ProductBenchmarksDevelopers EnterprisePricingDownload ResearchFAQStatus GitHub Trust & Legal 한국어로 보기
Evidence record

VZK-WAN22-I2V-INT8-3090

Wan2.2 I2V (native int8 engine) · image-to-video

VEIZIK-MEASURED · SINGLE RUN 0 external reproductions 2026-07-17
Conditions

What was run

Condition · Value
GPURTX 3090 24GB
Power cap (W)420
Resolution848x480
Frames81
FPS16
Steps8
Precision profilenative int8 (sanitize) / int8+SmoothQuant
Warm runno
Measured

Results

Metric · Value
Wall time, 8-step (s)430
Wall time, SmoothQuant (s)600
Peak VRAM (GB)6.5
Reference (fp8 offload) peak VRAM (GB)17.7
Reference (fp8 offload) wall time (s)752
Reference (fp8 offload) temporal drift (%)-47
Reference (ComfyUI native fp8) wall time (s)1355
Reference (ComfyUI native fp8) peak VRAM (GB)18
Temporal drift, int8 (%)-11
Temporal drift, SmoothQuant (%)-8
Notes & caveats

Full context

INTERNAL same-machine head-to-head (NOT externally reproduced). Wan2.2 5s/81f/848x480 human CF. Native int8 path peaked 6.5GB vs fp8 CPU-offload reference 17.7GB (~2.7x less VRAM), finished ~430s (int8 8-step) vs 752s (fp8 ref) vs 1355s (ComfyUI native fp8), and held far better temporal stability (-8..-11% sharpness drift vs fp8 -47%). CAVEATS: step counts differ per row (int8 8-step vs 20-step production) so wall-time is NOT an identical-condition comparison; MAD-vs-fp8 is distance-from-fp8 not quality-loss; power cap unverified (INC-002); f32-truth validation int8 within 1.47%. NO fastest/lowest/SOTA claim; 0 external reproductions. Registered by [internal] session for the Head-to-head internal lab note. [2026-08-09 canon note] This row is the 8-step DISTILLED variant (cfg=1.0, no CFG). The commercial canon for Wan2.2-I2V-A14B is bymodel (gen_bymodel_plan.py CFG['veizik-wan22-i2v']): 480x832 (=832x480 landscape 480p) / 81f / 40 step / cfg(guidance) 3.5 / shift 5.0 -- kept SEPARATE, do NOT merge with the 40-step commercial table (5x step-count gap). See veizik-bench/suites/wan22_i2v_14b_3090.yaml 'distilled' variant; conflict unified at gen_bymodel_plan.CFG.

Reproduce

Run it yourself

Run the same fixed configuration on your card and open a result PR — see Run it yourself. Raw run manifests are attached to the internal evidence record and are released alongside external reproduction tooling.

← Back to all benchmarks