Canvas360Dataset for Panoramic Generation

• Canvas360Dataset is a large-scale paired dataset with 1M high-quality panoramic samples designed specifically for in-context panoramic generation tasks.
• It uses a dual-stage framework combining 100K RGB–depth geometry-aware pretraining samples with 900K in-context samples for style transfer, inpainting, outpainting, and editing.
• The dataset incorporates rigorous synthesis, quality control, and evaluation metrics to bridge the gap in task-driven paired data for panoramic generative modeling.

to=arxiv_search.search ฝ่ายขายออนไลน์ ／久久json {"query":"Canvas360Dataset arXiv (Feng et al., 9 Jul 2026) Canvas360 panoramic generation", "max_results": 5, "sort_by": "submittedDate", "sort_order": "descending"} RTLUക്കാര to=arxiv_search.search code 天天送钱json {"query":"360-VSumm (Kontostathis et al., 2024) 360+x (Chen et al., 2024) 360-Indoor (Chou et al., 2019)", "max_results": 10, "sort_by": "relevance", "sort_order": "descending"} RTLR to=arxiv_search.search code 】【。】【”】【json {"query":"A Human-Annotated Video Dataset for Training and Evaluation of 360-Degree Video Summarization Methods (Kontostathis et al., 2024)", "max_results": 3, "sort_by": "relevance", "sort_order": "descending"} Canvas360Dataset is a large-scale paired panoramic dataset introduced as the data substrate of the Canvas360 framework for in-context panoramic generation. It is presented as a collection of 1M high-quality paired panoramic samples spanning geometry-aware pretraining and downstream task supervision, with the explicit goal of supporting style transfer, inpainting, outpainting, and editing in a unified panoramic setting (Feng et al., 9 Jul 2026). Within the accompanying framework, the dataset serves two distinct roles: a 100K RGB–depth panorama subset for geometry-aware pretraining, and a 900K paired in-context subset for downstream adaptation. The paper positions this design as a response to the lack of large-scale, task-driven paired data for panoramic generation, in contrast to prior panoramic datasets that primarily target perception, multimodal understanding, or embodied reasoning rather than paired generative supervision (Feng et al., 9 Jul 2026).

1. Name, scope, and task definition

Canvas360Dataset is the official dataset name used in the paper titled "Enhancing In-context Panoramic Generation via Geometric-aware Pretraining" (Feng et al., 9 Jul 2026). It is not described as a generic 360-degree benchmark; rather, it is explicitly constructed for in-context panoramic generation, where a model conditions on a panorama or masked panorama together with text and predicts a target panorama. The paper defines four downstream task families: style transfer, inpainting, outpainting, and editing, and treats them as a common class of in-context panoramic generation problems (Feng et al., 9 Jul 2026).

The dataset is embedded in a two-stage framework. In the first stage, the model is pretrained for text-to-panorama generation with geometric supervision using paired RGB and pseudo-depth panoramas. In the second stage, depth supervision is removed and the model is fine-tuned on paired in-context samples for the four downstream tasks. This organization makes the dataset both a pretraining resource and a task-specific supervision resource.

A central claim of the paper is that existing panoramic datasets are not designed around paired, task-level supervision for panoramic editing and completion. The dataset is therefore framed as infrastructure for a unified model that can handle heterogeneous panoramic tasks through a common input-output structure. A plausible implication is that the dataset’s design is less about broad scene coverage as an end in itself than about enabling a single generative interface across multiple panorama-specific tasks.

2. Corpus composition and scale

The dataset has two major components: a 100K pilot/pretraining set and a 900K downstream in-context set (Feng et al., 9 Jul 2026).

Component	Size	Role
RGB–depth pilot/pretraining set	100K	Geometry-aware pretraining
Downstream in-context set	900K	Fine-tuning on four tasks
Full dataset	1M	Combined resource

The downstream task distribution is given explicitly:

Task	Samples	Notes
Outpainting	250K	Perspective-to-panorama completion
Inpainting	250K	Masked panorama completion
Style transfer	200K	Panorama stylization
Editing	200K	Object-centric edits

The paper summarizes the full resource as “100K annotated RGB–depth panoramas and 900K in-context generation samples spanning four tasks” (Feng et al., 9 Jul 2026). It also states that the pilot set comprises “100K indoor and outdoor scenes.” No finer-grained taxonomy is reported for those scenes: there is no detailed indoor/outdoor percentage split, source-domain percentage table, category histogram, or geographic breakdown.

The paper does not describe a formal train/validation/test split for Canvas360Dataset itself. It states that, for in-context panoramic generation tasks, training and evaluation are conducted on the constructed dataset, but it does not specify held-out subset sizes, canonical split files, or benchmark server protocol. It likewise does not provide file organization details or per-task directory structure.

