Canvas360: Unified Panoramic Generation

• Canvas360 is a unified two-stage framework that leverages geometry-aware RGB–depth pretraining and task-specific fine-tuning for full 360° equirectangular panoramas.
• It supports diverse tasks such as style transfer, inpainting, outpainting, and editing while ensuring horizontal wrap-around continuity and correction of ERP distortions.
• Empirical results show improved geometric consistency and seam continuity, validated by metrics like FAED, IS, and NIQE across a tailored 1M-sample panoramic dataset.

Canvas360 is a two-stage framework for in-context panoramic generation that combines geometry-aware pretraining with downstream task-specific fine-tuning for 360° equirectangular panoramas. It is designed to extend text-to-panorama generation into a unified model that also supports style transfer, inpainting, outpainting, and editing, with particular emphasis on panoramic geometry: horizontal wrap-around continuity at the left/right boundary and latitude-dependent distortion under equirectangular projection (ERP). Its central claim is that strong panoramic priors can be learned during RGB–depth pretraining and then transferred to a single downstream model via token-level concatenation, yielding improved geometric consistency, seam continuity, and broad task coverage (Feng et al., 9 Jul 2026).

1. Problem setting and conceptual scope

Canvas360 addresses in-context panoramic generation rather than generic image editing. The target domain is full 360° ERP imagery, where the left and right boundaries are adjacent on the sphere and objects near the poles are stretched and curved in ways that do not arise in ordinary perspective-image datasets. The framework therefore treats panoramas as spherical scenes represented in ERP, not merely as unusually wide images. This distinction motivates the model’s geometry-aware design: it aims to capture wrap-around continuity, object distortion details, and radial scene structure suitable for panoramic generation (Feng et al., 9 Jul 2026).

The framework is organized into two stages. The first stage, geometry-aware text-to-panorama pretraining, jointly generates RGB panoramas and depth maps from RGB–depth panorama pairs. The second stage, unified downstream in-context finetuning, adapts the pretrained panorama backbone to a single model supporting four tasks—style transfer, inpainting, outpainting, and editing—using only RGB context inputs plus text. Depth is removed during downstream finetuning; the intended mechanism is that geometry-aware priors have already been internalized during pretraining and can then support flexible in-context generation without explicit geometric supervision at deployment (Feng et al., 9 Jul 2026).

Within the broader panoramic literature, Canvas360 occupies a specific methodological position. Earlier perspective-to-360 generation systems such as Argus formulate panoramic completion by explicitly projecting perspective video into ERP coordinates and aligning frames with estimated camera parameters (Luo et al., 10 Apr 2025). By contrast, 360Anything eliminates explicit camera information and learns perspective-to-ERP mapping through token-sequence conditioning in pre-trained diffusion transformers (Wu et al., 22 Jan 2026). Canvas360 differs from both because it is not primarily a perspective-to-panorama lifting method; it is a unified panoramic in-context generation framework built on panoramic pretraining and downstream panoramic editing tasks. This suggests its primary contribution lies in representation learning and task unification within the panoramic domain, rather than in recovering panoramic geometry from perspective conditioning alone.

2. Canvas360Dataset and supervision design

A central enabling component is Canvas360Dataset, described as a 1M-sample dataset tailored specifically to in-context panoramic generation (Feng et al., 9 Jul 2026). It is divided into a 100K pilot/pretraining subset of high-quality RGB panoramas paired with depth and text annotations, and a 900K paired in-context subset spanning four downstream tasks.

