GaussianGPT uses autoregressive next-token prediction to generate 3D Gaussian scenes
rsasaki0109 · x · 2026-07-28
GaussianGPT generates 3D scenes with autoregressive next-token prediction
GaussianGPT is an ECCV 2026 paper that explores a fully autoregressive alternative to diffusion and flow-matching for 3D generative modeling. The model directly generates 3D Gaussians via next-token prediction, enabling full scene generation in a step-by-step manner.
How it works
- First compresses Gaussian primitives into a discrete latent grid using a sparse 3D convolutional autoencoder with vector quantization.
- Serializes the resulting tokens and feeds them to a causal transformer with 3D rotary positional embeddings.
- Learns spatial structure and appearance sequentially rather than refining the scene holistically.
What this enables
- Scene completion
- Outpainting
- Controllable generation
- Larger-scale 3D scene synthesis
More from Multimodal
- User shares a new Midjourney style with exact prompt settings — azed_ai · 2026-07-28
- A 5 MB McBess-style LoRA for Krea2 trained on 120 captioned images — Winter_unmuted · 2026-07-28
- Topview launches Film Studio with 3D blocking and micro-expression controls — XFreeze · 2026-07-28
- Testing Style LoRAs: How to Isolate Style from Content in Image Generation — Dark_Sytze · 2026-07-28
- Oxygen-TryOn is a fashion-native model for multi-item virtual try-on — JD-company · 2026-07-28
- FilmBench evaluates cinematic video generation with film-school shot lists and 35 metrics — Shengyi Wang · 2026-07-28