Insta360 Research debuts Canvas360 with 1M-sample panorama dataset

Insta360's research arm has co-authored a paper with collaborators from Tsinghua, Beihang and Wuhan University introducing Canvas360, a two-stage system for generating and editing 360-degree panoramas. The paper on Hugging Face describes a pretraining stage that produces RGB and depth in parallel, followed by fine-tuning that unifies style transfer, inpainting, outpainting and editing inside a single model.

The scale claim is the interesting part. The team assembled Canvas360Dataset, one million paired panoramic samples, with 100K RGB-depth pairs earmarked for the geometry-aware pretraining stage and 900K downstream examples split across four tasks: 200K for style transfer, 250K each for outpainting and inpainting, and 200K for editing. That is a large purpose-built corpus for a fairly niche modality, and it is being packaged as a contribution alongside the model.

Under the hood the base is FLUX.1-dev, adapted with LoRA. Two design choices carry most of the panorama-specific weight. Velocity circular padding enforces boundary continuity, since the left and right edges of an equirectangular image have to wrap around a sphere, and a similarity loss regularizer keeps the RGB and depth streams from collapsing into each other. On the panorama-specific FAED metric, Canvas360 reports 2.33 against DiT360's 2.91, and the authors' user study puts Canvas360 first on boundary continuity, panorama awareness and overall quality against DiT360, HunyuanWorld and Matrix-3D.

The honest caveat is that this is a preprint reporting the authors' own numbers, and FID is not a clean sweep — DiT360 edges Canvas360 on standard FID in the same table. What the paper does not give you is a licensing or release timeline for the dataset or weights, nor a compute footprint for training. Much of the downstream 900K is also model-generated rather than human-labeled, which is worth holding in mind when reading the fidelity gains.

Where this matters is the identity of the lead affiliation. A hardware-adjacent research lab publishing its own panorama foundation approach, unified editing framework and a million-sample corpus is a stronger signal about where the 360 imaging market is going than yet another academic-only benchmark paper. Whether Canvas360's weights and data actually ship is what determines whether VR and virtual-tour teams get to use this or only cite it.