SoL-Refiner: Speed-of-Light One-Step Refinement for High-Resolution Video

Haozhe Liu; Tian Ye; Shuchen Xue; Yitong Li; Junsong Chen; Haopeng Li; Jincheng Yu; Duomin Wang

cs.CVSep 29, 2026

SoL-Refiner: Speed-of-Light One-Step Refinement for High-Resolution Video

Authors: Haozhe Liu, Tian Ye, Shuchen Xue, Yitong Li, Junsong Chen, Haopeng Li, Jincheng Yu, Duomin Wang, +4 more

Organizations: NVIDIA

Abstract

High-resolution video generation is expensive, as its cost grows rapidly with the number of spatiotemporal tokens. A practical alternative first generates a lower-resolution video and then applies a refiner, but conventional multi-step refinement introduces a second sampling bottleneck. We present SoL-Refiner, a one-step video refiner that transforms low-resolution model outputs into 4K videos with a single denoising step. Our three-stage recipe combines high-resolution continual training, reinforcement learning (RL) post-training, and a final one-step distillation. We introduce Refiner-Bench, a video refinement benchmark constructed from the outputs of different video generators, and use a shared-input protocol to compare refiners at approximately 2K output resolution. At 2K, the one-step SoL-Refiner outperforms all external refiners on the VBench and UniPercept averages, while at $3840\!\times\!2176$ it improves both metrics over the three-step LTX-2.3 Refiner. With the complete acceleration stack, SoL-Refiner achieves an $8.91\times$ speedup in refinement latency over the same baseline in our 2K latency setting.

Figures & tables

Panel of Figure 1 — Figure 1 : Overview of SoL-Refiner. (a) SoL-Refiner improves video resolution and visual quality while repairing artifacts in generated videos. (b) For a 5-second, $1344\!\times\!768$ , 24-fps video, the published full-resolution 50-step SGLang MiniMax H3 baseline [ 1 ] takes 152.3 s on one GB200, whereas distilled four-step $896\!\times\!512$ MiniMax H3 generation followed by the one-step SoL-Refiner takes 5.64 s using one GPU per stage ( $27.03\times$ ). The two-stage pipeline spends 4.06 s in H3 generation and 1.56 s in Refiner service. (c) Qualitative examples show how SoL-Refiner improves low-resolution videos generated by MiniMax H3.

Figure 2 : Training and inference pipeline of SoL-Refiner. Left, paired LQ/HQ videos and text/reference conditioning support continual training, RL post-training, and DMD-GAN distillation. Middle, the insets detail the truncated-

\sigma

flow path, reward feedback after a no-gradient rollout, and the DMD gradient and DiT discriminator that compress a multi-step teacher into a one-step student. Right, a tiny autoencoder (TAE) [ 14 ] replaces the full video variational autoencoder (VAE) to reduce encoding and decoding cost; latent upsampling and noise initialize the one-step refiner, and Sol Video Inference Engine (Sol-Engine) [ 15 ] executes one refiner NFE.

Method	Eval. Res.	# Steps	VBench $\uparrow$							$\Delta$ $\uparrow$
Method	Eval. Res.	# Steps	SC	BC	MS	DD	AQ	IQ	AVG
Resized Input	$1024\!\times\!576$	–	0.91805	0.95261	0.98354	0.70000	0.60621	0.63158	0.79866	0.00000
LingBot Stage-2 Refiner	$1920\!\times\!1088$	8	0.91715	0.95354	0.98398	0.74000	0.57762	0.63792	0.80170	0.00304
LTX-2.3 Refiner	$2048\!\times\!1152$	3	0.93004	0.95354	0.98873	0.70000	0.59541	0.65553	0.80388	0.00522
LTX-2.0 Refiner	$2048\!\times\!1152$	3	0.93036	0.95909	0.98656	0.72000	0.60679	0.64497	0.80796	0.00930
SEEDVR2	$2048\!\times\!1152$	1	0.91804	0.94783	0.97723	0.71333	0.60129	0.66337	0.80351	0.00485

Table 1 : VBench results on Refiner-Bench under a shared-input protocol. All methods receive the same

1024\!\times\!576

resized inputs. Resolution denotes the evaluated video resolution. SC, BC, MS, DD, AQ, and IQ denote subject consistency, background consistency, motion smoothness, dynamic degree, aesthetic quality, and imaging quality.

