“‘NiuLai’ is finally a thing of the past! Hi3D raises 3D modeling to new heights, v1.0 now open‑source.

Editor | Cold Cat

Since the phenomenon‑level success of the movie “Bull Come,” that crudely modeled, yellowish little calf has become many people’s first reaction when mentioning 3‑D modeling.

The director’s claim of “pure hand‑crafting” unexpectedly awakened netizens’ modeling spirit. Some are eager to try, while others can’t help but joke: nowadays, letting AI generate a 3D model is probably far more refined than “Bull Come.”

Jokes aside, how far has AI‑generated 3D modeling actually evolved? Let’s look at two comparison sets to feel the current level of detail in AI modeling.

Original image / ChatGenerated by GPT‑6 model / Generated by Hi3D Agent

These results come fromShumei万物’s self-developed 3D generation model Hi3D V3.0, which is pushing geometry and appearance representation to an incredibly fine scale.Achieved an industry-first nominal geometry refinement resolution of 2048³, with up to 8K texture output. The 3D model generation capability of Hi3D V3.0 far surpasses the general model GPT-6.

From these two sets of examples, it is clear that Hi3D V3.0 delivers a striking performance on the two major challenges of 3D modeling: highly complex structures and animal fur. Delicate shallow relief with subtle bumps, clear structural boundaries between adjacent parts, the natural growth patterns of animal hair… These seemingly minor details are precisely the aspects that test the skill in 3D modeling. The delicacy and realism of a 3D piece often lie hidden in these details.

And while our self‑developed models continue to iterate forward,Shumei Wanwu will bring the production‑validated Hi3D V1.0 model to open source.

• GitHub link: https://github.com/hi3d-official/Hi3Dv1.0

• Hugging Face demo link: https://huggingface.co/spaces/ilcve21/Sparc3D

• Hugging Face code weights link: https://huggingface.co/hi3d-official/Hi3Dv1.0

• Official project link: https://www.hi3d.ai/hi3d-1-0-open-source

If Hi3D V3.0 demonstrated the ongoing breakthrough potential of 3D generation technology, then the open‑source release of Hi3D V1.0 allows the accumulated technical achievements to enter the actual workflows of more developers.

Approximately 2 million 3D assets for training scale, up to 1536³ effective sparse surface extraction mesh resolution, and validation through real user workflows before public release, together form the technical foundation of Hi3D V1.0. Building on this foundation, developers can further create their own creative tools, professional applications, and workflows, extending AI 3D generation capabilities to more real-world scenarios.

We believe that from now on, the “NiuLai”-style rough 3D production will become a thing of the past.

Continuously improving model capabilities also enables verified results to be used by more people — the Hi3D V1.0 open‑source program connects Shumei’s technological advancements with broader industry co‑creation.

Make fine geometry the foundation for continued creation

As everyone knows, the biggest challenge in text-to-3D is:getting a model to start from a single image and generate a complete 3D asset that looks reasonable from every angle, is richly detailed, and can actually be used for downstream creation. Solving this problem requires tackling the underlying geometric representation and generation method.

In 2025, Sparc3D attracted widespread attention in the 3D geometry modeling field thanks to its high‑resolution sparse representation and sparse convolution auto‑encoding pipeline. Its core idea is not to represent the entire 3D space with a dense voxel grid, but to allocate computational resources only around the predicted surface region, encoding geometric information as features on sparse coordinates.

Sparc3D pipeline diagram

Hi3D V1.0 builds on this architecture and accomplishes three major tasks.

First, dramatically expand the training scale.The 3D assets used for training grew from Sparc3D’s previous size of about 500,000 to roughly 2,000,000, an increase of about fourfold.Why is data volume so critical? Single-image 3D generation is inherently an under-constrained problem; the input image only provides partial information about the object’s surface, and the remaining geometry must be inferred by the model from the 3D priors learned during training. Therefore, expanding the scale and geometric coverage of the training data helps the model learn a wider variety of object shapes and improves its ability to handle complex structures and long-tail samples.

Second, introduce higher‑resolution training data. Based on Sparc3D’s existing 512³ and 1024³ levels,Hi3D V1.0 adds training data at the 1536³ level. The increase in resolution means the model can learn and represent geometric changes at finer spatial scales. For example, shallow relief on complex armor surfaces, tiny gaps between mechanical parts, the layered structure of feathers, and dense textures on curved surfaces all require sufficient spatial resolution to preserve their original structure as much as possible.

It is worth noting that the 1536³ here refers to an effective sparse grid resolution; the actual flow model does not directly store a dense 1536³ latent space. In the first level, latent features are distributed over a sparse coordinate set of a nominal 64³ structured grid, and the decoder extracts geometric details through three learnable 2× expansions. This means the final surface reconstruction is performed at an effective resolution of 512³, without incurring the computational cost of a dense high‑resolution grid.

