specific retraining. Using burn masks derived from the CalFireSeg-50 dataset (Martin et al., 2025), we design and evaluate six controlled experimental configurations that systematically vary: (i) pipeline architecture (mask-only full generation vs. inpainting with pre-fire context), (ii) prompt engineering strategy (three hand-crafted prompts and a VLM-generated prompt via Qwen2-VL), and (iii) a region-wise color-matching post-processing step. Quantitative assessment on 10 stratified test samples uses four complementary metrics: Burn IoU, burn-region color distance (ΔC_burn), Darkness Contrast, and Spectral Plausibility. Results show that inpainting-based pipelines consistently outperform full-tile generation across all metrics, with the structured inpainting prompt achieving the best spatial alignment (Burn IoU = 0.456) and burn saliency (Darkness Contrast = 20.44), while color matching produces the lowest color distance (ΔC_burn = 63.22) at the cost of reduced burn saliency. VLM-assisted inpainting is competitive with hand-crafted prompts. These findings provide a foundation for incorporating generative data augmentation into wildfire detection pipelines. Code and experiments are available at: https://www.kaggle.com/code/valeriamartinh/genai-all-runned

Abstract

Diffusion Sampling

Explore similar work

Generating Satellite Imagery Data for Wildfire Detection through Mask-Conditioned Generative AI

Deterministic Mode Proposals: An Efficient Alternative to Generative Sampling for Ambiguous Segmentation

Diffuse2Seg: Diffusion Models Can Segment Anything Without Supervision