Component	Size	Role
RGB–depth pretraining subset	100K	Geometry-aware pretraining
Outpainting pairs	250K	Downstream supervision
Inpainting pairs	250K	Downstream supervision
Style transfer pairs	200K	Downstream supervision
Editing pairs	200K	Downstream supervision

The 100K pretraining subset is assembled from Matterport3D, the web, and state-of-the-art panoramic generation models, with panorama-level captions produced using vision-LLMs. The remaining 900K paired samples are synthesized for task-specific supervision. For style transfer, each example contains an original panorama and a stylized target panorama produced with FLUX.2-dev under style-specific prompts; the paper mentions 12 styles. For outpainting, the pipeline samples random yaw, pitch, field of view, image height, and width to define perspective views inside a panorama, then inverse-projects the visible perspective region back to ERP to form masks. For inpainting, it generates random rectangular masks with random area ratios, aspect ratios, and locations, and supports both global prompts describing the whole scene and local prompts describing the masked region, with local prompts generated using Qwen3-VL-30B-A3B-Instruct. For editing, VLM captioning, Florence2, and GroundingDINO are used to ground objects with boxes and segmentation masks, FLUX.2 removes targeted objects, and NanoBanana refines difficult cases; original/edited pairs are inverted to obtain both object erasure and object addition samples (Feng et al., 9 Jul 2026).

The dataset-construction process emphasizes geometric fidelity because standard 2D editors are not trained for ERP geometry. Inpainting and outpainting targets always come from real panoramas, editing relies on grounded erasure plus pair inversion rather than unconstrained panoramic addition, and style-transfer samples are filtered by comparing line drawings between source and stylized images to remove cases with large structural inconsistency. The authors manually inspect a 50K subset and report that 48K are clean/valid, and they further report little difference in editing performance between cleaned and randomly selected subsets, indicating that the remaining noise does not strongly harm final results (Feng et al., 9 Jul 2026).

This dataset design contrasts with earlier panoramic infrastructure. The object-centric benchmark 360-Indoor provides 3,335 real-world indoor ERP images with 89,148 BFoV object annotations for panoramic recognition and detection (Chou et al., 2019), whereas Canvas360Dataset is constructed for generative supervision across editing and completion tasks. The two datasets therefore address different layers of the panoramic stack: one dense recognition benchmark and one large-scale in-context generative corpus.

3. Geometry-aware pretraining

Canvas360 is built on FLUX.1-dev, framed as a flow-matching latent transformer fine-tuned with LoRA (Feng et al., 9 Jul 2026). The base formulation follows Rectified Flow interpolation:

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

with velocity target

$v = x_1 - x_0,$

and objective

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

This flow-matching objective is retained throughout the framework (Feng et al., 9 Jul 2026).

The defining addition in the pretraining stage is parallel depth generation. The model jointly generates RGB panoramas and depth panoramas from large-scale RGB–depth panorama pairs. The paper argues that depth in panoramas should be understood as radial distance from the camera center, not as perspective depth along the Cartesian $Z$-axis. To inject that prior, pseudo-depth is estimated using DAP (Depth Any Panoramas), then RGB and depth are encoded by the VAE into token sequences and concatenated along the token dimension. Rather than channel-wise fusion, token-wise concatenation is used so the design remains compatible with pretrained FLUX-style models (Feng et al., 9 Jul 2026).

To prevent RGB and depth tokens from occupying indistinguishable positions, Canvas360 introduces modality offsets in 3D RoPE coordinates. If $\mathbf{u}=(T,H,W)$ denotes positional coordinates, RGB and depth use

$$\mathbf{u}_0 = (0,H,W), \qquad \mathbf{u}_1 = (T_d,H,W), \quad T_d>0,\; T_d\in\mathbb{N}.$$

The two modalities therefore share spatial correspondence in $(H,W)$ but are separated along the embedding’s time/modality axis. According to the paper, this improves stability over direct RGB–depth co-generation and preserves compatibility with FLUX.1-Kontext-like concatenation mechanisms (Feng et al., 9 Jul 2026).

A second pretraining component is velocity circular padding, introduced to handle ERP seam consistency directly in the flow-matching target. After reshaping $x_t$ into $H \times W \times d$, the method appends two ghost longitude columns and synchronizes them with their circular counterparts:

$$\tilde{x}_t^{0} = x_t^{W}, \qquad \tilde{x}_t^{j} = x_t^{j},\; j=1,\ldots,W, \qquad \tilde{x}_t^{W+1} = x_t^{1},$$

with the same synchronization applied to target velocities:

$v = x_1 - x_0,$0

The loss is then computed on this padded velocity field. The intended effect is that seam-adjacent longitudes become local neighbors in the transformer input and are supervised to produce boundary-consistent velocities (Feng et al., 9 Jul 2026).

