JetBlock v4
Full Nemotron 3 Optimization
Production-grade optimization for Nemotron 3's hybrid architecture—Mamba-2 state-space models, MoE routing, and attention layers. V2 only handled 12% of layers. V4 optimizes all 52.
GitHubOpen source tools built to bridge creative intent and technical execution.
Nemotron 3 hybrid architecture
Full Nemotron 3 Optimization
Production-grade optimization for Nemotron 3's hybrid architecture—Mamba-2 state-space models, MoE routing, and attention layers. V2 only handled 12% of layers. V4 optimizes all 52.
GitHubNon-Linear Inference Optimizer
Speculative generation with branch pruning: generate 4 candidates, score at checkpoints, keep the best. High-quality output at roughly half the compute cost.
GitHubHardware-Specific Acceleration
TensorRT auto-conversion, FP8 TensorCore utilization, and memory management tuned for RTX 4090. Includes real-time GPU monitoring and VRAM optimization.
GitHubBatch-invariant inference
Guaranteed Reproducibility
Temperature=0 doesn't guarantee determinism—batch-size variance does. These nodes enforce batch_size=1 with fixed RNG states. Same seed, same prompt, identical output. Always.
GitHubECHO · CSQMF · PRISM
AI framework nodes implementing 4-tier context memory (ECHO), deterministic expert routing (CSQMF), and 6-perspective analysis (PRISM). Built on ThinkingMachines research.
GitHubHoudini ↔ ComfyUI ↔ RTX Remix
AI-Houdini Bridge
Connects AI assistants directly to Houdini sessions. Synapse handles communication, Engram maintains project memory. AI that understands your VFX context.
GitHubBidirectional Pipeline
Pass renders from Houdini to ComfyUI for AI enhancement, receive processed results back. Automated batch processing for production pipelines.
GitHubTexture Pipeline Integration
Load captured textures by hash, generate PBR maps, export to mod folders with correct naming. Full RTX Remix asset pipeline in ComfyUI.
GitHubInterested in collaboration? Get in touch