Skip to content
Help Build a Better ComfyUI Knowledge Base Become a Patron
NewsStepFun Open Sources Step1X-3D - High-Fidelity 3D Asset Generation Framework

IndexTTS 1.5 Release: High-Quality Chinese and English Text-to-Speech Model

05/23/2025

StepFun Open Sources Step1X-3D High-Fidelity 3D Asset Generation Framework

StepFun has officially open-sourced Step1X-3D, a comprehensive framework for high-fidelity 3D asset generation. This framework can generate 3D models with fine geometric structures and diverse textures from a single image, and is the first to achieve direct transfer of 2D control techniques to 3D generation.

Step1X-3D Showcase

Key Features

Step1X-3D adopts an innovative two-stage generation architecture that decomposes the 3D generation process into geometry generation and texture synthesis - two independent but coordinated stages. The framework features the following core capabilities:

High-Quality Data Processing Pipeline

The team constructed a training dataset of 2 million high-quality 3D assets through rigorous data cleaning and filtering processes from over 5 million original 3D assets. This dataset achieves high standards in geometric precision, texture quality, and topological integrity.

Advanced Geometry Generation Technology

The geometry generation module employs a hybrid VAE-DiT architecture capable of generating watertight Truncated Signed Distance Function (TSDF) representations. Through perceiver encoding and sharp edge sampling techniques, the system effectively preserves geometric details and generates topologically sound 3D meshes.

System Architecture

Precise Texture Synthesis

The texture synthesis module is fine-tuned based on Stable Diffusion XL, providing geometric guidance through normal maps and position maps to ensure precise alignment between generated textures and 3D geometry. The system supports multi-view consistency and can generate high-resolution texture maps.

Flexible Control Mechanisms

Step1X-3D supports parameter-efficient fine-tuning techniques like LoRA, allowing users to control object symmetry, geometric detail levels, and other attributes through tags. This provides users with more creative control options.

Technical Advantages

Compared to existing open-source solutions, Step1X-3D excels in multiple aspects:

Generation Quality: In benchmark tests, Step1X-3D’s geometry and texture generation quality surpasses existing open-source baselines, achieving performance comparable to commercial solutions in certain metrics.

Complete Open Source: Unlike many projects that only release model weights, Step1X-3D provides complete training code, data processing pipelines, and adaptation modules, facilitating reproduction and improvement by researchers.

Ecosystem Compatibility: By supporting the transfer of 2D control techniques to 3D, Step1X-3D forms good compatibility with existing image generation ecosystems.

Open Source Contents

This open source release includes:

  • Model Weights: Including geometry generation model (1.3B parameters) and texture synthesis model (3.5B parameters)
  • Training Code: Complete training code for VAE, diffusion models, and multi-view generation
  • Dataset: UID list of 800K high-quality 3D assets
  • Online Demo: Interactive demonstration on HuggingFace Spaces
  • Adaptation Tools: Adaptation modules supporting LoRA fine-tuning

Use Cases

Step1X-3D is suitable for various application scenarios:

Content Creation: Rapid 3D asset generation for game development, film production, and other fields Product Design: Quick 3D prototype generation based on concept images Education & Training: Auxiliary tools for 3D modeling and design education Research & Development: Foundation platform for 3D generation algorithm research

Technical Details

Geometry Generation Pipeline

The system first uses a 3D shape variational autoencoder to compress point clouds into latent space, then performs geometry generation through a FLUX-inspired diffusion transformer. This process employs sharp edge sampling and dual cross-attention mechanisms to enhance geometric detail preservation.

Texture Synthesis Pipeline

Texture generation uses a multi-stage pipeline: first post-processing geometry to ensure topological consistency, then creating textures through multi-view image generation models, and finally completing texture mapping through UV baking and repair.

Performance Results

In user studies, Step1X-3D achieved high scores in geometric rationality, texture clarity, and overall quality, demonstrating its potential for practical applications.

Community Response

Since its release, Step1X-3D has attracted widespread attention in the open-source community. The project has gained significant developer attention on GitHub, and the online demonstration on HuggingFace has attracted many users to experience it.

Many researchers have stated that Step1X-3D’s complete open-source strategy provides valuable resources for research in the 3D generation field, helping to advance the entire domain.

Future Plans

According to the project roadmap, the team plans to release more features in the future:

  • Support for additional control conditions like multi-view, bounding boxes, and skeletons
  • ComfyUI workflow integration support
  • More controllable generation models
  • Performance optimization and inference acceleration

Related Links