Hi3D Live 3D Edits

Turning one flat image into a usable 3D object usually takes several steps. Hi3D is now showing a live editing workflow that starts from one picture and updates the 3D result as the image is edited. (x.com) The underlying idea is image-to-3D: a model predicts how an object should look from other angles, then rebuilds shape and texture from those views. In its 2024 paper, the Hi3D team described a two-stage system that first generates multi-view images from one input image and then reconstructs a mesh from those views. (arxiv.org) That paper says Hi3D uses a video diffusion model with camera-pose conditioning, then a second refiner to sharpen texture detail. The authors reported 1,024-by-1,024 multi-view image generation before reconstruction, aiming to keep different angles consistent with each other. (arxiv.org) The new demo pushes that research toward editing instead of one-shot generation. In the clip posted on X, Hi3D shows background removal, image enhancement, and text-guided edits that propagate across multiple views instead of changing only a single frame. (x.com) That multi-view step is the hard part in 3D work. If an edit changes only the front view of an object, the side and back can break, which is why artists often rebuild assets by hand for games, augmented reality, and visual effects pipelines. (arxiv.org) Hi3D’s earlier paper framed the problem the same way: many image-to-3D systems can make a plausible single view but struggle with cross-view geometry and fine texture detail. The Hi3D code repository for the ACM Multimedia 2024 paper is public on GitHub, which suggests the company is building on a research base that was already disclosed. (github.com) The product pitch in the new demo is speed. Hitem3D’s public web app already offers “Image to 3D,” background removal, and image enhancement controls, and the live-edit demo suggests those preprocessing steps are being folded into a more interactive asset workflow. (hitem3d.ai) That puts Hi3D in a crowded image-to-3D market where companies are trying to move from static generation to production tools. Krea, for example, markets image, video, and 3D editing in one suite, showing how vendors are competing on workflow integration as much as raw model quality. (krea.ai) What Hi3D has shown so far is a demo, not a benchmarked product release with public latency, pricing, or export specs. But the direction is clear in the footage: fewer handoffs between image cleanup, view synthesis, and 3D asset creation from a single source image. (x.com)