A third component is similarity loss regularization. During RGB–depth joint training, the authors observe a degenerate local optimum in which RGB and depth predictions become excessively similar, producing dark outputs or artifacts. They therefore regularize the squared cosine correlation between predicted RGB and depth velocity fields:

$v = x_1 - x_0,$1

and optimize

$v = x_1 - x_0,$2

This regularizer acts on velocity predictions rather than on pixels or latent tokens, and its reported role is to prevent modality collapse and stabilize optimization (Feng et al., 9 Jul 2026).

For reproducibility, the paper specifies that DAP depth values are clipped before normalization—at 100 for outdoor scenes and 10 for indoor scenes—because very large values in distant regions destabilize training. This preprocessing choice is presented as sufficient for training despite residual discrepancies from available ground-truth depth (Feng et al., 9 Jul 2026).

4. Unified in-context finetuning and model interface

After geometry-aware pretraining, Canvas360 performs unified downstream finetuning without depth inputs. All downstream tasks are cast into a common training tuple containing a context panorama latent, a text prompt, and a target panorama latent. For style transfer and editing, the context is the full input panorama; for inpainting and outpainting, it is a masked panorama latent with the missing region removed. The paper’s central claim is that heterogeneous tasks can therefore be reduced to a shared conditioning interface rather than requiring separate task-specific architectures (Feng et al., 9 Jul 2026).

The in-context mechanism uses token-level concatenation in the style of FLUX.1-Kontext. Target tokens and context tokens are concatenated into one transformer sequence,

$v = x_1 - x_0,$3

and separated through positional offsets:

$v = x_1 - x_0,$4

Thus the model jointly attends over target and context while preserving token-role identity through coordinate offsets. The result is a unified in-context panoramic generation framework that supports style transfer, inpainting, outpainting, and editing with one shared backbone and no separate task-specific architecture (Feng et al., 9 Jul 2026).

The implementation details are concrete. Fine-tuning is parameter-efficient via LoRA inserted into attention blocks and in/out embedding layers, with rank $v = x_1 - x_0,$5, scaling factor $v = x_1 - x_0,$6, and LoRA dropout $v = x_1 - x_0,$7. Training uses FP16 mixed precision on 8 NVIDIA H20 GPUs. Only LoRA parameters are optimized, with AdamW, learning rate $v = x_1 - x_0,$8, $v = x_1 - x_0,$9, $$\mathcal{L}(\theta) = \mathbb{E}_{t,\, x_0 \sim \pi_0,\, x_1 \sim \pi_1} \left[ \left\| v_\theta(x_t, t) - (x_1-x_0) \right\|^2 \right].$$0, $$\mathcal{L}(\theta) = \mathbb{E}_{t,\, x_0 \sim \pi_0,\, x_1 \sim \pi_1} \left[ \left\| v_\theta(x_t, t) - (x_1-x_0) \right\|^2 \right].$$1, and weight decay $$\mathcal{L}(\theta) = \mathbb{E}_{t,\, x_0 \sim \pi_0,\, x_1 \sim \pi_1} \left[ \left\| v_\theta(x_t, t) - (x_1-x_0) \right\|^2 \right].$$2. Training runs for 25 epochs with per-GPU batch size 1 and gradient accumulation 3, giving effective batch size 24. The schedule uses a constant learning rate with 10% warmup by steps, and training guidance scale is 1.0. Main experiments are run at $$\mathcal{L}(\theta) = \mathbb{E}_{t,\, x_0 \sim \pi_0,\, x_1 \sim \pi_1} \left[ \left\| v_\theta(x_t, t) - (x_1-x_0) \right\|^2 \right].$$3, while ablations use $$\mathcal{L}(\theta) = \mathbb{E}_{t,\, x_0 \sim \pi_0,\, x_1 \sim \pi_1} \left[ \left\| v_\theta(x_t, t) - (x_1-x_0) \right\|^2 \right].$$4. Inference uses classifier-free guidance scale 3.0 and 28 sampling steps (Feng et al., 9 Jul 2026).

