Accelerating LTX-2.5 with Sol Engineup to 4.68× multi-step pipeline speedup & 1.90× distilled pipeline speedup

LTX-2.5 is Lightricks' 22B diffusion transformer for generating video and synchronized audio together. Its released inference package combines the 22B DiT with a convolutional video VAE, an audio VAE, and a Gemma text encoder. For the 1536×1024 benchmark configuration used in this article, the pipeline runs a 30-step guided first stage at 768×512, upsamples the latent representation, and finishes with a two-step distilled stage at 1536×1024.

To accelerate video generation without changing the Multi-step Pipeline's high-quality operating point, we optimize the full inference stack at three complementary levels.

1. Diffusion Cache. The Multi-step Pipeline executes 30 steps in Stage 1, leaving redundancy across adjacent diffusion-denoising steps. We apply First Block Cache to reuse stable residuals and skip unnecessary forwards, delivering nearly 50% higher Stage 1 performance while preserving the overall visual result.
2. Parallelism Optimization. Because the first-stage token sequence is relatively compact, sequence parallelism leaves too much communication overhead. At the same time, the high-quality Multi-step Pipeline contains four CFG branches, which naturally allows the guidance workload to be distributed across four GPUs with substantially less communication.
3. Kernel optimization. Cache-aware compilation and kernel fusion reduce launch overhead and avoid unnecessary memory traffic. Block 0 remains the cache decision point, while cache misses execute the remaining compiled transformer blocks.

Performance across different resolutions and durations

Warm end-to-end latency on 4× B200 at BF16 precision, with a 30-step first stage and a 2-step second stage.

Video comparison

Setting: 4× B200 · BF16 precision · 30-step first stage · 2-step second stage · same prompt and seed.

Our setting uses the NVFP4 Distilled Pipeline, with an 8-step Stage 1 and a 3-step Stage 2. Stage 1 operates on a relatively compact token sequence, and its denoising schedule is already well optimized; consequently, our focus shifts to Stage 2. At 4K for five seconds, attention becomes the dominant cost in this higher-resolution refinement stage. We therefore apply Sol-Attn, a training-free sparse attention method, to target this bottleneck without additional training or calibration.

Stage 2 Optimization — Core Technology: Sol-Attn

At high resolution and long duration, with audio tokens interleaved into the same attention, sequence length dominates the cost. Attention maps in video diffusion are also extremely sparse: most query blocks attend meaningfully to only a small fraction of key blocks. This structure calls for an efficient sparse attention method that reduces computation without a catastrophic loss in quality.

Sol-Attn is training-free and calibration-free. It folds block selection into the online-softmax pass itself, avoiding a separate routing stage:

• Query-dependent thresholding. Instead of a fixed top-k or top-p, each query row derives a threshold from its block proxy-score statistics, τᵢ = μᵢ + βσᵢ. One β value controls density across layers, steps, and resolutions without learning or offline calibration.
• On-the-fly routing. Pooled keys stream chunk-wise and are compared with the threshold during softmax accumulation. No proxy map is materialized, keeping routing overhead low at long sequence lengths.
• Proxy-score reuse. Rejected blocks are approximated with a Taylor expansion around the row-wise mean score, recovering probability mass that hard selection discards and preserving quality at high sparsity.

We use Sol Engine to optimize video inference across different hardware and configurations:

• Hardware: NVIDIA DGX Spark, B200, and other GPU platforms.
• Configurations: different resolutions and durations, including 4K and 1080p workloads.

For each deployment target, Sol Engine combines:

• Parallelism optimization
• Diffusion cache
• Kernel fusion
• Sparse attention

Sol Engine closes that gap by search instead of by hand. It runs one optimization search per technique family, concurrently — kernel, sparse attention, cache — each sweeping its own knobs against the real checkpoint on the real GPUs. The results cannot simply be concatenated, because the techniques interact: caching changes which steps sparse attention has to serve. So the surviving candidates go to an integrator that searches over combinations rather than over knobs, with a human validator signing off on the output video.

Citation

If this work is useful to you, please cite Sol Engine and Sol-Attn.

@misc{li2026solvideoinferenceengine,
title = {Sol Video Inference Engine: Agent-Native Full-Stack Acceleration Framework for Efficient Video Generation},
author = {Yitong Li and Junsong Chen and Haopeng Li and Haozhe Liu and Jincheng Yu and Ligeng Zhu and Ping Luo and Song Han and Enze Xie},
year = {2026},
eprint = {2606.23743},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2606.23743},
}

@misc{li2026solattnacceleratingvideogeneration,
title = {Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification},
author = {Haopeng Li and Yitong Li and Junsong Chen and Tian Ye and Haozhe Liu and Jincheng Yu and Duomin Wang and Ruihua Zhang and Zeke Xie and Enze Xie and Song Han},
year = {2026},
eprint = {2607.24027},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2607.24027},
}