\Delta

is the AVG difference from the resized input. Best results are in bold ; second-best results are underlined .

Method	Eval. Res.	# Steps	UniPercept $\uparrow$
Method	Eval. Res.	# Steps	IAA	IQA	ISTA	AVG
Resized Input	$1024\!\times\!576$	–	59.3540	60.8943	42.0890	54.1124
LingBot Stage-2 Refiner	$1920\!\times\!1088$	8	58.9189	58.0827	42.1403	53.0473
LTX-2.3 Refiner	$2048\!\times\!1152$	3	61.2239	61.9438	45.2052	56.1243
LTX-2.0 Refiner	$2048\!\times\!1152$	3	60.5074	60.4902	42.3823	54.4600
SEEDVR2	$2048\!\times\!1152$	1	62.5692	64.4827	43.4057	56.8192

Table 2 : UniPercept results on Refiner-Bench under a shared-input protocol. All methods receive the same

1024\!\times\!576

resized inputs. Resolution denotes the evaluated video resolution. Best results are in bold ; second-best results are underlined .

Figure 3 : Performance on Refiner-Bench across output resolutions. We compare the three-step LTX-2.3 Refiner with the one-step SoL-Refiner using VBench AVG (left) and UniPercept AVG (right). Red

\Delta

labels report the relative improvement over LTX-2.3 Refiner. VBench scores are shown on a 0–100 scale. Both vertical axes use truncated ranges for readability. Higher is better.

Figure 4 : Quality and acceleration across base generators. (a,b) VBench and UniPercept averages for direct high-resolution generation, low-resolution generation, and low-resolution generation followed by SoL-Refiner. (c) End-to-end latency on one H100 GPU, split into base generation and refinement; speedups are relative to direct high-resolution generation. WAN uses 81 frames, 50 generation steps, and

832\!\times\!480\!\rightarrow\!1280\!\times\!704

refinement; WAN-1.3B’s direct baseline is

1280\!\times\!720

. Cosmos-Nano uses 189 frames, 35 generation steps, and 480p

\!\rightarrow\!1280\!\times\!720

refinement.

Training Stage	VBench $\uparrow$							UniPercept $\uparrow$
Training Stage	SC	BC	MS	DD	AQ	IQ	AVG	IAA	IQA	ISTA	AVG
Stage 1: Continual (Multi-Step)	0.92756	0.95361	0.98201	0.72000	0.60526	0.66484	0.80888	63.0807	64.5693	42.3359	56.6620
Stage 2: Post-Training (Multi-Step)	0.92988	0.94935	0.98027	0.72000	0.62344	0.69851	0.81691	66.7284	72.3977	44.8249	61.3170
Stage 3: DMD-GAN (One-Step)	0.92191	0.94591	0.98054	0.71333	0.60921	0.69196	0.81048	65.7322	70.6682	44.8445	60.4150

Table 3 : Ablation of the three-stage training recipe on Refiner-Bench. We report VBench and UniPercept scores after each stage. Best results are in bold ; second-best results are underlined .

Configuration	VBench $\uparrow$
Configuration	SC	BC	MS	DD	AQ	IQ	AVG
w/o regularization recipe	0.94438	0.95297	0.98074	0.62667	0.64757	0.71044	0.81046
w/o multiple reward models	0.93553	0.94765	0.98443	0.70000	0.62616	0.69111	0.81415
Ours	0.92988	0.94935	0.98027	0.72000	0.62344	0.69851	0.81691

Table 4 : Reward-learning ablation on Refiner-Bench. We report VBench scores. The “w/o multiple reward models” variant uses only HPSv3++. Best results are in bold ; second-best results are underlined .

