Normal bridging
Uses 2.5D normal maps to reduce RGB ambiguities (lighting/texture vs shape) before 3D generation.
Stable-X / research collaborators · Open image-to-3D
Hi3DGen
High-fidelity 3D geometry via normal bridging
Hi3DGen generates high-fidelity 3D geometry from images by treating normal maps as an intermediate bridge between RGB and 3D. The approach is detailed in “Hi3DGen: High-fidelity 3D Geometry Generation from Images via Normal Bridging” (arXiv:2503.22236).
What is Hi3DGen?
Direct RGB-to-3D mapping confuses lighting, texture, and shape. Hi3DGen instead estimates a sharp normal map (surface orientation field), then generates geometry conditioned on that geometric signal.
The paper describes three pillars: (1) NiRNE—an image-to-normal estimator with noise injection and dual-stream training to separate low- and high-frequency patterns; (2) NoRLD—normal-regularized latent diffusion for geometry; (3) a data synthesis pipeline producing DetailVerse assets to complement human-created 3D data.
Authors are affiliated with The Chinese University of Hong Kong, Shenzhen; ByteDance; and Tsinghua University (project page). The public demo is hosted as Stable-X/Hi3DGen on Hugging Face Spaces with the short description “High-fidelity 3D Geometry Generation from single view image.”
When to use Hi3DGen
Pick Hi3DGen when geometric detail and surface orientation matter more than flashy baked textures—hard-surface props, characters with crisp silhouettes, and any workflow where you will retexture later.
How Hi3DGen works
STEP 1
Estimate normals from RGB
NiRNE predicts a normal map that encodes per-pixel surface orientation as a geometric intermediate.
STEP 2
Diffuse geometry with normal regularization
NoRLD generates 3D structure while staying consistent with the normal signal.
STEP 3
Review mesh fidelity
Inspect silhouettes and surface detail in the Gradio demo; export according to Space capabilities.
Key capabilities
NiRNE normal estimator
Noise-injected dual-stream training aims for generalizable, stable, sharp normal estimation from single images.
NoRLD geometry diffusion
Normal-regularized latent diffusion supplies explicit geometric supervision during training.
Open research demo
Stable-X/Hi3DGen Space plus paper/code references (ByteDance/Hi3DGen GitHub).
Technical snapshot
Figures below come from the linked papers, model cards, or project pages—not from unbenchmarked third-party marketing claims.
- Paper
- arXiv:2503.22236
- Core idea
- Image → normal (NiRNE) → geometry (NoRLD)
- Dataset note
- DetailVerse synthesis pipeline (paper)
- Demo
- huggingface.co/spaces/Stable-X/Hi3DGen
- Code reference
- github.com/bytedance/Hi3DGen
Who it is for
Geometry-first asset creation
When you need clean shape drafts before manual UV/texturing in Blender or Substance.
Detail-sensitive objects
Paper focus on high-frequency geometric detail versus blurry image-to-3D baselines.
Research comparison
Useful geometry-centric baseline next to TRELLIS.2 / Hunyuan texture-heavy pipelines.
Limitations & practical notes
- Primary emphasis is geometry fidelity; textured appearance quality depends on the demo build and may lag full PBR systems.
- Normal estimation can still fail on transparent, highly specular, or heavily occluded inputs.
- Space hardware and queues vary; results are not a substitute for measured production QA.
Frequently asked questions
What problem does Hi3DGen solve?+
It reduces the difficulty of mapping RGB images (with lighting and texture cues) directly to 3D by inserting a normal-map bridge, as argued in arXiv:2503.22236.
Is Hi3DGen only for single images?+
The official Space describes single-view image input. Multi-view variants are outside what we embed here unless the Space exposes them.
Sources & references
Landing claims are grounded in the following primary materials. Always prefer the original paper or model card for citations.
Try Hi3DGen in the browser
Jump into the playground and generate a 3D model from your own image—no install required.