The paper’s main experiments are conducted at $1024 \times 2048$, with ablations at $512 \times 1024$. Because the work consistently refers to ERP panoramas, this suggests that dataset samples are represented in standard equirectangular format, although the paper does not compress that point into a single formal dataset-specification sentence (Feng et al., 9 Jul 2026).

3. Data sources and synthesis pipeline

The 100K pilot set is said to be sourced from Matterport3D, the web, and state-of-the-art panoramic generation models (Feng et al., 9 Jul 2026). For that subset, the authors generate full-panorama captions with VLMs and pair RGB panoramas with pseudo-depth maps predicted by DAP, which the paper describes as a state-of-the-art panoramic metric-depth estimator.

The paper reports a specific preprocessing step for pseudo-depth stabilization: depth values are clipped at 100 for outdoor scenes and 10 for indoor scenes, then normalized to a fixed range. It further states that comparison to available ground-truth depth shows only small discrepancy after processing, which is presented as evidence that the pseudo-depth is adequate for geometry-aware supervision (Feng et al., 9 Jul 2026).

The 900K downstream subset is synthesized from the pilot data through task-specific generation procedures:

• Style transfer: generated with FLUX.2-dev using style-specific prompts; Figure 1 states that stylization spans 12 styles. The paper does not enumerate those styles in text.
• Outpainting: built by sampling yaw, pitch, field of view, height, and width from predefined priors, projecting a centered perspective view after a sampled yaw shift, and inverse-projecting it into ERP space to obtain a visible-region mask.
• Inpainting: built from rectangular masks with random area ratios, aspect ratios, and locations, under both global-prompted and local-prompted regimes.
• Editing: built through object localization and removal pipelines based on Florence2, GroundingDino, FLUX.2, and NanoBanana, following SE360 for the overall construction procedure (Feng et al., 9 Jul 2026).

For local-prompted inpainting, masked-region captions are generated with Qwen3-VL-30B-A3B-Instruct. For editing, the authors emphasize object removal and addition. They first localize objects using VLM captioning and fused grounding, obtain bounding boxes and segmentation masks, remove targeted objects, and then invert original-edited pairs to obtain both erasure and addition samples (Feng et al., 9 Jul 2026). The paper presents this inversion strategy as a practical response to the weakness of planar image editing models on panorama-aware object addition.

A recurring design principle is that panoramic geometry should be preserved during data synthesis. For inpainting and outpainting, the targets are always derived from real panoramic images. For style transfer, structural consistency is checked by comparing extracted line drawings between original and stylized images, and samples with large structural inconsistencies are removed. For editing, hard cases involving small objects, fine structures, or complex backgrounds are refined with NanoBanana to preserve background consistency (Feng et al., 9 Jul 2026).

4. Sample representation and training role

The paper formalizes downstream supervision as a task-agnostic structure consisting of a context panorama latent, a text prompt, and a target panorama latent (Feng et al., 9 Jul 2026). In prose, the intended representation is:

• $\mathbf{x}_{\text{cxt}}$: post-VAE latent of the context panorama image
• $c$: text prompt
• $\mathbf{x}_{\text{tgt}}$: post-VAE latent of the target panorama image

This common structure is shared across all four downstream tasks. For style transfer and editing, the context is the full input panorama. For inpainting and outpainting, the context is a masked panorama latent. The target is always the desired full panorama. The paper further states that the model consumes target and context tokens through sequence concatenation, with a positive positional offset $T_c$ used to separate their semantic roles while preserving spatial alignment (Feng et al., 9 Jul 2026).

The pretraining subset is organized differently. RGB and depth panoramas are encoded into token sequences $\mathbf{x}_{\text{rgb}}$ and $\mathbf{x}_{\text{depth}}$, again concatenated with modality-specific positional offsets. The pretraining objective is standard flow matching with an added similarity regularizer. The paper gives the rectified-flow interpolation as

$x_t = (1 - t) x_0 + t x_1, \quad t \in [0,1].$

and the flow-matching loss as

$$\mathcal{L}(\theta) = \mathbb{E}_{t,\, x_0 \sim \pi_0,\, x_1 \sim \pi_1} \left[ \left\| v_\theta(x_t, t) - (x_1-x_0) \right\|^2 \right].$$

For RGB–depth pretraining, this loss is computed independently for each modality, and the overall objective is described as

$512 \times 1024$0

where $512 \times 1024$1 is explained as a squared cosine-correlation penalty between predicted RGB and depth velocity fields (Feng et al., 9 Jul 2026). This makes the first 100K examples structurally different from the downstream 900K examples: the former encode geometric correspondence, whereas the latter encode task-conditioned panoramic transformation.