Third, complete validation in a real production environment. Before the official open‑source release, the model checkpoint corresponding to Hi3D V1.0 had already entered a production workflow serving real users and had been tested in actual usage scenarios. The technical report specifically emphasizes that Hi3D V1.0 does not introduce new generation targets, decoders, or surface representation methods; all improvements are made within the existing Sparc3D framework.

Hi3D V1.0 demonstrates a highly practical technical roadmap: building on a mature geometric generation architecture, it further pushes the model’s capability ceiling by continuously scaling data size and resolution.

Engineering Design of Geometry

So, how does Hi3D V1.0 manage to generate fine geometry while controlling the overhead of high‑resolution computation? The answer lies in its staged generation process.

The entire pipeline is divided into two stages. In the first stage, an upstream module generates a coarse mesh from the input image, establishing the object’s global structure. In the second stage, Sparc3D’s refinement process takes over, using both the reference image and the coarse mesh: image features provide appearance and shape cues, while the coarse mesh provides a geometric coordinate foundation. Then, the sparse flow model generates latent features on these coordinates, which are progressively expanded by a sparsity‑aware VAE decoder with self‑pruning capability, and finally a sparse iso‑surface extraction yields a finer 3D mesh.

The key design here is that the generative model does not need to directly handle a massive high‑resolution dense 3D space.

Hi3D V1.0 defines three progressively refined geometry reconstruction levels:

• Level 1: Computed on sparse coordinates of a nominal 64³ structural grid, ultimately extracting a geometry surface at an effective resolution of 512³.

• Level 2: The structural grid is upgraded to 128³, corresponding to an effective surface extraction resolution of 1024³.

• Third level: further increased to 192³, corresponding to a maximum effective surface extraction resolution of 1536³.

Within each level, the decoder performs three learnable 2× upsamplings, achieving an 8× scale increase from the structural grid to the surface extraction grid.

This division of labor allows the system to maintain high resolution while controlling computational cost. The flow transformer does not need to directly maintain a dense high‑dimensional latent space; instead, it operates on low‑dimensional sparse coordinates and progressively enlarges them via the decoder. This also explains why scaling the data size and increasing resolution levels can be accomplished within the same framework without redesigning the architecture.

This design gives high‑resolution geometry generation better scalability. More importantly, the expansion of training data scale and the improvement of resolution correspond precisely to two distinct challenges in 3D generation.

The scale of the data determines how rich a geometry a model can learn, while the resolution affects how finely the model can express geometric structures.

Three‑dimensional assets contain many long‑tail structures, such as intricate cutouts, interleaved thin‑walled parts, and asymmetric organic shapes. Because training samples are limited, models often lack sufficient geometric priors, making it difficult to infer occluded regions and leading to structural collapse or over‑smoothing. Larger‑scale training data helps models learn a wider variety of geometric forms, improving their ability to reconstruct complex structures.

The introduction of an effective resolution of 1536³ addresses another issue. High‑resolution training data enables the model to learn finer geometric details and distinguish true surface structures from noisy artifacts. When a thin‑walled component or sharp edge might be swallowed by quantization error at 512³ resolution, the 1536³ training signal provides the model with much more precise supervision.

Comparison of local geometric details in seven typical regions. Source: Technical Report

The improvements in both directions are combined; broader geometric experience together with higher‑precision detail learning form the core competitiveness of Hi3D V1.0.

Quality Evaluation Logic

A 3D asset cannot be evaluated solely by a single rendered image, and front‑view similarity does not represent the overall 3D quality.

Based on this idea, the research team goes beyond traditional image semantic consistency metrics and further evaluates the generated results from three dimensions.

First is Multi‑View Geometric Consistency (MV Fidelity).The researchers render the generated asset from six standard viewpoints—front, back, left, right, top, and bottom—to examine whether the object maintains a reasonable structure from different directions, while checking for defects such as missing parts, floating fragments, or duplicated geometry. This test assesses the model’s understanding of the complete 3D structure, not just its ability to reproduce the input image.

Next is geometric detail richness (Detail Richness).This part evaluates while minimizing interference from textures and materials, using gray‑scale rendering and surface normal maps to examine subtle surface undulations, structural hierarchy, the preservation of small components, and geometric details in regions of the input image that are not visible.

Finally, the geometry’s practical usability.The research team also looks for detached fragments, surface boundary defects, and insufficient wall thickness in the mesh, and assesses the model’s compatibility with downstream 3D‑printing workflows using metrics such as slicing success rate.

This more comprehensive evaluation method also provides concrete data support for the advantages of Hi3D V1.0. Among the nine methods compared in the technical report, Hi3D V1.0 achieved a multi‑view geometric consistency score of 0.75 and a geometric detail richness score of 0.841, both of which are the highest in the table.

Quantitative comparison of different models on four metrics: geometric resolution, multi‑view fidelity (MVF), geometric detail richness (DR), and slice success rate (SSR). Source: Technical Report

In addition, the research team designed a user preference study comprising 562 human judgments, comparing Hi3D V1.0 with representative Image-to-3D methods such as Pixal3D, TRELLIS.2, Hi3DGen, and TripoSG. The evaluation criteria go beyond a single positive similarity metric and focus on the following aspects:

• Geometric completeness from unseen viewpoints: whether the model’s structure remains reasonable when rotated to side, back, or bottom views

• Preservation of thin walls and fine structures: details that are easily lost such as weapon blades, character accessories, mechanical parts, etc.

