AI Toolkit author trains AR+diffusion hybrid image model on one GPU in 24 hours
AI Toolkit author ostrisai spent a weekend training an AR+diffusion hybrid image model—Qwen3-VL-4B generating image codes decoded by a frozen BFL Klein 4B—on a single RTX 6000 Pro in 24 hours, sharing first samples and a real-time generation demo.
2026-09-20 ~ 2026-09-20 · 4 related posts
- AI Toolkit author trains AR-diffusion hybrid image model in 24 hours on a single RTX 6000 Pro — ostrisai · 2026-09-20
- First samples from Ostris's AR-diffusion hybrid show Qwen3-VL-4B grasping prompt-to-image generation — ostrisai · 2026-09-20
- Ostris shows DiT code-to-image reproduction progress in AR-diffusion hybrid trained on one RTX 6000 Pro — ostrisai · 2026-09-20
- Real-time generation demo shows hybrid image model generating whole-image codes first — ostrisai · 2026-09-20