Image captioning with zero image-text pairs: one rotation maps DINO features into text space
randall_balestr · x · 2026-10-30
Follow-up on the PRH theorem thread: the authors fine-tuned a small LLM decoder on text only, on top of their text-JEPA. A photo goes through a DINO encoder, one rotation maps it into text space, and the decoder describes it — no network was ever trained on image-caption pairs. Sample captions shown in the post.
More from Multimodal
- Comfy Studio: open-source tool turns ComfyUI workflows into creative apps with no frontend coding — Realistic_Win_9200 · 2026-10-30
- A minimal abstract monochrome silhouette prompt template for image generation — azed_ai · 2026-10-30
- Kling AI launches Originals program with Halloween short film made on Kling 4.0 — azed_ai · 2026-10-30
- CreatorOS Studio: open-weights local ComfyUI video suite hits 100+ features, removes people from clips — Key_Cap_8111 · 2026-10-30
- Nightshift: open-source ComfyUI video studio drafts by day, finishes the same take overnight locally — Huge_Relation2996 · 2026-10-30
- Open-source ComfyUI light stage: drag lights around your subject for AI video relighting — Holiday-Opposite8514 · 2026-10-30