SPIF

Style-Preserved Inference Flow — Technical Architecture

System Pipeline

🎤 Voice
User input
→
🧠 Prompt Engine
AI translation
→
🌀 ALDO
Dream logic
→
⚡ SPIF
Stylus · 72ms
→
🖼 TAESD
Encode · 1.76ms
→
📡 WebRTC
VP8 / H.264
→
🖥 Browser
2K live

Each frame feeds back as input → continuous dream evolution · Touch input injects noise at screen coordinates

Performance

72ms
Model Inference
1.76ms
Frame Encode
15 FPS
Framerate
2K
Resolution
~2.5B
Params (from 12B)
80%
Compression

Model Pipeline: SDXL → Stylus

SDXL 12B
Full model
→
LoRA Training
Lock artist style
→
Structural Pruning
Style-locked
→
Distillation
1-step / 4-step
→
Quantization
Vocab pruning
→
Stylus ~2.5B
Deploy-ready

By locking to one style, we prune ~80% of parameters while increasing fidelity.

Dream Seed Engine

311,000,000,000+
Unique prompt combinations · 7 bags × 200 entries · 30+ style presets with prefix/suffix wrapping

ALDO — Artistic Lucid Dream Orchestrator

🌊 Tadpole Stream
High chaos. The dream leads. Emergent surprise dominates user input.
✨ Lucid Stream
User leads. The dream follows. Precise control over generation.

Stack

LayerTechnology
ML FrameworkPyTorch, diffusers, CUDA
APIFastAPI + WebSocket
Streamingaiortc (WebRTC), VP8/H.264
TrainingRunPod A100
InferenceRunPod 4090

SPIF — Making AI art a live medium, not a prompt slot machine.