We present Dehaze-then-Splat, a two-stage pipeline for multi-view smoke removal and novel view synthesis developed for Track~2 of the NTIRE 2026 3D Restoration and Reconstruction Challenge. In the first stage, we produce pseudo-clean training images via per-frame generative dehazing using Nano Banana Pro, followed by brightness normalization. In the second stage, we train 3D Gaussian Splatting (3DGS) with physics-informed auxiliary losses -- depth supervision via Pearson correlation with pseudo-depth, dark channel prior regularization, and dual-source gradient matching -- that compensate for cross-view inconsistencies inherent in frame-wise generative processing. We identify a fundamental tension in dehaze-then-reconstruct pipelines: per-image restoration quality does not guarantee multi-view consistency, and such inconsistency manifests as blurred renders and structural instability in downstream 3D reconstruction.Our analysis shows that MCMC-based densification with early stopping, combined with depth and haze-suppression priors, effectively mitigates these artifacts. On the Akikaze validation scene, our pipeline achieves 20.98,dB PSNR and 0.683 SSIM for novel view synthesis, a +1.50,dB improvement over the unregularized baseline.
Figures & tables
Figure 2: Per-frame dehazing quality versus multi-view consistency on Akikaze. (a, b) Ground-truth clean images exhibit consistent color and texture across adjacent views. (c, d) Nano Banana Pro outputs, despite global brightness normalization, show visible differences in color rendition and texture detail. Red boxes highlight the plant region where color shifts are most apparent. Such cross-view inconsistencies degrade downstream 3DGS reconstruction.
Scene
Init Points
BG Color
Notes
Akikaze
50k
255
Validation (has GT)
Futaba
50k
255
Bright outdoor
Koharu
50k
255
High brightness var.
Midori
50k
255
Low contrast
Hinoki
20k
0
Dark, dense smoke
Natsume
50k
255
Test scene
Table 1: Per-scene configuration.
Configuration
PSNR ↑
SSIM ↑
LPIPS ↓
Step
Baseline (Default, STOP=5k)
19.48
0.632
0.432
20k
+DCP+Depth
20.27
0.644
0.396
20k
+MCMC (STOP=15k)
20.53
0.651
0.450
20k
+MCMC+DCP+Depth (STOP=15k)
20.57
0.655
0.450
20k
+MCMC+DCP+Depth (STOP=3k)
20.92
0.680
0.598
2k
+Dual-source Gradient
20.98
0.683
0.600
2k
Table 2: Ablation on Akikaze NVS (4 test views). Best-PSNR checkpoint per row. λDCP=0.01 , λdepth=0.1 , λgrad=0.1 .
Figure 3: Training progression without early stopping (DefaultStrategy, STOP=15k, no auxiliary losses). Novel-view quality peaks at early steps and degrades catastrophically with continued training. By step 18k, Gaussians have exploded from 104k to 933k, producing severe floater artifacts, color bleeding, and structural collapse.
STOP=5k
STOP=15k
PSNR
SSIM
PSNR
SSIM
Default
20.39
0.667
19.99 †
0.667 †
MCMC
20.86
0.677
20.58
0.674
Table 3: Decoupling MCMC vs. STOP (no auxiliary losses, best-PSNR checkpoint). † Catastrophic degradation: PSNR drops to 17.84, Gaussians explode to 933k by step 18k.
Method
PSNR
Brightness Std ↓
Nano Banana Pro + GT norm
20.07
0.012
Nano Banana Pro (raw)
19.25
0.120
Nano Banana 2
18.72
0.094
MB-TaylorFormer-L
17.00
0.048
DCP (classical)
∼ 11.9
—
Original smoky
11.01
—
Table 4: 2D dehazing quality on Akikaze (per-frame PSNR vs. GT, averaged over 25 training views).
Figure 4: Comparison of 2D dehazing methods on Akikaze (view 0001). Nano Banana Pro with GT normalization achieves the highest per-frame PSNR (20.07 dB), substantially outperforming MB-TaylorFormer (17.00 dB) and classical DCP ( ∼ 11.9 dB).
Configuration
PSNR
SSIM
LPIPS
Step
Full pipeline (all losses)
20.98
0.683
0.600
2k
w/o gradient loss
20.92
0.680
0.598
2k
Baseline (no aux. losses)
19.48
0.632
0.432
20k
Route B (Dehaze3DGS)
10.28
0.573
0.780
20k
Table 5: Final NVS results on Akikaze (4 test views).
Figure 5: Qualitative comparison of novel view synthesis on Akikaze (test view 0026). (a) Ground truth. (b) Our full pipeline produces sharp renders with accurate color. (c) Baseline without auxiliary losses shows moderate blur and color shift. (d) End-to-end scattering decomposition (Route B) fails to separate smoke from scene content. Bottom row: zoom crops of color checker (red box) and bear face (cyan box) regions.
School of Computer Science and Technology, Hangzhou Dianzi University · School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen