Most AI image-to-3D models start from pixels and guess the geometry. Hi3D V1.0 works the other way around: it takes a rough 3D mesh you already have, looks at a reference photo of the same object, and refines that mesh into high-resolution geometry. On October 23, 2026, the team behind it announced the model's open release, with inference code on GitHub (hi3d-official/Hi3Dv1.0, GitHub, retrieved 2026-10-24) and weights on Hugging Face under the same name (PR Newswire announcement, published October 23, 2026, accessed 2026-10-24). The detail that matters most is the license. The Hi3D Open License Agreement 1.0 permits commercial use, but products passing 200,000 monthly active users or USD 1,000,000 in monthly revenue need a separate written license. It is not an OSI-approved open source license, and that distinction shapes who this release is for.
Key Takeaways
Hi3D V1.0 is an image-conditioned 3D geometry refinement model released openly on October 23, 2026. It refines a coarse mesh you supply, guided by a reference image, rather than generating geometry from nothing.
The vendor reports training on about two million curated 3D assets, refinement up to a 1536³ sparse extraction grid, the highest multi-view fidelity scores in its reported evaluation, and a 90% slicing success rate (PR Newswire announcement). These are vendor-reported figures.
It ships under the custom Hi3D Open License Agreement 1.0, not an OSI-approved license. Commercial use is allowed, but products crossing 200,000 monthly active users or USD 1,000,000 in monthly revenue must apply for a separate license within 60 days, and commercial integrations carry "Powered by Hi3D 1.0" branding.
The release is inference only: no training code, no evaluation harness, no benchmark splits. The README states the repository alone cannot reproduce the paper's comparative tables.
The reference build needs Linux, Python 3.10, an NVIDIA GPU with compute capability 8.0 or newer, and a CUDA toolkit from 12.0 to 12.5. There is no CPU inference path, and the top setting was validated on a 96 GB NVIDIA H20. A community MLX port runs on Apple silicon.
What Is Hi3D V1.0?
Hi3D V1.0 is an openly released image-conditioned 3D geometry refinement model. Geometry refinement is the task of taking an existing rough mesh and improving its surface detail, as opposed to generating a mesh from scratch. You give the model two inputs, a reference image and a coarse mesh of the same object, and it returns a higher-resolution mesh. The model builds on the Sparc3D research work: it progressively samples sparse latents, decodes them with a self-pruning sparse VAE, and extracts a mesh, according to the project README (hi3d-official/Hi3Dv1.0, GitHub, retrieved 2026-10-24). The V1.0 lineage, previously called Sparc3D, gained recognition in 2025 for high-resolution geometry modeling. Its training set grew from about 500,000 to two million curated 3D assets before this release (PR Newswire announcement).
Refinement runs in up to three passes: 512 runs one pass, 1024 runs two, and 1536 runs all three, reaching an effective sparse extraction grid of 1536³. Each pass uses 25 scheduler steps by default. The README cautions that these labels describe nominal sparse coordinate and extraction grids, not a dense 1536³ tensor or a fixed triangle count.
Aspect | Hi3D V1.0 |
|---|---|
Announced | October 23, 2026, Beijing (PR Newswire) |
Model type | Image-conditioned 3D geometry refinement |
Training data | About 2 million curated 3D assets, up from about 500,000 |
Max resolution | 1536³ sparse extraction grid |
Output | Geometry only: mesh, no textures or materials |
License | Hi3D Open License Agreement 1.0 (custom) |
Code | Inference code and weights; no training code |
The repository is small and new. It was created on October 16, 2026 and, at the time of writing, holds 9 commits, one release, 185 stars, and 15 forks (hi3d-official/Hi3Dv1.0, GitHub, retrieved 2026-10-24). The eight model artifact files live on Hugging Face in safetensors and JSON format, and a live demo runs on Hugging Face Spaces. The conditioning encoders, DINOv2-large and CLIP ViT-L/14, load separately from Hugging Face on first use.
The License Is the Fine Print
Calling this an open source release is accurate in the everyday sense: the code and weights are published and free to download. It is not accurate in the strict sense. Hi3D V1.0 ships under the Hi3D Open License Agreement Version 1.0, which is not an OSI-approved open source license. Independent coverage of the release makes exactly this point, and the README states the terms plainly (SelfHostPilot analysis, published October 23, 2026; hi3d-official/Hi3Dv1.0 README, GitHub, retrieved 2026-10-24). Our what-is-open-source explainer covers where that line is drawn.
Commercial use is permitted.
If a related product or service exceeds 200,000 monthly active users or reaches USD 1,000,000 in monthly revenue, you must notify the licensor and apply for a separate written license within 60 days.
Commercial integrations must display "Powered by Hi3D 1.0" with a link to the official site where supported.
Standalone outputs that contain no Hi3D materials or model derivatives do not by themselves trigger the branding or scale obligations.
Apache-2.0 and MIT terms continue to cover the original and third-party portions they apply to.
For a small operator this is usually fine: a print shop, a small studio, or an internal tool is nowhere near 200,000 monthly users, and a branding line is a small cost. The difference shows up at scale. Cross a threshold and you are on a 60-day clock to apply for a written license whose terms you cannot see today. If your business case only works at scale, answer the threshold question before you build. Read the Agreement itself before relying on any summary, including this one (SelfHostPilot, October 23, 2026). For license inquiries, the README says to contact contact@hi3d.ai.
