BenthicFlow: Generating Extensible Underwater Environments via Flow Matching

Joaquín Figueira; Camile Lendering; Manfred Gonzalez-Hernandez; Giacomo D'Amicantonio; Erkut Akdag; Egor Bondarev

cs.CVAug 24, 2026

BenthicFlow: Generating Extensible Underwater Environments via Flow Matching

Authors: Joaquín Figueira, Camile Lendering, Manfred Gonzalez-Hernandez, Giacomo D'Amicantonio, Erkut Akdag, Egor Bondarev

Organizations: AIMS Group, Department of Electrical Engineering, Eindhoven University of Technology, Eindhoven, The Netherlands

Abstract

Computer vision applications for 3D scene understanding in underwater environments remain challenging due to the lack of high-quality 3D data and the inability of surface-trained models to generalize to underwater scenes. To address this challenge, an emerging trend is to employ generative models to close the data domain gap. However, existing methods assemble large scenes by stitching independently generated tiles post hoc with separately trained models, while demonstrating heterogeneous landscapes only within individual survey sites. We introduce BenthicFlow, a unified framework based on a single conditional flow-matching model that jointly generates aligned textures and depth maps. A MultiDiffusion-inspired sampling procedure reconciles overlapping windows throughout the generative trajectory, enabling spatially extensible RGBD mosaics without a separate stitching model. The generated mosaics are subsequently lifted into explicit 3D benthic environments using surface-aligned Gaussian surfels. Experiments across geographically distinct survey sites demonstrate that BenthicFlow preserves site-specific appearance while generating coherent, large-scale 3D scenes that closely match the target distributions. Code and trained models are available at https://github.com/jacomof/BenthicFlow.

Figures & tables

Figure 1: Overview of BenthicFlow’s generative pipeline. First, reference images are encoded and interpolated to form a large conditioning embedding grid. A large Gaussian noise canvas is then sampled and convolved in overlapping windows by a CFM model ( $v_{\theta}$ ) trained to denoise small latent windows ( $W_{i}$ ). This process is repeated over multiple integration steps, and, on each step, the flow trajectories of overlapping tokens are averaged. Finally, the complete denoised canvas is decoded in a single forward pass into a large RGBD mosaic, which is splatted to render the final scene.

Model	Location	FID $\downarrow$	Cos. Sim. $\uparrow$	KID $\downarrow$
FLUX.2-dev	Batemans	110.3	0.573	$0.0966\pm 0.0018$
	Hawaii	137.1	0.593	$0.1247\pm 0.0024$
	Scott Reef	76.7	0.745	$0.0601\pm 0.0014$
	Average	99.7	0.649	$0.0850\pm 0.0017$
BenthicFlow	Batemans	20.0	0.805	$0.0140\pm 0.0012$
	Hawaii	13.9	0.833	$0.0070\pm 0.0005$

Table 1: Quantitative comparison of image generation quality broken down by geographic region. Campaigns from Batemans Bay (3) and Scott Reef (3) are averaged to summarize regional performance, alongside Hawaii (1) (2000 images are sampled per campaign). KID is reported as the mean

\pm

the pooled standard deviation from subset bootstrapping. BenthicFlow-DPF refers to the model using DPF-Net [ 32 ] pre-processing and Cos. Sim. denotes DINOv2 cosine similarity. Both BenthicFlow variants use 50 sampling steps. Best overall Average results are highlighted in bold .

Real Data
	F1
Model	Batemans	Hawaii	Scott	Bal. Acc. $\uparrow$	Macro F1 $\uparrow$
CLAHE	0.928	0.834	0.850	0.873	0.871
DPF-Net	0.910	0.783	0.869	0.865	0.854
Generated Data
FLUX.2-dev	0.260	0.667	0.697	0.568	0.541

Table 2: Discriminative capacity evaluation using a

k

-NN classifier (

k=20

2,000

training samples per reef). Classifiers are trained on either real data or synthetic data generated by the respective models, and evaluated on the test set. Reef columns report class-specific F1-scores (Scott refers to the Scott Reef location and Batemans abbreviates Batemans Bay), while global performance is captured via Balanced Accuracy (Bal. Acc.) [ 4 ] and Macro F1-score. Best results of the generative models are in bold .

