Sol-H3: Recursive Self-Improvement for MiniMax-H3 Inference Acceleration on Sol-Engine across Cloud and Edge

Yitong Li; Jincheng Yu; Junsong Chen; Haopeng Li; Shuchen Xue; Haozhe Liu; Ping Luo; Song Han

cs.AISep 28, 2026

Sol-H3: Recursive Self-Improvement for MiniMax-H3 Inference Acceleration on Sol-Engine across Cloud and Edge

Authors: Yitong Li, Jincheng Yu, Junsong Chen, Haopeng Li, Shuchen Xue, Haozhe Liu, Ping Luo, Song Han, +1 more

Organizations: NVIDIA Research, Efficient AI Team & Singapore Lab.

Abstract

Video diffusion models are rapidly scaling and exhibiting enhanced generation capabilities. Among these recent advancements, MiniMax-H3 stands out as a highly capable, production-level open-source model. However, its 33-billion parameters and multi-step iterative denoising process introduce substantial computational overhead. Consequently, their practical production is hindered by generation latency in the cloud deployment like NVIDIA-GB200, alongside strict memory limits that pose further challenges at the edge device like DGX-Spark. To address these diverse hardware bottlenecks from cloud to edge device, we present a full-stack inference pipeline that integrates efficient algorithmic design with optimized operator implementations. Algorithmically, we introduce a cross-resolution two-stage generation scheduler that exploits the step-wise nature of diffusion: early low-resolution steps rapidly establish the global layout, while later high-resolution steps focus refinements of local and perceptual details. These stages are connected by a learned latent-to-latent mapping module, completely eliminating the computationally expensive VAE decode-reencode cycle for resolution transferring cross different resolutions. For operator implementation, we deploy a Recursive Self-Improvement (RSI) loop that searches kernel fusions and memory layouts, evaluating latency together with numerical agreement. Together, these optimizations deliver up to 30x end-to-end speedup and 20% lower memory: a 5-second 1344x768 video with audio is generated 3.5x faster than real time on an 8xGB200 node, and in under a minute fully memory-resident on a single DGX Spark.

Figures & tables

Figure 1 : Sol-H3: two-stage video generation with recursive self-improvement. A cross-resolution two-stage inference pipeline reduces the overhead of full-resolution denoising. Recursive Self-Improvement (RSI) profiles, optimizes, and verifies implementations under restricted generation settings, producing optimized kernels and effective memory layouts. The combination of this inference algorithm design and efficient operator implementation enables MiniMax-H3 to achieve approximately 30x acceleration and reduce GPU memory consumption by nearly 20%.

Figure 2 : System overview. We accelerates the generation process through a two-stage pipeline and a latent-to-latent connector. Stage 1 generates content at 672 $\times$ 384 for 4-steps; an H3 latent upscaler doubles spatial resolution, and a separate adapter translates the result into LTX latent coordinates; Stage 2 refines at 1344 $\times$ 768 for 3-steps.

Panel of Figure 3 — Figure 3 : Cross-VAE latent translation. Left: both frozen encoders process the same source video; latent and decoded-pixel supervision train the residual translator. The spatial upscaler is separate from this training task. Right: conversion cost for an earlier checkpoint and held-out quality for the final checkpoint of the same architecture. The profile excludes the upscaler and refiner ( Sec. 6.5 ).

Figure 4 : Attention computation with reference KV cache. Left: full self-attention over all tokens, the native inference mode of MiniMax-H3 [ 4 ] . Right: after the first denoising step, image and video reference query rows are skipped (gray). Text and generation queries attend to cached reference K,V (cross-attention) and to each other (self-attention).

Figure 5 : Latency profiles of Sol-H3 inference. The same 5 s

1344\times 768

workload on one GB200 (top), one DGX Spark (middle) and one RTX 5090 (bottom), each against a full-resolution 49-step MiniMax-H3 baseline [ 35 , 36 ] . The optimized request completes in 6.226 s on GB200, 56.17 s on Spark and 39.80 s on RTX 5090, a speedup of more than 20

\times

in every case. The profiles differ in where the time goes: Stage-1 generation dominates on GB200 at 60.6%, Stage-2 refinement dominates on Spark at 44.0%, and on RTX 5090 prompt encoding is the largest single block at 34.1%, since Qwen is loaded and released per request under CPU offload.

Panel of Figure 6 — Figure 6 : Qualitative comparison between Sol-H3 and the baseline. Sol-H3(right) drafts at a lower resolution and refines back to the target resolution. While preserving scene structure, subject identity and motion continuity with minor visual differences, Sol-H3 delivers a significant speedup of more than 20 $\times$ [ 35 ] .

GPUs	5 s output			10 s output			15 s output
GPUs	Base H3	Sol-H3	Speedup	Base H3	Sol-H3	Speedup	Base H3	Sol-H3	Speedup
1	129.898	9.475	13.71 $\times$	376.942	24.309	15.51 $\times$	746.885	44.958	16.61 $\times$
4	35.328	2.530	13.97 $\times$	100.440	6.212	16.17 $\times$	194.930	11.471	16.99 $\times$
8	18.250	1.434	12.73 $\times$	50.660	3.350	15.12 $\times$	99.513	6.062	16.41 $\times$

Table 1 : Sol-H3 latency on GB200s. Baseline of MiniMax-H3 and Sol-H3 latency for 5, 10, and 15 s outputs on NVIDIA GB200 GPUs. It delivers more significant speedups with longer duration of generated videos.

