First samples from Ostris's AR-diffusion hybrid show Qwen3-VL-4B grasping prompt-to-image generation
ostrisai · x · 2026-09-20
Ostris shares first samples from his AR/diffusion hybrid image model (Qwen3-VL-4B autoregressive model + frozen BFL Klein 4B diffusion decoder), generated with AI Toolkit's default prompts. The AR model has clearly learned to produce images matching the prompts.
More from Multimodal
- Experimental video mixes visual styles, generated with Seedance 2.5 on Adobe Firefly — umesh_ai · 2026-11-04
- Hailuo AI claims H3 scored perfect on language tests; Japanese creator praises its Japanese signage fidelity — Hailuo_AI · 2026-11-04
- YANAMI: 400K-dialog multimodal dataset shows models still fail at stateful multiturn interaction — Shengyu Wang · 2026-11-04
- Open-source Idol model turns a single image into 3D in 45s on 8GB VRAM — Promptmethus · 2026-11-04
- ComfyUI v0.44.0 released with new Patina texture nodes — Fcking_Chuck · 2026-11-04
- Krea 3 technical report released: a deep dive into data, architecture and training — bdsqlsz · 2026-11-04