Compressing 2.7B Pixels into a 790k-Parameter Neural Network
thisdudelikesAI · x · 2026-08-12
A developer successfully encoded the 2.7 billion pixels (1,620 frames) of the 'Bad Apple' animation into a tiny 3.2MB neural network with just 790k parameters.
Core Concept: Instead of storing frames or pixels like traditional codecs, the video is represented as a continuous mathematical function. The network takes a 3D coordinate (time, x, y) as input and computes the exact pixel value, effectively 'memorizing' the video in its weights.
Key Optimizations:
- Time-stretch: Scaling the time coordinate by 4x before the first layer, providing 4x more temporal capacity for free.
- Motion-focused sampling: Since 90% of the animation is static black, uniform sampling starved the moving edges of gradients. The fix was dedicating half of every training batch to pixels that actually changed between frames.
Results: Validation MSE dropped from 0.0795 to 0.0090 (9x better). 398 out of 400 sampled frames improved.
While 3MB is larger than the 700KB source file, the breakthrough proves the paradigm of 'learning' rather than 'storing' video works, opening doors for scaling up to real footage and audio.
More from Research
- Atlas Discovery Unveils ClinicBench to Evaluate AI Agents' Clinical Decisions — BraydonDymm · 2026-08-12
- Developer solves image degradation in multi-turn editing for SOTA image generation models — imdigitalashish · 2026-08-12
- InSight-doc: Adaptive Visual Resolution for Long-Document Understanding — Kaican Li · 2026-08-12
- AI Needs Internal 'Social World Models' for Social Intelligence, Says MIT Researcher — xuanalogue · 2026-08-12
- Fields Medalist Tim Gowers Asks: What Sort of Maths Are LLMs Good At? — ColinWright · 2026-08-12
- Running Qwen3.5-0.8B on a Sub-$20 CPU with Under 512MiB Memory — buryhuang · 2026-08-12