StochSync – an AI image generation technology that generates high-quality images for 360° panoramas and 3D mesh textures.

What is StochSync?

StochSync (Stochastic Diffusion Synchronization) is an innovative image generation technique specifically designed for generating images in complex spaces, such as 360° panoramas or 3D surface textures. It combines the advantages of diffusion synchronization (DS) and fractional distillation sampling (SDS), balancing image detail and coherence by introducing maximum randomness during the inverse generation process of the diffusion model. StochSync requires no additional training, performs exceptionally well in panorama and 3D texture generation, and can generate high-quality images even without image conditions.

StochSync's main functions

• Zero-sample generationNo additional training is required on the data in the target space, thus avoiding the overfitting problem.
• Balancing details and coherenceBy introducing maximum randomness in the reverse diffusion process of diffusion synchronization (DS) and combining it with multi-step denoising methods, the detail quality of DS is preserved, and the coherence of fractional distillation sampling (SDS) is achieved.
• 360° panoramic image generationStochSync can generate high-quality panoramas even without image conditions, outperforming previous fine-tuning methods.
• 3D Mesh Texture GenerationGiven a depth map, StochSync can generate textures comparable to existing methods.

StochSync's technical principles

• Diffusion synchronization (DS)DS generates high-quality images by jointly performing the backdiffusion process in different projection spaces and synchronizing intermediate outputs in the target space. However, DS performs poorly in terms of global coherence when sufficient conditional information (such as depth maps or image conditions) is lacking.
• Fractional distillation sampling (SDS)SDS updates the target space data step by step through gradient descent, which can generate more coherent images, but often lacks detail.
• Combining the two methodsStochSync first revealed the connection between DS and SDS by introducing maximum randomness in the back diffusion process of DS (i.e. setting the noise level σ_t to the highest value at each step) and combining it with multi-step denoising methods to achieve the detail quality of DS and the coherence of SDS.

StochSync's technological advantages

• Maximum randomnessStochSync introduces maximum randomness in each step of the denoising process, enhancing global coherence.
• Multi-step noise reductionBy replacing the traditional single-step prediction with a multi-step denoising process, the realism of the generated image is further improved.
• Non-overlapping view samplingStochSync uses non-overlapping view sampling to avoid inconsistencies between views, while achieving time synchronization through view overlap.

StochSync project address

• Project official website:https://stochsync.github.io/
• arXiv technical paper:https://arxiv.org/pdf/2501.15445v1

Application scenarios of StochSync

• High-resolution image generation:It can generate 8K resolution panoramic images and avoid the occurrence of repeating patterns through optimization techniques.
• Texture mapping of complex geometries:StochSync can be used to generate textures for 3D Gaussian models, demonstrating its potential for application on complex geometries.