The design is related to, but distinct from, other panoramic generation interfaces. 360Anything also uses token-sequence conditioning with pre-trained diffusion transformers, but it treats the perspective input and ERP target as separate token streams and eliminates explicit camera information (Wu et al., 22 Jan 2026). Canvas360 instead assumes panoramic context inputs at downstream time and uses geometry-aware pretraining to embed panoramic priors before in-context finetuning. This suggests that Canvas360’s unification is primarily task-unification within the panorama domain, rather than view-lifting from perspective media.

5. Empirical results and ablation evidence

Canvas360 is evaluated on five tasks: text-to-panorama generation, style transfer, inpainting, outpainting, and editing (Feng et al., 9 Jul 2026). For text-to-panorama evaluation, the reported baselines include PanFusion, SMGD, PAR, WorldGen, LayerPano3D, HunyuanWorld, and DiT360. For in-context tasks, comparisons include FLUX.1-Fill-dev, FLUX.1-Kontext-dev, FLUX.2-dev, NanoBanana, SE360, Omni2, PAR, and PanoDiffusion, depending on the task.

The metric suite is panorama-specific as well as conventional. In addition to FID, the paper reports $$\mathcal{L}(\theta) = \mathbb{E}_{t,\, x_0 \sim \pi_0,\, x_1 \sim \pi_1} \left[ \left\| v_\theta(x_t, t) - (x_1-x_0) \right\|^2 \right].$$5 and $$\mathcal{L}(\theta) = \mathbb{E}_{t,\, x_0 \sim \pi_0,\, x_1 \sim \pi_1} \left[ \left\| v_\theta(x_t, t) - (x_1-x_0) \right\|^2 \right].$$6, FAED, Inception Score computed with a Places365-pretrained ResNet, CLIP Score, Q-Align quality and aesthetics, BRISQUE, NIQE, and appendix-only LRCE-RGB for left-right boundary consistency. This metric choice reflects the paper’s argument that FID alone is not well matched to omnidirectional fidelity because it relies on an Inception network trained on perspective images (Feng et al., 9 Jul 2026).

The main text-to-panorama result is that Canvas360 achieves the best FAED at 2.33, improving over DiT360 and HunyuanWorld, both at 2.91. It also reports the best IS at 1.76, best QA aesthetic at 4.20, best NIQE at 3.70, and second-best results on FID, $$\mathcal{L}(\theta) = \mathbb{E}_{t,\, x_0 \sim \pi_0,\, x_1 \sim \pi_1} \left[ \left\| v_\theta(x_t, t) - (x_1-x_0) \right\|^2 \right].$$7, $$\mathcal{L}(\theta) = \mathbb{E}_{t,\, x_0 \sim \pi_0,\, x_1 \sim \pi_1} \left[ \left\| v_\theta(x_t, t) - (x_1-x_0) \right\|^2 \right].$$8, QA quality, and BRISQUE; specifically, FID is 44.17 versus DiT360’s 42.88. In the appendix, LRCE-RGB is best for Canvas360 at 0.0063, supporting improved seam consistency (Feng et al., 9 Jul 2026).

For downstream tasks, the paper reports strong quantitative results across the board. In style transfer, Canvas360 achieves the best content preservation (CP) at 0.502 and the best overall vision (OV) at 0.497, while FLUX.2-dev is slightly better on style resemblance (SR) at 0.503 versus 0.491. In inpainting, Canvas360 reports LPIPS 0.096, FAED 0.371, and PSNR 25.87. In outpainting, it reports LPIPS 0.416, FAED 1.791, and PSNR 17.16. In editing, against FLUX.1-Kontext-dev, FLUX.2-dev, NanoBanana, SE360, and Omni2, it reports LPIPS 0.084, FAED 0.358, and PSNR 26.40 (Feng et al., 9 Jul 2026).

