Local Scenes Diffusion Method (LSDM)

• LSDM is a diffusion method that reformulates panoramic generation as a spatially continuous process to address seam artifacts from stitched data.
• The technique synchronizes circular rotations of panorama-aligned inputs and BEV features to maintain structural consistency and control.
• Empirical results on nuScenes-360 show that LSDM improves image coherence and layout fidelity, enhancing downstream BEV segmentation performance.

Local Scenes Diffusion Method (LSDM) is the coherence-oriented component of Percep360, a framework for controllable panoramic street-view generation for autonomous driving. In that usage, LSDM denotes a panorama-specific latent diffusion strategy that treats the $360^\circ$ image as a cyclic spatial domain and uses synchronized circular rotations of panorama-aligned inputs and Bird’s Eye View (BEV) features to mitigate seam artifacts inherited from stitched pinhole supervision. Its central purpose is not generic diffusion modeling, but the conversion of stitched, locally inconsistent panoramic training targets into a setting in which diffusion can learn a more coherent panoramic distribution while preserving control from BEV, depth, mask, and text signals (Teng et al., 9 Jul 2025).

1. Definition and nomenclature

In the Percep360 literature, LSDM refers specifically to the Local Scenes Diffusion Method. It is introduced as the mechanism that “reformulates the panorama generation as a spatially continuous diffusion process, bridging the gaps between different data distributions,” with the term “local scenes” referring to the locally coherent portions of a stitched panorama that remain structurally meaningful despite global seam artifacts (Teng et al., 9 Jul 2025).

This usage is not interchangeable with several unrelated acronym expansions elsewhere on arXiv. The 2016 scene-recognition paper "Locally-Supervised Deep Hybrid Model for Scene Recognition" defines LS-DHM or LSDHM, not “Local Scenes Diffusion Method,” and it is unrelated to diffusion (Guo et al., 2016). The 2026 paper "Semi-Supervised Generative Learning via Latent Space Distribution Matching" uses LSDM for Latent Space Distribution Matching, a framework for semi-supervised conditional generative modeling rather than local-scene diffusion (Chong et al., 4 Mar 2026). The 2023 ultrasound-tracking paper "LSDM: Long-Short Diffeomorphic Motion for Weakly-Supervised Ultrasound Landmark Tracking" uses LSDM for Long-Short Diffeomorphic Motion in medical image analysis (Liu et al., 2023). In encyclopedia use, the unqualified phrase Local Scenes Diffusion Method therefore most precisely denotes the Percep360 mechanism.

2. Problem setting: stitched panoramas as incoherent supervision

Percep360 studies panoramic street-view generation for autonomous driving under a practical constraint: the panoramas are not native panoramic captures, but stitched compositions of multiple pinhole-camera views. The paper argues that such stitched supervision contains two structural defects. First, pinhole cameras introduce sampling information loss: “there are parts of the true panoramic visual field that are not faithfully observed.” Second, stitching introduces spatial misalignment and discontinuities at seams, so that adjacent panorama regions need not follow a smooth common distribution (Teng et al., 9 Jul 2025).

The paper formalizes this by partitioning the panoramic domain $\Omega \subset \mathbb{R}^2$ into six coherent regions $\{\mathcal{R}_i\}_{i=1}^6$ and six aliased regions $\{\mathcal{A}_i\}_{i=1}^6$:

$$\Omega = \left( \bigcup_{i=1}^6 \mathcal{R}_i \right) \cup \left( \bigcup_{i=1}^6 \mathcal{A}_i \right), \quad \mathcal{R}_i \cap \mathcal{A}_j = \emptyset,\quad \forall i,j.$$

The stitched image distribution is then written as

$$f(x) = \sum_{i=1}^6 \mathbb{I}_{\mathcal{R}_i}(x)\cdot f_i(x) + \sum_{i=1}^6 \mathbb{I}_{\mathcal{A}_i}(x)\cdot \tilde{f}_i(x),$$

where $f_i(x)$ denotes the local coherent distribution in region $\mathcal{R}_i$ and $\tilde f_i(x)$ denotes the discontinuous distribution in seam region $\mathcal{A}_i$. The target controlled coherent distribution is modeled as

$\Omega \subset \mathbb{R}^2$0

with the objective of transforming the discontinuous stitched distribution toward a coherent one over the full panorama (Teng et al., 9 Jul 2025).

