"Be horse." – a diffusion language model on an M2 Air
Parallel token decoding beats autoregressive LLMs on throughput, if the math holds up.

Temporally-aware diffusion model faithfully extends cel animation without cropping.
Computer vision researchers and animation fans
Runway ML · Pika Labs · Stable Video Diffusion
Parallel token decoding beats autoregressive LLMs on throughput, if the math holds up.
Animated diffusion step explorer—clear pedagogy, but one interactive doesn't make a course.
Diffusion model generates endless novel climbing routes for home walls—solves real boredom.
Everything you type and generate stays on your machine — model downloads, LoRA mixing, upscaling via Real-ESRGAN and inference run locally while the app shows real-time progress and backend logs. The app leans into practical UX (hardware-based model recommendations, visual memory allocation, generation history) which makes setup and iteration less painful. It isn't reinventing the local-UI wheel — AUTOMATIC1111 and ComfyUI still dominate — but ExoGen is a tidy, approachable alternative for people who want a desktop, privacy-first experience.
Dance choreography is genuinely unexplored in AI UGC; quality still early, but the gap exists.
Runs SDXL locally on iPhone when most competitors require cloud APIs.