Open source tools built to bridge creative intent and technical execution.

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.

Speculative generation with branch pruning: generate 4 candidates, score at checkpoints, keep the best. High-quality output at roughly half the compute cost.

TensorRT auto-conversion, FP8 TensorCore utilization, and memory management tuned for RTX 4090. Includes real-time GPU monitoring and VRAM optimization.

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.

AI framework nodes implementing 4-tier context memory (ECHO), deterministic expert routing (CSQMF), and 6-perspective analysis (PRISM). Built on ThinkingMachines research.

Connects AI assistants directly to Houdini sessions. Synapse handles communication, Engram maintains project memory. AI that understands your VFX context.

Load captured textures by hash, generate PBR maps, export to mod folders with correct naming. Full RTX Remix asset pipeline in ComfyUI.