What the Release Does Not Include
The release is inference only. Training code, the evaluation harness, benchmark splits, and comparison meshes are not included. The README says plainly that the repository alone cannot reproduce the paper's comparative tables. That matters because the headline numbers, the highest multi-view fidelity and detail richness scores in the reported evaluation, and the 90% slicing success rate, come from the vendor. Downloadable weights let you run the model; they do not let you verify the claims that made you interested, as independent coverage notes (SelfHostPilot).
The output is geometry only: no textures, materials, or vertex colors. It runs at batch size one. Output meshes are recentered and normalized to unit maximum radius, so input world coordinates and absolute scale are not preserved. And because the model refines a coarse mesh, severe structural errors in that input can persist, such as inverted normals or disconnected parts. A broken input is likely to produce a detailed, broken output. This is one step in a pipeline, not a one-click asset machine: you still need something to produce the coarse mesh, and something after it for texturing, cleanup, or slicing. Slicing is the step that turns a finished mesh into layer-by-layer printer instructions.
The Hardware Reality
The reference build is demanding and specific. The README requires Linux, Python 3.10, an NVIDIA GPU with compute capability 8.0 or newer, a CUDA toolkit in the 12.0 to 12.5 range, and flash-attn, which has no fallback attention path. There is no CPU inference path. The measured stack was validated on a 96 GB NVIDIA H20, and the README states that peak memory and feasibility on smaller cards have not been established. Nobody is promising that a 24 GB consumer card can run the 1536 setting (hi3d-official/Hi3Dv1.0 README, GitHub, retrieved 2026-10-24).
There is a community alternative: on the same day as the release, an MLX port appeared as mlx-community/Hi3Dv1.0-mlx for Apple silicon. It is inference only, installs from source, and is a community project rather than the vendor's reference build. For Mac users with large unified memory, it is the most practical way to try the model without an NVIDIA card (SelfHostPilot).
What Independent Coverage Says
Independent coverage has been analytical rather than celebratory. SelfHostPilot's writeup treats Hi3D as a case study in reading an open source AI announcement, with four checks: the license by name, reproducibility, hardware, and scope. Its conclusion is a routine worth stealing: run those four checks before you clone anything. If all four answers fit, you have a tool to build on; if not, you have a demo to look at. The same piece stresses that the vendor's fidelity and slicing numbers stay vendor-reported, because nobody outside can rerun the evaluation on the same splits (SelfHostPilot, October 23, 2026).
The announcement also sits inside a larger commercial picture. Hi3D continues to sell Hi3D V3.0, described as the first commercially available AI 3D model with 2048³ voxel resolution, alongside a creator-focused Hi3D Agent and MCP integration. Alongside the open release, the company announced a reduction of more than 50% in credit consumption on its platform. The open model is a foundation layer; the paid products are where the company expects to make money (PR Newswire announcement).
Who Should Try It, and Who Should Skip
Try it if you run a 3D printing workflow and care about slicing success. It also fits indie game developers who need geometry cleanup on drafted assets, or small studios whose products sit far below the license thresholds. For example, a small print shop selling custom parts or a solo developer cleaning up drafted game assets would sit comfortably below the 200,000-user line. If your interest is AI-generated characters for animation rather than printed geometry, our Mesh Avatar Studio article covers a different approach to the same creative itch.
Skip it if you need textures out of the box, if your shop is CPU-only or runs older NVIDIA cards, if your legal team requires an OSI-approved license, or if you need to reproduce and verify benchmark claims before committing. And if your product is anywhere near 200,000 monthly users or USD 1,000,000 in monthly revenue, get the license question answered in writing first.
Frequently Asked Questions
Is Hi3D V1.0 really open source? The code and weights are published and free to download, and commercial use is permitted. But the governing document is the custom Hi3D Open License Agreement 1.0, not an OSI-approved license. Scale triggers and a branding requirement apply, so read the Agreement before building a business on it.
What hardware do I need to run it? The reference build needs Linux, Python 3.10, an NVIDIA GPU with compute capability 8.0 or newer, and a CUDA toolkit from 12.0 to 12.5. There is no CPU inference path, and the top 1536 setting was validated on a 96 GB H20. A community MLX port runs on Apple silicon.
Does it generate textures? No. The output is geometry only: a refined mesh with no textures, materials, or vertex colors. You will need a separate texturing step in your pipeline.
Can I use it in a commercial product? Yes, with conditions. Commercial integrations must show "Powered by Hi3D 1.0" branding with a link, and products that exceed 200,000 monthly active users or USD 1,000,000 in monthly revenue must apply for a separate written license within 60 days.
How does V1.0 differ from Hi3D's paid models? V1.0 is the open refinement model: inference only, up to a 1536³ extraction grid, geometry without textures. The commercial V3.0 offers 2048³ voxel resolution with finer detail and up to 8K textures, plus the Hi3D Agent and MCP integrations. The company also cut platform credit consumption by more than 50% alongside the open release.
The Bottom Line
Hi3D V1.0 is a useful, honestly scoped release: a real model, real code, and real weights for one real job, turning coarse meshes into detailed geometry. Its limits are stated in the open, from the inference-only code to the hardware demands to the custom license. Treat it like a power tool with a rental agreement attached. Check the four things independent coverage flags, license, reproducibility, hardware, scope, and you will know within an hour whether this is a tool to build on or a demo to admire. Start with the GitHub repository and the Hugging Face model page, or try the live demo on Hugging Face Spaces, and read the Indonesian version of this article.
Editorial note: the figures in this article were fact-checked against the linked primary sources on October 24, 2026.