ADAPTIVE VIDEO GENERATION ENGINE // FREEVIDEO-RS

FREEVIDEO-RS // BOUNDED VRAM PLANNER & TWO-PASS DMD8 STUDIO

Hardware-aware adaptive execution planner for H3 3D-VAE ($17n+5$ temporal alignment), two-pass DMD8 latent upscaling (vdn8-lbh-tail2), longest-gap bisection SafeTensors 50-block memory interleaving, and SQLite v6 telemetry forecasting. Delivers a 60.8x speedup over Python/CUDA probe planners with zero GPU context overhead.

PLANNING SPEEDUP

60.8x

TWO-PASS SAVINGS

60.0%

RESIDENT BLOCKS

21 / 50

VIDEO TOKENS

72,576

1. H3 VAE CANVAS GEOMETRY (17n+5)

RESOLUTION PRESET1344x768

REQUESTED FRAMES (Snaps to 17n+5)243

2. HARDWARE VRAM & ARCHITECTURE

GPU ARCHITECTUREada

DETECTED VRAM CAPACITY (GiB)16 GiB

SYSTEM HOST RAM (GiB)64 GiB

3. TWO-PASS DMD8 SCHEDULE (VDN8-LBH-TAIL2)

PASS 1 BASE STEPS (Half-Res)8

PASS 2 REFINE STEPS (Full-Res)2

STAGE // 50-BLOCK LONGEST-GAP BISECTION RESIDENCY & TWO-PASS DMD8 LATENT PIPELINE POLICY: ADMITTED // NATIVE-FP8

PASS 1 -> PASS 2

672x384 -> 1344x768

ATTENTION HEAD CHUNK

16 HEADS (SAGE2)

PINNED HOST CACHE

12.54 GB

RESIDUAL OFFLOAD

DISABLED (IN-VRAM)

PLANNING LATENCY

0.021 ms (60.8x)

EMPIRICAL PERFORMANCE BENCHMARKS // FREEVIDEO-RS VS PYTHON/CUDA PLANNER

Subsystem / Operation	Legacy Python Planner	freevideo-rs (Rust)	Speedup	Peak Memory
H3 VAE (17n+5) Geometry & Token Grid (10k calls)	48.20 ms	0.62 ms	77.7x	0.2 MB (vs 148 MB)
Two-Pass vdn8-lbh-tail2 Schedule + Savings (10k calls)	64.50 ms	0.95 ms	67.9x	0.3 MB (vs 162 MB)
SafeTensors Header & Bisection Interleaver (50 blocks)	310.40 ms	5.10 ms	60.9x	1.8 MB (vs 420 MB)
24-Tier GPU Capacity Grid Simulation (simulate_capacity_grid)	845.00 ms	14.20 ms	59.5x	2.4 MB (vs 610 MB)
End-to-End Adaptive Resource Policy & SQLite History Check	1,268.10 ms	20.87 ms	60.8x	2.4 MB (vs 610 MB)