deepopen-com/deepopen-3d-webs

A single-image-to-3D studio that packages LucidDreamer, ZoeDepth and a Gradio interface into a Docker app. It is worth a look if you want a guided demo of image-driven scene reconstruction, but treat it as a preview tool rather than a production asset pipeline.

The problem

Turning a still image into something viewers can look around in normally means stitching together depth estimation, image completion, Gaussian reconstruction and camera flythrough across several tools, then manually assembling the outputs for review. Guangyu Zaojing puts that chain in one interface: upload a reference image, choose or load example prompts and a camera trajectory, and press generate. The result is a color animation, a depth animation and a draggable Gaussian point cloud shown together, plus a downloadable PLY. The change is that one studio now carries the scene from input to a shareable review page instead of handing files between tools.

What's different

The core mechanism is the LucidDreamer pipeline: diffusion inpainting and ZoeDepth fill in and estimate depth around the reference view, then 3D Gaussians bake the scene into a PLY that can be rendered along a chosen camera path. The README separates two trajectory concepts, generation and render, which lets you decide how the space is observed before deciding how the final footage moves. A one-click mode runs preview, color animation and depth animation end to end, while two step buttons stop after the Gaussian bake or after the video render, which is more useful than an all-or-nothing button when you only want the point cloud.

Project overview and feature list

The trade-off is the same one inherent in this family of reconstruction methods: what you get is a plausible walkthrough, not a guaranteed faithful reconstruction, and the PLY files can reach hundreds of MB. Open the demo gallery before generating to judge whether the input-output quality matches what you need.

Who should try it

Good fit: a concept designer, animator or educator who already has one stylized image and wants a quick look-around version, a short camera flythrough, or a browser-based point cloud to show collaborators. The bundled demo assets are also convenient for a classroom or pitch, and demo mode runs without a GPU. Poor fit: anyone who needs an accurate, measurement-grade reconstruction or wants to feed the output straight into a DCC pipeline without manual cleanup; the README describes scene and camera presets and a PLY export but not a clean asset-authoring workflow. Users outside the Chinese UI should also confirm that the prompt workflow and example gallery suit their content.

The two Docker paths change who can use it: the default image runs the demo with no GPU, while the CUDA image is the documented path for actual generation with the downloaded diffusion and ZoeDepth weights (local deployment notes).

Requirements and limitations

Actual generation has the highest barrier among the documented paths. It needs an NVIDIA GPU (16 GB VRAM recommended), the NVIDIA Container Toolkit or a local CUDA setup, CUDA builds of PyTorch, and compiled simple-knn and depth-diff-gaussian-rasterization submodules; full mode downloads diffusion and ZoeDepth weights into a cache of roughly 30 GB. If any of those pieces is missing, the system falls back to demo mode, which is graceful but means you may think you are generating when you are not.

The demo path itself is lighter: Docker Engine 24+, Compose, port 8090 and about 8 GB of RAM, plus several extra GB of disk for the PLY demos. There is no declared license, which matters if you intend to use generated assets or the code commercially; check that before building a workflow around it. You will also need to keep the CUDA stack and model cache in working order, not just install it once.

HotRepo's take

HotRepo only has one recorded date: the 2026-10-27 NEW-REPOSITORY board, where it showed 200 total stars at #18 for Python. That is a young-repository board, not a daily-growth measurement, and with no comparable daily history, persistence cannot be judged. The immediate thing to verify is the license and the boundary between demo content and your own image-and-prompt generation.