This formulation is significant because it places LSDM in a specific class of diffusion problems: not merely wide-image generation, but generation from supervision that is already internally inconsistent. The method therefore addresses coherence at the level of spatial parameterization rather than by introducing a new denoising loss.

3. Mathematical formulation and algorithmic mechanism

LSDM preserves the standard latent diffusion formulation and changes the spatial parameterization of conditioning data. Percep360 uses the standard latent encoding and reverse process

$\Omega \subset \mathbb{R}^2$1

$\Omega \subset \mathbb{R}^2$2

$\Omega \subset \mathbb{R}^2$3

$\Omega \subset \mathbb{R}^2$4

and

$\Omega \subset \mathbb{R}^2$5

The paper explicitly states that no new diffusion objective beyond LDM is introduced; the novelty lies in “how the data and controls are spatially reparameterized and aligned” (Teng et al., 9 Jul 2025).

The core LSDM step is to sample a random rotation angle

$\Omega \subset \mathbb{R}^2$6

convert it to a horizontal shift

$\Omega \subset \mathbb{R}^2$7

and apply a circular roll to each panorama-aligned 2D input $\Omega \subset \mathbb{R}^2$8:

$\Omega \subset \mathbb{R}^2$9

with

$\{\mathcal{R}_i\}_{i=1}^6$0

The same angular reparameterization is applied to BEV features. After BEV encoding

$\{\mathcal{R}_i\}_{i=1}^6$1

the feature map is shifted by

$\{\mathcal{R}_i\}_{i=1}^6$2

The paper characterizes this synchronized roll as the concrete implementation of the “spatially continuous diffusion process” (Teng et al., 9 Jul 2025).

The effect is to randomize the angular origin of the panorama during training. A seam-corrupted region in one parametrization becomes an interior region in another, and the model is forced to learn structure that is stable under circular shifts rather than memorize fixed seam coordinates. This suggests a cyclic equivariance constraint, although the paper does not introduce that term formally.

4. Conditioning structure and role inside Percep360

LSDM operates inside a side-controlling latent diffusion architecture inspired by ControlNet and ControlNeXt. The conditioning inputs are the panoramic RGB image $\{\mathcal{R}_i\}_{i=1}^6$3, binary mask $\{\mathcal{R}_i\}_{i=1}^6$4, depth map $\{\mathcal{R}_i\}_{i=1}^6$5, BEV map $\{\mathcal{R}_i\}_{i=1}^6$6, and text prompt $\{\mathcal{R}_i\}_{i=1}^6$7. The 2D inputs are encoded by a frozen VAE encoder, the BEV map by a dedicated BEV encoder following BEVControl, and the text prompt by a CLIP text encoder (Teng et al., 9 Jul 2025).

Within this architecture, LSDM is the coherence mechanism rather than the controllability mechanism. The paper distinguishes it from the Probabilistic Prompting Method (PPM): LSDM “improves coherence by reconfiguring the relation between the network and image-BEV pairs,” whereas PPM “dynamically selects the most relevant control cues,” especially depth and mask priors, to strengthen layout and semantic control (Teng et al., 9 Jul 2025).

This division of labor matters technically. The paper reports that seam compensation by itself can reduce controllability, because the model may alter layout while trying to smooth stitched boundaries. LSDM therefore does not replace structured control; it makes such control usable in the presence of stitched panoramic supervision. A plausible implication is that LSDM should be interpreted as a geometric conditioning strategy over a cyclic domain, not as a full conditional-generation interface on its own.