Condition	Eff. ref. tokens	KV cache speedup
No reference	0	—
1 image	14,344	1.14 $\times$
2 images	28,432	1.35 $\times$
3 images	42,776	1.56 $\times$
1 video	31,792	1.41 $\times$
2 videos	63,584	1.92 $\times$

Table 2 : Additional cost introduced by reference tokens and acceleration of reference KV caching. Left: DiT time relative to reference-free generation against the effective reference-token count. The latency increases dramatically as the reference materials growing. Right: effective reference-token counts for every condition and the speedup obtained by reference KV caching strategy, which reducing the computation overhead up to 2

\times

Panel of Figure 7 — Figure 7 : Reference KV cache: qualitative comparison. Each row displays a single frame from the input reference video on the left, two frames from the full-compute teacher baseline in the middle, and two frames from our cached reference KV generation on the right. It demonstrates that this approximation only introduces minor visual differences while preserving visual quality and instruction alignment

Checkpoint	Latent MSE $\downarrow$	PSNR (dB) $\uparrow$	SSIM $\uparrow$
Earlier decoder-aware checkpoint	0.11058	29.413	0.8763
Final checkpoint	0.09783	30.618	0.8962

Table 3 : Held-out latent translation quality. Both checkpoints use the same 194.76M-parameter architecture. Metrics compare decoded adapter predictions with decoded paired LTX teacher latents on the same 256 test clips; they do not include the refiner.

Conversion	Params (M)	Latency (ms)	Peak (GiB)	Speedup
Full VAE round trip	2,742.47	12,748.6	13.788	$1.0\times$
Tiny AutoEncoder round trip	23.11	201.9	23.846	$63.1\times$
Latent adapter	194.76	63.1	0.785	$\mathbf{202.0\times}$

Table 4 : Isolated H3-to-LTX conversion cost. One H100 80GB, BF16, batch one, 192 source frames at

1344\times 768

. Latency is the median synchronized wall time after warmup; memory is peak allocated device memory for each conversion arm. The adapter includes geometry alignment but excludes the separate

\times 2

upscaler.

LoRA execution	Time (s) $\downarrow$	Bitwise agreement
Merged BF16 weights	24.785	No
Native separate branches	26.283	Reference
Consumer fusion	25.589	Yes (tested inputs)

Table 5 : LoRA precision and execution cost. Median pipeline latency for eight denoising evaluations on four GB200 GPUs. Numerical agreement is relative to native separate branches under the same other acceleration settings. Merged weights provide a speed reference with different outputs.

Panel of Figure 8 — Figure 8 : LoRA execution changes the generation trajectory. The same eight-step checkpoint, prompt, seed, and initial noise are evaluated on four GB200 GPUs at $1344\times 768$ , with the other acceleration settings fixed. The upper rows compare identical timestamps; the bottom row magnifies the butterfly burst at the indicated times. All images come from lossless frames exported before video encoding. Native branches and consumer fusion match byte for byte in the decoded video and audio; merging the weights changes the scene and introduces grid-like patterns around the butterflies in this example.

Explore similar work

Sol Video Inference Engine: Agent-Native Full-Stack Acceleration Framework for Efficient Video Generation
Jun 21, 2026Yitong Li, Junsong Chen, Haopeng Li +6Video Diffusion Models
NVIDIA Research, Efficient AI Team & Singapore Lab.
Dynamic Video Generation: Shaping Video Generation Across Time and Space
May 20, 2026Shikang Zheng, Jingkai Huang, Jiacheng Liu +5Video Generation
Shanghai Jiao Tong University · South China University of Technology · Tsinghua University
Efficient Video Diffusion Models: Advancements and Challenges
Apr 17, 2026Shitong Shao, Lichen Bai, Pengfei Wan +2Video Diffusion Models Diffusion Models
The Hong Kong University of Science and Technology (Guangzhou), China · Kling Team, Kuaishou Technology, China · The Hong Kong University of Science and Technology, Hong Kong

Jun 21, 2026, cs.CVJ/K move · Enter open · S save

Sol Video Inference Engine: Agent-Native Full-Stack Acceleration Framework for Efficient Video Generation

Yitong Li, Junsong Chen, Haopeng Li, Haozhe Liu+5

NVIDIA Research, Efficient AI Team & Singapore Lab.

Video Diffusion Models

Open paper AskPDF arXiv

Modern video diffusion models achieve higher generation quality through scaling, but this also increases inference cost. Although many acceleration methods have been proposed, a central challenge is that the most effective acceleration strategy is highly instance-specific: a recipe that works well for one combination of model, hardware, and inference configuration often does not transfer to another. Different models vary in architecture, numerical sensitivity, and attention concentration patterns. Inference settings differ in spatial and temporal resolution and video duration, while hardware platforms differ in memory hierarchy, supported numerical formats, and kernel throughput. These factors create a large tuning space, making manual performance engineering costly. We present Sol Video Inference Engine, an agentic, native, training-free acceleration framework for video diffusion models. It organizes five broadly applicable techniques, cache, sparse attention, token pruning, quantization, and kernel fusion, into an agentic acceleration stack for instance-specific optimization. For a concrete deployment target defined by a model, hardware platform, and serving configuration, parallel skill agents optimize the implementation of each technique, an agent integrator composes them into a global acceleration stack, and a human validator provides feedback on generation quality. We instantiate this workflow on three video models with different sizes and architectures: 64B Cosmos3-Super, 22B LTX-2.3, and 2B SANA-Video. With little human effort, the full stack achieves more than 2x end-to-end acceleration while maintaining near-lossless VBench quality, demonstrating the effectiveness of the agent framework for video diffusion acceleration.

Abstract

Explore similar work

Sol Video Inference Engine: Agent-Native Full-Stack Acceleration Framework for Efficient Video Generation

Dynamic Video Generation: Shaping Video Generation Across Time and Space

Efficient Video Diffusion Models: Advancements and Challenges