Open Replication of DeepMind's Vision Banana Beats Original on Cityscapes
y_m_asano · x · 2026-11-13
FunAI Lab fully rebuilt Google DeepMind's Vision Banana (which remained closed data/weights) using only public data, via FLUX.2 Klein 4B/9B finetunes.
- Beats the original: 73.9 vs 69.9 mIoU zero-shot on Cityscapes.
- Releases code and 148 checkpoints.
- New test-time scaling recipe "legend-ensembling": strong gains by adding compute at inference with no retraining.
Related event: Open Vision Banana: Fully Open Replication Beats Google's Original(2 posts)→
More from Multimodal
- Custom motion-director skill turns any voiceover into 1080x1920 motion-graphics animations with Gemini Spark — mijozakinho_2 · 2026-11-13
- MiniMax's H3 Video Model Lands on Envato, Free and Unlimited for 24 Hours from Nov 19 — cneuralnetwork · 2026-11-13
- Envato offers free unlimited MiniMax H3 access for 24 hours starting Nov 19 — EXM7777 · 2026-11-13
- Synthesia launches interactive AI clones: record a short video, get a talking you in 2 minutes — synthesiaIO · 2026-11-13
- AI-generated 'Pokémon Hood' video: a bazooka-wearing turtle strolls through a gritty retro street — ImportantBrief1900 · 2026-11-13
- AI-generated Three Kingdoms teaser: Sun Ce dies at 26 and wakes up five years earlier — KoreanIndieDev · 2026-11-13