---
title: "3D Generation"
description: "2 posts about 3D Generation, summarised, each linking to the original."
---

# 3D Generation
> 2 posts about 3D Generation, summarised, each linking to the original.

## Articles

### [Split-and-Scale Latent 3D Representations](https://yomu.fyi/post/split-and-scale-latent-3d-representations.md)
- Company: [Apple](https://yomu.fyi/company/apple.md)
- Published: Nov 16, 2026
- Type: Problem & solution

Latent 3D representations enable compact encoding of scenes and support generative models, but they struggle to capture fine details due to limited latent capacity. Scaling capacity at training is computationally expensive and often infeasible under memory limits, while naively increasing latents at inference fails to use added capacity effectively. Split-and-scale addresses both challenges by representing 3D scenes as continuous functions of 3D coordinates, allowing a single pretrained tokenizer to be applied at inference to arbitrarily small sub-regions. This allocates the same latent capacity to a smaller spatial extent, capturing far higher detail. Extensive ablations show inference-time scaling achieves quality competitive with much larger models that exceed single-GPU memory limits, and the method supports high-quality generative modeling with an autoregressive 3D generator competitive with state-of-the-art baselines.


### [The Definitive Fine-Tuning Guide for 3D Generation Models](https://yomu.fyi/post/the-definitive-fine-tuning-guide-for-3d-generation-models.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Suvaditya Mukherjee, merve
- Published: Nov 12, 2026
- Type: Tutorial

Fine-tuning a general-purpose 3D generation model like TRELLIS.2 for a specific object category can correct geometry errors, such as inaccurate shoe soles. The authors add LoRA adapters only to the sparse-structure flow DiT and the shape flow DiT, freezing the texture flow DiT to avoid training on diffuse color maps that include baked lighting. Using 188 footwear objects from the Google Scanned Objects dataset, the fine-tuned model improves the mean F-score from 0.276 to 0.803 on nine held-out objects. Chamfer distance falls from 0.174 to 0.033, and normal consistency rises from 0.530 to 0.804. The approach uses supplied Docker containers and Hugging Face Jobs, with the sparse-structure run taking 7.2 hours and the shape run taking 3.6 hours on one A100 80 GB GPU.