5. Evaluation use and quantitative role

Canvas360Dataset is used not only for training but also for evaluating the four in-context tasks. The appendix states that 500 results are generated for each in-context experiment for evaluation (Feng et al., 9 Jul 2026). The reported metrics depend on task type:

• Style transfer: CP, SR, and OV, following SRQE
• Inpainting, outpainting, and editing: LPIPS, FAED, and PSNR

The paper emphasizes FAED as a panorama-specific metric and argues that it better reflects panoramic fidelity than perspective-biased FID (Feng et al., 9 Jul 2026). For the upstream text-to-panorama stage, evaluation is conducted on Matterport rather than Canvas360Dataset for fairness, using FID, FID$512 \times 1024$2, FID$512 \times 1024$3, FAED, IS, CLIP Score, Q-Align, BRISQUE, and NIQE.

The reported downstream comparisons place Canvas360Dataset at the center of the benchmark protocol. For example, on the dataset-backed evaluations:

• Inpainting: the paper reports Ours 0.096 / 0.371 / 25.87 for LPIPS / FAED / PSNR, against FLUX.1-Fill-dev 0.171 / 0.461 / 24.23, PAR 0.158 / 0.455 / 23.76, and PanoDiffusion 0.147 / 0.523 / 23.93 (Feng et al., 9 Jul 2026).
• Outpainting: Ours 0.416 / 1.791 / 17.16, compared with FLUX.1-Fill-dev 0.509 / 1.916 / 16.32, PAR 0.553 / 1.849 / 16.71, and PanoDiffusion 0.674 / 1.989 / 15.21.
• Editing: Ours 0.084 / 0.358 / 26.40, compared with FLUX.1-Kontext-dev 0.102 / 0.458 / 25.77, FLUX.2-dev 0.099 / 0.410 / 26.17, NanoBanana 0.094 / 0.395 / 25.91, SE360 0.138 / 0.386 / 25.16, and Omni2 0.105 / 0.392 / 25.03.
• Style transfer: the paper reports CP 0.502, SR 0.491, OV 0.497 for the proposed method against FLUX.1-Kontext-dev, FLUX.2-dev, and Qwen-Image-Edit.

These numbers are presented as evidence that the dataset is sufficiently structured to support cross-task benchmarking under a unified generative setting.

6. Quality control, limitations, and relation to other 360-degree datasets

The paper devotes unusual attention to synthesis noise and quality control. A manual inspection of a 50K subset found 48K clean and valid samples, implying roughly 96% acceptance in that inspected subset (Feng et al., 9 Jul 2026). A separate cleaning ablation on 20K editing samples reports similar performance for randomly selected versus manually cleaned data: random samples give LPIPS 0.096, FAED 0.396, PSNR 25.34, while cleaned samples give LPIPS 0.093, FAED 0.391, PSNR 25.70. The paper interprets this as indicating that residual noise has limited effect on final performance.

Several limitations are also acknowledged. The training corpus is described as imbalanced across scene types, and the paper notes that high-quality panoramic data remain scarce for some categories. It reports underperformance on high-resolution human faces and text-rich signage, especially in strongly distorted ERP regions (Feng et al., 9 Jul 2026). The paper also states that the release will include documentation on data sources, preparation procedures, filtering steps, and intended-use restrictions, but it does not provide concrete license terms, repository structure, or direct dataset download instructions in the supplied text. The project page is given as https://zry000.github.io/Canvas360/.

Canvas360Dataset should be distinguished from several unrelated 360-degree datasets that address different tasks. 360-VSumm is a human-annotated dataset for transforming 360-degree videos into concise 2D-video summaries and is explicitly not named Canvas360Dataset (Kontostathis et al., 2024). 360-Indoor is an equirectangular indoor object-detection benchmark with BFoV annotations, not a generative paired dataset (Chou et al., 2019). 360+x is a panoptic multi-modal scene-understanding dataset centered on synchronized panoramic, egocentric, and audio modalities rather than in-context image generation (Chen et al., 2024). This distinction is important because the term “360 dataset” has been used across several neighboring subfields, including summarization, object detection, multimodal scene understanding, viewport analysis, and VQA, whereas Canvas360Dataset is specifically generation-oriented.

In that landscape, the defining features of Canvas360Dataset are its paired supervision, its four-task in-context formulation, and its coupling of RGB–depth geometric pretraining with downstream panorama editing and completion (Feng et al., 9 Jul 2026). This suggests that its primary historical significance lies not in being a general-purpose panoramic corpus, but in formalizing panoramic generation as a unified, multi-task, data-centric problem.