5. Empirical evaluation

Percep360 is evaluated on nuScenes-360, with 28,130 training images and 6,019 validation images, and trained on two A6000 GPUs for ten days (Teng et al., 9 Jul 2025). The main baselines are a BEVControl-derived baseline and a modified MagicDrive. The reported full-system results are:

• MagicDrive*: BRISQUE $\{\mathcal{R}_i\}_{i=1}^6$8, PIQE $\{\mathcal{R}_i\}_{i=1}^6$9, SSIM $\{\mathcal{A}_i\}_{i=1}^6$0, FID $\{\mathcal{A}_i\}_{i=1}^6$1, Driv. $\{\mathcal{A}_i\}_{i=1}^6$2, Mean. $\{\mathcal{A}_i\}_{i=1}^6$3
• Baseline: BRISQUE $\{\mathcal{A}_i\}_{i=1}^6$4, PIQE $\{\mathcal{A}_i\}_{i=1}^6$5, SSIM $\{\mathcal{A}_i\}_{i=1}^6$6, FID $\{\mathcal{A}_i\}_{i=1}^6$7, Driv. $\{\mathcal{A}_i\}_{i=1}^6$8, Mean. $\{\mathcal{A}_i\}_{i=1}^6$9
• Percep360: BRISQUE $$\Omega = \left( \bigcup_{i=1}^6 \mathcal{R}_i \right) \cup \left( \bigcup_{i=1}^6 \mathcal{A}_i \right), \quad \mathcal{R}_i \cap \mathcal{A}_j = \emptyset,\quad \forall i,j.$$0, PIQE $$\Omega = \left( \bigcup_{i=1}^6 \mathcal{R}_i \right) \cup \left( \bigcup_{i=1}^6 \mathcal{A}_i \right), \quad \mathcal{R}_i \cap \mathcal{A}_j = \emptyset,\quad \forall i,j.$$1, SSIM $$\Omega = \left( \bigcup_{i=1}^6 \mathcal{R}_i \right) \cup \left( \bigcup_{i=1}^6 \mathcal{A}_i \right), \quad \mathcal{R}_i \cap \mathcal{A}_j = \emptyset,\quad \forall i,j.$$2, FID $$\Omega = \left( \bigcup_{i=1}^6 \mathcal{R}_i \right) \cup \left( \bigcup_{i=1}^6 \mathcal{A}_i \right), \quad \mathcal{R}_i \cap \mathcal{A}_j = \emptyset,\quad \forall i,j.$$3, Driv. $$\Omega = \left( \bigcup_{i=1}^6 \mathcal{R}_i \right) \cup \left( \bigcup_{i=1}^6 \mathcal{A}_i \right), \quad \mathcal{R}_i \cap \mathcal{A}_j = \emptyset,\quad \forall i,j.$$4, Mean. $$\Omega = \left( \bigcup_{i=1}^6 \mathcal{R}_i \right) \cup \left( \bigcup_{i=1}^6 \mathcal{A}_i \right), \quad \mathcal{R}_i \cap \mathcal{A}_j = \emptyset,\quad \forall i,j.$$5 (Teng et al., 9 Jul 2025)

The LSDM-specific ablation is more diagnostic. The paper compares Baseline, + Mask, + M.Crs, + Rota, and + LSDM F. The reported values are:

• Baseline: BRISQUE $$\Omega = \left( \bigcup_{i=1}^6 \mathcal{R}_i \right) \cup \left( \bigcup_{i=1}^6 \mathcal{A}_i \right), \quad \mathcal{R}_i \cap \mathcal{A}_j = \emptyset,\quad \forall i,j.$$6, PIQE $$\Omega = \left( \bigcup_{i=1}^6 \mathcal{R}_i \right) \cup \left( \bigcup_{i=1}^6 \mathcal{A}_i \right), \quad \mathcal{R}_i \cap \mathcal{A}_j = \emptyset,\quad \forall i,j.$$7, FID $$\Omega = \left( \bigcup_{i=1}^6 \mathcal{R}_i \right) \cup \left( \bigcup_{i=1}^6 \mathcal{A}_i \right), \quad \mathcal{R}_i \cap \mathcal{A}_j = \emptyset,\quad \forall i,j.$$8, Driv. $$\Omega = \left( \bigcup_{i=1}^6 \mathcal{R}_i \right) \cup \left( \bigcup_{i=1}^6 \mathcal{A}_i \right), \quad \mathcal{R}_i \cap \mathcal{A}_j = \emptyset,\quad \forall i,j.$$9, Mean. $$f(x) = \sum_{i=1}^6 \mathbb{I}_{\mathcal{R}_i}(x)\cdot f_i(x) + \sum_{i=1}^6 \mathbb{I}_{\mathcal{A}_i}(x)\cdot \tilde{f}_i(x),$$0
• + Mask: BRISQUE $$f(x) = \sum_{i=1}^6 \mathbb{I}_{\mathcal{R}_i}(x)\cdot f_i(x) + \sum_{i=1}^6 \mathbb{I}_{\mathcal{A}_i}(x)\cdot \tilde{f}_i(x),$$1, PIQE $$f(x) = \sum_{i=1}^6 \mathbb{I}_{\mathcal{R}_i}(x)\cdot f_i(x) + \sum_{i=1}^6 \mathbb{I}_{\mathcal{A}_i}(x)\cdot \tilde{f}_i(x),$$2, FID $$f(x) = \sum_{i=1}^6 \mathbb{I}_{\mathcal{R}_i}(x)\cdot f_i(x) + \sum_{i=1}^6 \mathbb{I}_{\mathcal{A}_i}(x)\cdot \tilde{f}_i(x),$$3, Driv. $$f(x) = \sum_{i=1}^6 \mathbb{I}_{\mathcal{R}_i}(x)\cdot f_i(x) + \sum_{i=1}^6 \mathbb{I}_{\mathcal{A}_i}(x)\cdot \tilde{f}_i(x),$$4, Mean. $$f(x) = \sum_{i=1}^6 \mathbb{I}_{\mathcal{R}_i}(x)\cdot f_i(x) + \sum_{i=1}^6 \mathbb{I}_{\mathcal{A}_i}(x)\cdot \tilde{f}_i(x),$$5
• + M.Crs: BRISQUE $$f(x) = \sum_{i=1}^6 \mathbb{I}_{\mathcal{R}_i}(x)\cdot f_i(x) + \sum_{i=1}^6 \mathbb{I}_{\mathcal{A}_i}(x)\cdot \tilde{f}_i(x),$$6, PIQE $$f(x) = \sum_{i=1}^6 \mathbb{I}_{\mathcal{R}_i}(x)\cdot f_i(x) + \sum_{i=1}^6 \mathbb{I}_{\mathcal{A}_i}(x)\cdot \tilde{f}_i(x),$$7, FID $$f(x) = \sum_{i=1}^6 \mathbb{I}_{\mathcal{R}_i}(x)\cdot f_i(x) + \sum_{i=1}^6 \mathbb{I}_{\mathcal{A}_i}(x)\cdot \tilde{f}_i(x),$$8, Driv. $$f(x) = \sum_{i=1}^6 \mathbb{I}_{\mathcal{R}_i}(x)\cdot f_i(x) + \sum_{i=1}^6 \mathbb{I}_{\mathcal{A}_i}(x)\cdot \tilde{f}_i(x),$$9, Mean. $f_i(x)$0
• + Rota: BRISQUE $f_i(x)$1, PIQE $f_i(x)$2, FID $f_i(x)$3, Driv. $f_i(x)$4, Mean. $f_i(x)$5
• + LSDM F: BRISQUE $f_i(x)$6, PIQE $f_i(x)$7, FID $f_i(x)$8, Driv. $f_i(x)$9, Mean. $\mathcal{R}_i$0 (Teng et al., 9 Jul 2025)

The authors’ interpretation is explicit: masking or cross-attention strategies can improve image-quality metrics but damage structure and controllability, while naive rotation also degrades quality and layout fidelity; LSDM “effectively achieves a balance between image quality and layout controllability” (Teng et al., 9 Jul 2025).

The downstream evidence is also notable. When Percep360-generated data are used to augment OneBEV, the reported BEV segmentation results are:

• w/o synthetic data: Driv. $\mathcal{R}_i$1, Mean. $\mathcal{R}_i$2
• w/ Baseline: Driv. $\mathcal{R}_i$3, Mean. $\mathcal{R}_i$4
• w/ Percep360: Driv. $\mathcal{R}_i$5, Mean. $\mathcal{R}_i$6 (Teng et al., 9 Jul 2025)

These results indicate that the generated panoramas are not only visually smoother than stitched supervision, but more useful for downstream panoramic BEV segmentation than baseline synthetic data. The paper explicitly notes a 2.5\% mIoU improvement from $\mathcal{R}_i$7 to $\mathcal{R}_i$8 (Teng et al., 9 Jul 2025).