Figure 2: Qualitative comparison of the generative quality of the BenthicFlow models compared to FLUX.2-dev. Each column pair contains the output of a different reference image, with the left column containing the image and the right its predicted depth. The depth column is left blank for FLUX.2-dev, as it does not generate that modality.

Panel of Figure 3 — Figure 3: Analysis of the extensible generation strategy. (a) Visualization of the seam gradient ratio metric across different window overlaps. (b) t-SNE visualization of the interpolation windows between a set of images, demonstrating progressive semantic evolution between references.

Model	RMSE $\downarrow$	Pearson $r$ $\uparrow$
BenthicFlow d-RAE	0.006	0.997
BenthicFlow-DPF d-RAE	0.009	0.992

Table 3: Depth generation consistency evaluated against Depth Anything V2 pseudo-ground truth on the test dataset using root mean squared error (RMSE) and Pearson correlation. Best results in bold .

Figure 4: Two underwater scenes synthesized with BenthicFlow (left) and BenthicFlow-DPF (right) at a scale 32 times larger than the CFM’s training resolution. The reference seeds used to render them are visible within frames on the corners of the scenes.

Surf.	Ref.	Perc.	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
✗	✗	✗	29.20	0.863	0.149
✓	✗	✗	30.03	0.884	0.138
✓	✓	✗	30.96	0.925	0.046
✓	✓	✓	30.94	0.925	0.048
✗	✓	✓	31.33	0.927	0.045

Table 4: Lifting-stage ablation over

300

generated samples (front-view reprojection vs. reference; paired, single view). We report Peak Signal to Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) and LPIPS. The three components of the lifting-stage are ablated: the use of surfels (Surf.), anchored refinement (Ref.) and perceptual color polishing (Perc.). Best per column in bold .

Model	Campaign	FID $\downarrow$	Cos. Sim. $\uparrow$	KID $\downarrow$		Depth RMSE $\downarrow$	Depth $r$ $\uparrow$
BenthicFlow d-RAE	Batemans201011	14.8	0.874	0.0117 $\pm$	0.0011	0.0044	0.997
	Batemans201211	17.3	0.876	0.0130 $\pm$	0.0010	0.0074	0.997
	Batemans201411	23.1	0.870	0.0216 $\pm$	0.0010	0.0041	0.999
	Hawaii201801	15.0	0.883	0.0095 $\pm$	0.0009	0.0052	0.998
	ScottReef200907	8.8	0.929	0.0070 $\pm$	0.0008	0.0061	0.998
	ScottReef201108	7.1	0.920	0.0049 $\pm$	0.0006	0.0072	0.997

Table S1 : Per-campaign d-RAE reconstruction results. Generative fidelity metrics: FID and KID (lower is better), DINOv2 cosine similarity (Cos. Sim.) of mean-pooled tokens (it measures semantic alignment, higher is better); reconstruction is evaluated only for the two flow-based pipelines. Depth consistency metrics: RMSE (lower is better) and Pearson correlation

r

(higher is better) compare reconstructed to reference depth. For each campaign, we highlight the best results between the two models in bold .

Model	Campaign	FID $\downarrow$	Cos. Sim. $\uparrow$	KID $\downarrow$
BenthicFlow	Batemans201011	27.1	0.817	0.0234 $\pm$	0.0017
	Batemans201211	16.0	0.797	0.0075 $\pm$	0.0007
	Batemans201411	16.9	0.801	0.0110 $\pm$	0.0008
	Hawaii201801	13.9	0.833	0.0070 $\pm$	0.0005
	ScottReef200907	8.4	0.888	0.0046 $\pm$	0.0003
	ScottReef201108	9.0	0.870	0.0043 $\pm$	0.0004

Table S2 : Per-campaign conditional generation results. FID and KID measure distributional fidelity (lower is better); DINOv2 cosine similarity (Cos. Sim.) of mean-pooled tokens measures semantic alignment (higher is better). For each campaign the best results between the two models are shown in bold .

