Cyclone: Diffusion Model for Cycle-Consistent Weather Editing from Unpaired Driving Data

arXiv 2026

Thang-Anh-Quan Nguyen1,2, Moussab Bennehar1, Luis Roldão1, Nathan Piasco1,
Dzmitry Tsishkou1, Laurent Caraffa3, Jean-Philippe Tarel2, Roland Brémond2
1Noah's Ark, Huawei Paris Research Center, France
2COSYS, Gustave Eiffel University, France
3LASTIG, IGN-ENSG, Gustave Eiffel University, France
arXiv Code

All-in-One Realistic Weather Editing. We introduce Cyclone, a diffusion-based framework for weather editing in autonomous driving. It can both synthesize high-fidelity diverse weather conditions (bottom) and remove adverse effects (top). Notably, this is achieved without requiring any paired data during training.

Abstract

Reliable perception under diverse weather conditions remains a major challenge for autonomous driving systems. A common strategy to improve robustness is either to synthesize adverse weather conditions for training perception models or to apply weather-removal techniques to recover clean inputs. However, existing approaches typically rely on synthetic data augmentation or physics-based, task-specific models that require paired training data and often struggle to generate realistic weather effects or generalize robustly to out-of-domain scenarios. Toward this problem, we present Cyclone, a unified framework for weather editing based on latent diffusion, equipped with cycle-consistent constraints and knowledge from image-text models. Cyclone enables the generation of multiple weather conditions across diverse scenes while eliminating the need for paired data. Experimental results show that our approach produces more realistic, structure-preserving outputs than existing baselines and leads to consistent improvements across several downstream driving perception tasks. Furthermore, we demonstrate that Cyclone can be distilled to a video diffusion model for temporally consistent weather editing.

Cyclone is/has:
  • ✅ All-in-one weather (clear, fog, rainy, snowy, night) editing model
  • ✅ No pair required for training
  • ✅ No additional discriminator training needed
  • ✅ No auxiliary visual cues (e.g., semantic, depth, or edge maps).
  • ✅ No identity overfitting
  • ✅ Implicit cross-domain invariance regularization
  • ✅ Strong generalization

Results

Adding and removing weather ☀️🌫️🌧️❄️🌙

Input Foggy day Foggy night
Input Snowy day Snowy night
Input Snowy day Clear night
Input Clear day Foggy day Rainy day

From single images to videos

Cyclone can be extended to video inference by generating pseudo GT for paired sequence and fine-tuning of a SVD model for temporally consistent editing.


From single-view to multi-view

In driving datasets, cameras are typically positioned in a structured left-front-right arrangement, providing shared information across views. We simply stack the multi-view images without any view-specific design as the self-attention layers ensure that each viewpoint aggregates information from the others.


On both synthetic and real life

We train and evaluate on both synthetic and real-world datasets, demonstrating its ability to generalize across different domains.


On famous driving datasets

Cyclone can be applied to many driving datasets (e.g., KITTI, nuScenes, Pandaset, Waymo, etc.), allowing perception systems to be tested on more scenarios.


And cities across the world, regardless of day or night


Improve upon driving downstream tasks

Removing adverse weather effects can improve the performance of downstream perception tasks by providing clearer inputs to the models.

We outperform SoTA unpair editing techiques

SoTA methods usually focus on object-centric editing and do not work percisely for weather in driving scenarios.


While being smaller, faster, and more efficient than general instruction-based editing models trained on paired data

Instruction-based editing models relied on hand-crafted paired data, which is infeasible to obtain in this case and often fails to capture the complexity of real-world weather conditions. Competitive results can be achieved with orders of magnitude lower resources: 20x smaller model size, 1000x less training data, and 100x faster inference speed. On the other hand, training with non-pixel-aligned tasks (e.g., novel view synthesis, object insertion/removal, and instant transfer) may weaken spatial alignment in weather editing, resulting in structural distortions when the outputs are displayed as videos.

Method

Our image translation pipeline includes a forward translation from $x_t$ to $\bar{y}$, conditioned on the text prompt $c_y$ and input image $x_0$, followed by a backward translation from $y_t$ to $\bar{x}$, conditioned on $c_x$ and $\bar{y}$. We enforce this cycle consistency using $\mathcal{L}_{\text{recon}}$ and $\mathcal{L}_{\text{cycle}}$. To further stabilize training and prevent collapse to the trivial identity mapping, we introduce a teacher model $D_\phi$ that provides regularization. The teacher is initialized with the same weights as the student model and is subsequently updated using an EMA of the student parameters. In particular, we use the same student model $D_\theta$ across multiple domains.
Cyclone quickly converges to a stable solution, while the baseline which has no regularization collapses to the trivial identity mapping, requiring early stopping.
We show that Cyclone's output can be used to distilled into SVD. SVD equipped with temporal attention not only maintains the editing quality of the original model but also improves consistency throughout the sequence.