Pixel-Space Diffusion Training Slower, Ideal for Distillation with 3x Speedup
bdsqlsz · x · 2026-08-20
A new paper, "An Empirical Study of Training Pixel-Space Text-to-Image Diffusion Models," investigates training efficiency in pixel versus latent spaces. It finds that direct large-scale pre-training in pixel space converges significantly slower than in latent space.
Key Findings & Strategy:
- Proposes a "latent-to-pixel" strategy: efficiently acquire generative priors in latent space, then transition to pixel space during post-training.
- Systematically examines key design choices like weight initialization, data composition, and prediction targets.
- Results: The resulting pixel-space models match or outperform latent-space counterparts while achieving 3.18x to 4.75x end-to-end inference speedups.
- Tip: Using a noise scale of 2 can reduce color drift.
Related event: Study: Pixel-Space Diffusion Training Better for Distillation(2 posts)→
More from Research
- Transformer trained on flying bat recordings yields hippocampus 'digital twin' revealing hidden structure — patrickmineault · 2026-10-16
- One year of Recursive Language Models: GPT-5-mini beats GPT-5 by 110% on 132k-token tasks, cheaper too — a1zhang · 2026-10-16
- MIT PhD Sharut Gupta joins DeepMind; Isola hails her representational alignment defense talk — phillip_isola · 2026-10-16
- 36 primitives, just 10 parameters each: a remarkably precise T-pose body fit — letandrewcook · 2026-10-16
- Interp researcher pushes back: if you can't simulate a neuron, you haven't explained it — nabla_theta · 2026-10-16
- MLP-Splatting (ECCV 2026) code released: MLPs as light-field primitives instead of Gaussians — AjdDavison · 2026-10-16