Panel of Figure 11 — Figure S1 : Two generated terrains and the variation of their gradient magnitudes in both axes. Cyan dashed lines are used to indicate the region of the image where we would expect noticeable seams to appear ( i.e . the center of the region where two windows overlap and where the pull of multiple trajectories equalizes).

Figure S2 : Comparison of the seam gradient ratio between employing a uniform weight window and the standard sine window of BenthicFlow.

Figure S3 : Additional generated scenes for BenthicFlow (left) and BenthicFlow-DPF (right). All reference seeds, visible in frames below the generated scenes, are sampled from different deployments and campaigns.

Campaign	Split	Deployment	#Images
Batemans201011	Train	r20101117_001912_batemans_03_site1sz	8,872
		r20101117_224623_batemans_06_site6guz	3,165
		r20101118_010326_site6guz_08_broadgrid	10,011
		r20101118_053354_site4sz_09_transect	4,188
		r20101119_224028_site6guz_11_densegrid3	2,411
		r20101120_002302_site5sz_12_densegrids	8,486

Table S3 : Deployment-level data split. Each AUV campaign is split by whole deployments : the validation and test sets each hold out one complete deployment per campaign, never seen during training, so no spatially overlapping imagery leaks across splits. Totals: 510,307 train / 76,480 val / 87,574 test images over 52/7/7 deployments.

3D Scene Generation

3D scene generation creates structured, explorable environments from text, images, maps, or other inputs.

See all 190 papers

WorldFlow3D: Flowing Through 3D Distributions for Unbounded World Generation
Mar 31, 2026Amogh Joshi, Julian Ost, Felix Heide3D Scene Generation 3D Scene Layout Generation
Princeton University · Torc Robotics
WaterGen: Decoupling Scene and Medium in Underwater Image Generation
Jun 30, 2026Jiayi Wu, Tianfu Wang, Tianyi Xiong +6Synthetic Data Augmentation Image Generation
University of Maryland · University of South Florida · University of Florida
Extend3D: Town-Scale 3D Generation
Mar 31, 2026Seungwoo Yoon, Jinmo Kim, Jaesik ParkImage-to-3D Scene Generation 3D Scene Generation
Seoul National University

Mar 31, 2026J/K move · Enter open · S save

WorldFlow3D: Flowing Through 3D Distributions for Unbounded World Generation

Amogh Joshi, Julian Ost, Felix Heide

Princeton University · Torc Robotics

3D Scene Generation 3D Scene Layout Generation Flow Matching

Open paper AskPDF arXiv

Unbounded 3D world generation is emerging as a foundational task for scene modeling in computer vision, graphics, and robotics. In this work, we present WorldFlow3D, a novel method capable of generating unbounded 3D worlds. Building upon a foundational property of flow matching - namely, defining a path of transport between two data distributions - we model 3D generation more generally as a problem of flowing through 3D data distributions, not limited to conditional denoising. We find that our latent-free flow approach generates causal and accurate 3D structure, and can use this as an intermediate distribution to guide the generation of more complex structure and high-quality texture - all while converging more rapidly than existing methods. We enable controllability over generated scenes with vectorized scene layout conditions for geometric structure control and visual texture control through scene attributes. We confirm the effectiveness of WorldFlow3D on both real outdoor driving scenes and synthetic indoor scenes, validating cross-domain generalizability and high-quality generation on real data distributions. We confirm favorable scene generation fidelity over approaches in all tested settings for unbounded scene generation. For more, see https://light.princeton.edu/worldflow3d.

Abstract

3D Scene Generation

Explore similar work

WorldFlow3D: Flowing Through 3D Distributions for Unbounded World Generation

WaterGen: Decoupling Scene and Medium in Underwater Image Generation

Extend3D: Town-Scale 3D Generation