Looped Diffusion Transformer reuses blocks during denoising for image generation
Radar briefing dated 3 October 2026 on a text-to-image model that reuses Transformer blocks at each denoising step, with the performance claim stated by the original post.
This Radar item describes a text-to-image model in which selected Transformer blocks are reused at each denoising step. According to the original post, published on 3 October 2026, a 260M model outperforms models up to 6.5 times larger. The post also states that extra loops help more than extra denoising steps with the same compute. The source presents block reuse as a way to improve predictions without adding unique parameters at inference. The available text gives no benchmark, dataset or evaluation methodology details.
To consult and verify, read the original article or post, look for the experiments section, the comparison models, the metrics used and how compute was measured. Check these against the 6.5 times claim before treating the result as confirmed. This item is informational only and does not indicate an image generation capability available on Rota Nacional.