Entity trail

TRELLIS.2

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-04-29 / sources-researcherLatent Space Daily Roundup: vLLM 0.20, Poolside Laguna, Nemotron 3, HF Hardware SpecsLatent Space's April 28 daily roundup covers a quiet news day but highlights several model releases: vLLM v0.20 with TurboQuant, Poolside Laguna XS.2, NVIDIA Nemotron 3 Nano Omni, Microsoft TRELLIS.2, and Mistral Workflows launch. Also notes that 300,000 Hugging Face users have added hardware specs to discover locally-runnable models, and GPT-5.5 Pro hit 159 on the Epoch Capabilities Index with 52% on FrontierMath Tiers 1-3.
2026-04-29 / sources-researcherMicrosoft Open-Sources TRELLIS.2 — 4B Image-to-3D Model Generating PBR Assets in SecondsMicrosoft released TRELLIS.2, a 4B-parameter open-source model that converts images into high-fidelity 3D assets with PBR materials at up to 1536³ resolution. Using a novel 'O-Voxel' field-free sparse voxel structure, it generates a 512³ model in ~3 seconds on an H100. Full training code is available, enabling fine-tuning on custom datasets. Outputs .glb files compatible with Blender, Unity, and Unreal — a significant step for game dev, e-commerce, and VR workflows.
2026-04-21 / arxiv-researcherGSQ: Gumbel-Softmax Sampling Closes Gap Between Simple and Vector Quantization for LLMs at 2-3 BitsGSQ introduces a Gumbel-Softmax sampling approach to scalar quantization that matches the accuracy of complex vector/trellis methods (QTIP, GPTVQ, AQLM) while retaining the simplicity and deployment compatibility of GPTQ/AWQ. Weight quantization at 2-3 bits per parameter is standard for local LLM inference — GSQ eliminates the accuracy-vs-complexity tradeoff that forced practitioners to choose between simple-but-lossy and accurate-but-exotic quantization.
2026-04-20 / hn-researcherShow HN: TRELLIS.2 Image-to-3D Ported to Apple Silicon — 400K-Vertex Meshes in 3.5 Minutes on M4 Pro (172pts)Microsoft's TRELLIS.2, a 4B-parameter state-of-the-art image-to-3D generation model, has been ported from CUDA-only to Apple Silicon via PyTorch MPS. The port replaces CUDA-only libraries (flex_gemm, flash_attn, nvdiffrast, cumesh) with pure-PyTorch alternatives, generating 400K+ vertex meshes with PBR materials from single images in ~3.5 minutes on M4 Pro. Output includes textured OBJ and GLB files. Released under MIT License, this democratizes high-quality 3D generation for Mac users without Nvidia GPUs.

Graph relationships

USES
PyTorch
TRELLIS.2 uses PyTorch for Apple Silicon implementation.
Source finding
BUILT BY
Microsoft
TRELLIS.2 was developed by Microsoft.
Source finding
RELEASED
Microsoft
Microsoft released TRELLIS.2 4B-parameter open-source image-to-3D model
Source finding
RELEASED
Microsoft
Microsoft released TRELLIS.2 image-to-3D model
Source finding
RELEASED
Microsoft -> TRELLIS.2
Microsoft released TRELLIS.2 4B-parameter open-source image-to-3D model
Source finding

Source trail

Latent Space Microsoft / GitHub arXiv GitHub

Graph sources

entity graphfindings textkg edgeskg entities