The user study also emphasizes panoramic consistency. Across 71 participants and 10 images, Canvas360 ranks first in boundary continuity (31.1%), panorama awareness (34.1%), and overall quality (35.9%), while remaining near-best in text alignment at 26.2%, slightly below HunyuanWorld’s 26.9% (Feng et al., 9 Jul 2026).

The ablation studies clarify the contribution of each pretraining component. Starting from a FLUX.1-dev baseline, FAED is 5.37. Adding depth generation improves FAED to 4.74 and boosts IS and CLIP Score, but worsens BRISQUE and can produce dark or artifact-prone outputs. Adding positional offsets improves aesthetic quality, while adding $$\mathcal{L}(\theta) = \mathbb{E}_{t,\, x_0 \sim \pi_0,\, x_1 \sim \pi_1} \left[ \left\| v_\theta(x_t, t) - (x_1-x_0) \right\|^2 \right].$$9 yields the best QA quality and aesthetics within that family and stabilizes outputs. Velocity circular padding improves BRISQUE and NIQE and slightly improves QA quality, though not necessarily FID or FAED; the paper interprets this as evidence that seam-specific operations target local boundary artifacts rather than the global feature statistics captured by FID-like metrics (Feng et al., 9 Jul 2026).

6. Position within panoramic research, applications, and limitations

Canvas360 belongs to a rapidly expanding panoramic systems landscape, but its scope is specific. Argus studies video-to-360° generation from perspective video using explicit projection into a shared ERP canvas and camera-pose-based frame alignment (Luo et al., 10 Apr 2025). 360Anything studies metadata-free lifting of perspective images and videos to 360° panoramas through geometry-free sequence concatenation and Circular Latent Encoding (Wu et al., 22 Jan 2026). CubeComposer studies native 4K perspective-to-360° video generation through cubemap decomposition, spatio-temporal autoregression, and cube-aware seam handling (Li et al., 4 Mar 2026). Disentangled360 studies single-image 360° view generation through a Gaussian Splatting backbone with isotropic and anisotropic scene embeddings (KV et al., 11 Dec 2025). 360CityGML integrates CityGML with 360° walkthrough videos for interactive urban visualization (Banno et al., 16 Oct 2025), while the MPEG-OMAF browser player work shows how viewport-dependent tiled 360 playback can be implemented with HTML5 MSE and WebGL despite extractor-track limitations (Podborski et al., 2019). Relative to these systems, Canvas360 is best characterized as a unified panoramic content-generation framework that uses pretraining to transfer panoramic priors into editing and completion tasks.

Its reported application domain follows directly from those capabilities: style transfer, inpainting, outpainting, editing, and text-to-panorama generation in ERP space (Feng et al., 9 Jul 2026). A plausible implication is that it is also relevant to immersive media authoring pipelines in which panoramic assets must remain geometrically coherent under seam wrapping and ERP distortion.

The paper is explicit about limitations. Canvas360Dataset remains imbalanced across scene types, and the model underperforms on underrepresented cases such as high-resolution human faces and text-rich signage, particularly in heavily distorted ERP regions. The authors also note that some scene categories still lack enough high-quality panoramic data and that future work should rebalance and expand the corpus (Feng et al., 9 Jul 2026). More generally, the system operates in the panoramic image domain rather than in explicit 3D scene representations. This distinguishes it from scene-understanding benchmarks such as 360-Indoor, which provide BFoV-annotated object instances in real ERP imagery for detection and recognition (Chou et al., 2019).

Taken together, the reported evidence presents Canvas360 as a geometry-aware panoramic generation system whose main technical identity lies in three linked design choices: RGB–depth pretraining for spherical structure, velocity-level seam supervision for ERP continuity, and a unified token-concatenation interface for multiple downstream tasks. The framework does not claim to solve all panoramic generation problems, but it defines a concrete route for transferring panorama-specific priors into a single in-context model with broad task coverage and strong performance on panorama-specific fidelity metrics (Feng et al., 9 Jul 2026).