MiniMax-H3: native 1344×768 vs 896×512 draft + SoL-Refiner (H3 vs LTX-2.3) — blind

NVlabs' two-stage protocol on our production H3 (minimax-h3-both, turbo-8, isk-sol), through image-server over HTTP on one RTX 5090: 10 T2VA prompts, seed 4242, 5 s @24 fps. input = native H3 rendered directly at 1344×768 (124 frames, trimmed to 121 by stream copy). A / B = the same 896×512 H3 draft refined to 1344×768 by SoL-Refiner-LTX-2.5-for-MiniMax-H3 or SoL-Refiner-LTX-2.3-One-Step — which is which varies per clip. A and B share the draft's content and audio; the input is its own sample, so different content vs the input is expected. Pick a pair per card — the slider keeps the timestamp. 1:1 = native pixels. Judge on whether you would ship it, not on resemblance. Votes stay in this browser — use download votes.

RTX 5090, 32 GB · H3 int8 ConvRot + turbo LoRA · refiner DiTs int8 ConvRot refiner prompt = the H3 prompt's visual description LTX-2.3 refiner at frame_rate 16 (24 ghosts motion) timings = warm median of 9 (clip 1 paid the cold load); each stage on its own warm server — a worker hosting both would add a model swap

pipeline	stage 1	stage 2	total	vs native
native H3 turbo-8 @1344×768 (= input)	64.3 s	–	64.3 s	1.00×
H3 turbo-8 draft @896×512 → SoL-Refiner H3 (LTX-2.5) @1344×768	28.3 s	21.1 s	49.4 s	1.30×
H3 turbo-8 draft @896×512 → SoL-Refiner LTX-2.3 one-step @1344×768	28.3 s	12.7 s	41.0 s	1.57×
NVlabs' claim (GB200, one GPU per stage; baseline = 50-step H3 152.3 s)	4.06 s	1.56 s	5.64 s	27.0×