6. Limitations and open questions

The paper is explicit that LSDM is a partial solution to panoramic generation from stitched data. It “focuses on image coherence and controllability of 2D features,” and identifies “label mapping from six-view images to panoramic representations” as a critical future direction (Teng et al., 9 Jul 2025). Thus, although LSDM addresses cyclic continuity in panorama space, it does not solve full annotation transfer or 3D-consistent panoramic supervision.

Several methodological constraints are also stated or implied. First, LSDM relies on the assumption that the panorama is horizontally periodic and that image-space rotation can be aligned with BEV-feature rotation. Second, it does not introduce an explicit seam-consistency loss; continuity is enforced implicitly through circular roll and aligned conditioning. Third, the ablation shows that “simple rotation alone is insufficient”: the benefit arises from synchronized reconfiguration of image-BEV correspondence, not from generic data augmentation (Teng et al., 9 Jul 2025).

A plausible implication is that LSDM is best understood as a geometry-aware training strategy for cyclic domains rather than a self-sufficient generative model family. In settings where the domain is not naturally periodic, or where control signals cannot be rotated consistently, the method would require reformulation.

7. Relation to adjacent local-scene diffusion research

Percep360’s LSDM belongs to a broader family of locality-aware diffusion ideas, but it is unusually specific in targeting stitched panoramic coherence. Closely related scene-generation work includes "Compositional 3D Scene Generation using Locally Conditioned Diffusion," which applies different prompts to different regions through locally composed denoising fields in a score distillation sampling loop (Po et al., 2023); "Gaussian Scenes: Pose-Free Sparse-View Scene Reconstruction using Depth-Enhanced Diffusion Priors," which uses a confidence-guided RGBD image-to-image diffusion prior inside a progressive Gaussian-SLAM-inspired process (Paul et al., 2024); and "Scaling Diffusion Models to Real-World 3D LiDAR Scene Completion," which reformulates point-cloud diffusion as a local point-wise offset process in metric scene space (Nunes et al., 2024).

Other adjacent formulations emphasize local editing or local structural control rather than panorama coherence. "LatentEditor: Text Driven Local Editing of 3D Scenes" uses latent-space masks and iterative dataset updates for localized NeRF editing (Khalid et al., 2023). "ScenDi: 3D-to-2D Scene Diffusion Cascades for Urban Generation" generates a bounded 3D foreground volume and refines it with 2D video diffusion, which places local urban scene structure explicitly in 3D (Guo et al., 21 Jan 2026). "Generating Driving Scenes with Diffusion" performs map-conditioned local traffic-scene generation in ego-centered BEV patches using latent diffusion plus a detection-style decoder (Pronovost et al., 2023). "LAW-Diffusion: Complex Scene Generation by Diffusion with Layouts" uses object region maps, location-aware cross-object attention, and latent grafting for local instance reconfiguration (Yang et al., 2023). "Enabling Local Editing in Diffusion Models by Joint and Individual Component Analysis" links regions of interest to bottleneck subspaces through Jacobians and JIVE, separating local from global semantics (Kouzelis et al., 2024). At the theoretical end, "Localized Diffusion Models for High Dimensional Distributions Generation" shows that, under locality structure, score estimation can be restricted to neighborhoods and still support global reverse diffusion (Gottwald et al., 7 May 2025).

Taken together, these works suggest that “locality” in diffusion has acquired several technically distinct meanings: region-wise prompt composition, local view repair, local offset denoising, patch-aligned layout control, ROI-conditioned latent editing, and graph-localized score estimation. Percep360’s LSDM is specific among them in one respect: it uses locality to counteract the discontinuous supervision induced by multi-camera panorama stitching, recasting a $\mathcal{R}_i$9 panorama as a cyclic domain on which denoising should be spatially continuous rather than seam-bound (Teng et al., 9 Jul 2025).