Frontier models are usually scored in isolation. Aiwake has to keep two of them under unscripted pressure, remember what was already said, and turn the collision into a phonetic 2D anime reel without a GPU render farm.
ENGINEERED OUTCOME
A parametric 2D anime engine: Rhubarb visemes, shot-reverse-shot rigs, and a RAM-to-FFmpeg pipe that finishes a 40s reel in 19s. The classic terminal reel stays available as the alternate player.
Each stage activates its matching runtime node. Inspect contracts, payload shape, and measured or expected latency without leaving the execution context.
Bridge / architectural decisions
The 19-Second Compile Barrier: Why We Left Cloud Diffusion
The player above is the current state: a broadcast 2D anime engine that compiles a reel in under 19 seconds. These two branches are the decisions that made it possible — leaving cloud diffusion, then replacing both crude vectors and photoreal mouth swaps with a parametric cel-shaded rig. The trace terminates in the terminal compiler that first proved 7× realtime throughput and a sub-10ms event bus, before a diffusion credit was ever spent.
BRANCH // 01 · COMPUTE
The Cloud Diffusion Trap
Challenge
Sora, Runway, and HeyGen took 5–10 minutes per video, burned $2–$5 per finished minute, and introduced temporal face flicker.
Breakthrough
A deterministic 2D parametric media pipeline replaced diffusion with Python raster composition: a broadcast-ready reel in 19 seconds, twice real-time speed, with zero GPU cost.
Naive geometry looked like a 2004 Flash toy. Photoreal 2D mouth swaps looked worse: texture seams, dead eyes, and uncanny facial motion.
Breakthrough
A 90s cel-shaded mecha language made the constraints intentional. Rhubarb C++ drives acoustic A–H visemes while shot-reverse-shot staging preserves cinematic lead room.
This is the same pipe before it had a face. A headless CPU rasterizer established the 7× compile rate and the sub-10ms bus. The anime rig above is that runtime wearing a cel-shaded body.
AIWAKE://TELEMETRY-RENDER-NODE9:16
7.0× REAL-TIME
~2.2s compile for a 15-second reel
ZERO DIFFUSION
Headless CPU rasterization: Pillow matrix + CRT scanlines
SUB-10ms BUS
FSM transitions, token latency, and dialectic friction on-screen