• Organization of local geometric details: whether surface features are presented as true geometry or degenerated into texture noise

• Downstream Production Compatibility: Whether the generated mesh can be used directly in slicing, editing, 3D printing, and other practical workflows

The results show that out of 562 judgments, Hi3D V1.0 achieved 469 wins, 28 ties, and 65 losses. Using the scoring rule of 1 point for a win and 0.5 points for a tie,the overall human preference score reached 85.9%.

User preference score. Source: Technical report

Continuous iteration, a new form of creation in the 3D world

From initially pursuing higher geometric precision, to perfecting textures and structures, and then bridging the practical application stage after model generation, Hi3D has undergone three major version evolutions, and its technical roadmap is becoming increasingly clear.

• V1.0 focuses on high‑precision geometry generation, raising the effective sparse surface extraction resolution to a maximum of 1536³, enabling more accurate reconstruction of complex structures and fine geometry.

• V2.0 extends capabilities to integrated geometry and texture generation, introducing intelligent occlusion completion, relighting removal, and other techniques to further improve model structural integrity and material appearance.

• Version 3.0 pushes geometric reconstruction resolution to 2048³, further improving complex structures, local details, and texture quality. At the same time, supporting capabilities such as intelligent part separation, multi‑color printing, and automatic layout are continuously refined, making the generated 3D models easier to integrate into real‑world creation and production workflows.

In the era of agents, having a powerful model alone is not enough. How to organize model capabilities so that AI truly understands creative intent, invokes specialized tools, and completes a series of complex tasks is becoming the new competitive focus.

Following this direction, Shumei has launched the Hi3D Agent.

Hi3D Agent is a “Creator-First 3D Agent”, the steering wheel always stays in the creator’s hands, while AI handles the accelerator. This concept is akin to the currently popular Vibe Coding. The latter allows ordinary people to develop software through natural language, whereas Hi3D Agent brings a similar experience to 3D creation, exploring a new “Vibe Creation” path.

Hi3D Agent uses a large language model to understand natural language goals, schedule tasks, and then orchestrates specialized Hi3D models and related tools to perform generation and processing. Its exploration covers existing asset adjustment, character animation handling, as well as disassembly and print preparation, enabling creators to continue advancing subsequent steps around their work.

Starting from a reference image, first define the character’s form, then gradually refine the details, and finally choose a path oriented toward display, interaction, or physical production—this creative workflow provides concrete scenarios for the integration of professional models and agents. The creator is responsible for selecting the direction and judging the results, while the system organizes the corresponding capabilities around these intentions.

Creators can build upon existing characters by adding props, constructing scenes, and then integrating these 3D assets into an interactive environment, progressively advancing toward a playable game demo. Perhaps the recent Steam New Releases event set a new record for the number of new games and ideas, many of which involve AI-generated 3D art.

At the same time, Shumei万物 is advancing the integration of Hi3D MCP, enabling external agents, specialized tools, and existing workflows to invoke Hi3D’s 3D generation capabilities. Developers therefore have the opportunity to build more scenario‑specific AI 3D applications within familiar environments.

From continuous iteration of underlying models, to the ongoing refinement of 3D creation tools, and to the new creation paradigm driven by agents, every step of Shumei万物 expands the capabilities and application boundaries of AI 3D.

From open models, infinite possibilities grow

The iteration of Hi3D models is always exciting, because the more ideas it encounters, the more brilliant technologies emerge. Each advancement in model capability means that ideas once limited by technical barriers now have the potential to become reality.

And the open‑source release of Hi3D V1.0 extends this possibility even further to a broader range of developers and creators.

Independent developers can build dedicated tools around their favorite subjects, professional studios can integrate 3D generation into existing production pipelines, and research teams can stand on the existing technical foundation to continue exploring new geometry generation methods.

This is also the key significance of Shumei Wanwu’s ongoing investment in model development and promotion of technology openness. From V1.0 to V3.0, from high‑precision geometry generation to agent‑driven creative workflows, Shumei Wanwu continuously expands the technical boundaries of AI 3D while striving to bring these capabilities into more real‑world scenarios.

Now, with Hi3D V1.0 open‑sourced, this technology stack, proven in production environments, becomes a starting point that more developers can use, study, and further innovate upon.

As the technical barrier gradually lowers, creators can devote more energy to character design, scene conception, and artistic expression. The model provides increasingly rich means of implementation, but the final appearance of the work still depends on human imagination.

Technological progress has no end, and the possibilities of creativity have no boundaries.

When high-precision 3D generation becomes a capability that anyone can invoke and explore, the next breakout work may appear in your hands.

© THE END

For reprints, please contact this public account for permission

Submit or request coverage: liyazhou@jiqizhixin.com