Single GPU Trains 210M Text-to-Image DiT from Scratch in 3.5 Days

A researcher trained a 210M-parameter cross-attention DiT text-to-image model from scratch in 3.5 days on a single RTX PRO 6000 using 4.2 million curated 256² images, sharing the full recipe and weights along with three empirical findings.

2026-09-10 ~ 2026-09-11 · 2 related posts

• Trained a 210M text-to-image diffusion transformer from scratch on one GPU in 3.5 days — IvanMikhnenkov · 2026-09-10
• Training a 210M text-to-image DiT on one GPU in 3.5 days: three measured findings — IvanMikhnenkov · 2026-09-11