Agent edits 0.00017% of SDXS-512 weights, no training, turning postcards into walkable scenes
IgorCarron · x · 2026-11-10
User TensorSlay demonstrated "weight space programming" on the image model SDXS-512: an agent directly edited just 0.00017% of the network's weights — with no training — transforming flat postcard images into scenes with depth you could seemingly walk into. A reproduction recipe was shared in the comments; Igor Carron reposted it with "And they fall..."
More from Multimodal
- AI-generated checkerboard art series: neon jester, painted parakeet portraits — OVolosin82152 · 2026-11-10
- Nuance launches Emma, an AI video-call persona that's nearly indistinguishable from a real human — FlorianGallwitz · 2026-11-10
- AI-generated 'Matrix of Medusa' art series fuses Greek myth with glitch aesthetics — LudovicCreator · 2026-11-10
- Claude 3 Sonnet co-designs abyssal sea turtle that sings with its breath — repligate · 2026-11-10
- One of these two people is AI: Nuance avatars get indistinguishable — mark_k · 2026-11-10
- Retro AI music video made with Fable 5.2 and a Claude Music Video skill — AIandDesign · 2026-11-10