Stage 2 sparse attention

LTX-2.5 Sol-Attn on B200

Matched Dense and Sol-Attn generations across three official LTX-2.5 checkpoints and native high-resolution workloads.

Hardware: 4× B200
Parallel: Sequence parallel
Video VAE: Conv VAE
Seed: 42

01 / Visual comparison

PromptA woman in a yellow raincoat waits at a neon-lit intersection in heavy rain. Wide, medium, and low-angle shots preserve her identity while a late arrival approaches through reflected neon and rain.

Dense Sol-Attn

Workload

Checkpoint

Dense1.00×

Schedule: 1.0 / 1.25 / 1.5 + L0 dense1.25×

AttentionStage 2 speedupE2E speedupPSNRSSIM

Schedule: 1.0 / 1.25 / 1.5 + L0 dense1.58×1.25×34.190.9708

02 / Warm latency

Pipeline breakdown

One complete warmup generation, then one measured run.

Warm end-to-end latency by checkpoint

Stage 1 Stage 2 Others*

Others* is measured E2E latency minus Stage 1 and Stage 2, including VAE and pipeline overhead.

Checkpoint	Attention	Stage 1	Stage 2	VAE	E2E	S2 speedup	E2E speedup	PSNR	SSIM

03 / Fixed policy

What changes

Sol scope

All Stage 2 video self-attention layers and denoising steps.

Dense scope

Stage 1, audio self-attention, and all cross-attention.

Kernel policy

Global threshold routing with a workload-specific conservative schedule.

Controlled variables

Prompt, seed, schedule, checkpoint, VAE, and four-GPU SP are matched.