Figure 5 : Qualitative effect of RL post-training. The rows show the low-quality input, the refiner without RL post-training, and the same refiner with RL post-training. Four aligned frames span 0.0 to 5.0 seconds. Each complete frame is followed by a magnified crop from the red box. In this selected example, RL post-training produces cleaner facial contours and sharper eyes, mouth, and jewelry across time.

Configuration	VBench $\uparrow$
Configuration	SC	BC	MS	DD	AQ	IQ	AVG
DMD-R	0.91385	0.93886	0.97779	0.66000	0.61929	0.69608	0.80098
RL then DMD	0.92191	0.94591	0.98054	0.71333	0.60921	0.69196	0.81048

Table 5 : Distillation ablation on Refiner-Bench. We compare DMD-R with RL followed by DMD using VBench. Best results are in bold ; second-best results are underlined .

Figure 6 : Latency analysis. Panel (a) compares Base (Multi-Step) with the same pipeline using Sol-Engine on one H100 GPU. The four groups report

2048\!\times\!1152

and

3840\!\times\!2176

outputs with 81 or 161 frames at 16 fps. Panel (b) reports component latency for

2048\!\times\!1152

outputs with 241 frames at 24 fps. Each stacked bar shows VAE encoding, denoising, and VAE decoding. The first bar is the three-step LTX-2.3 Refiner baseline. The second bar is the one-step SoL-Refiner, and the final two bars add TAE and Sol-Engine. Red arrows report the speedup between adjacent configurations. The final configuration reduces refinement latency from 57.461 to 6.447 seconds, an

8.91\times

speedup over the baseline. Lower latency is better.

Figure 7 : H3 acceleration on a single NVIDIA DGX Spark (GB10). (a) End-to-end latency for full-resolution four-step H3, its quantized variant, SoL-H3, and SoL-H3 with SoL-Refiner. Full-res denotes direct 768p generation; both SoL-H3 configurations generate at 384p and refine to 768p. Speedups use the 374-second full-resolution four-step baseline. (b) Stage-wise latency for the two SoL-H3 configurations. Stage 1 takes 21.60 seconds in both; Stage 2 decreases from 34.57 to 18.07 seconds with SoL-Refiner. The red annotation in panel (b) reports the 16.50-second reduction in total latency (29.4% relative to SoL-H3). The panels use separate linear scales.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Resolution	Configuration	GPU	VBench $\uparrow$							Latency (s) $\downarrow$
Resolution	Configuration	GPU	SC	BC	MS	DD	AQ	IQ	AVG	81 frames	161 frames
	Base (Multi-Step)	H100	0.9082	0.9364	0.9813	0.7067	0.5618	0.6713	0.7943	127.7	490
2K	w/ Sol-Engine	H100	0.9092	0.9365	0.9813	0.7200	0.5550	0.6601	0.7937	53.7	164
	Base (Multi-Step)	H100	0.9037	0.9367	0.9806	0.7267	0.5547	0.6719	0.7957	927.6	5123
4K	w/ Sol-Engine	H100	0.9043	0.9386	0.9817	0.7267	0.5511	0.6591	0.7936	284.8	1650

Appendix

Table 6 : Latency and quality comparison. Measurements use one H100 GPU. VBench metrics are evaluated on the corresponding 81-frame videos. Latency is reported in seconds per sample.

Source Generator	Videos	Native Res.	Aligned Crop	Main Refiner Input
WAN 2.1 1.3B	50	$832\!\times\!480$	$832\!\times\!448$	$1024\!\times\!576$
SANA-Video 2B	50	$1280\!\times\!736$	$1280\!\times\!704$	$1024\!\times\!576$
LTX-2 Stage 1	50	$768\!\times\!512$	$768\!\times\!512$	$1024\!\times\!576$

Appendix

Table 7 : Source preparation for the main 2K Refiner-Bench setting. Native and aligned resolutions are generator-specific, while every aligned video is resized to the shared

1024\!\times\!576

refiner input.

Content Group	Videos	Main Visual Challenges
Human daily activity	18	Hands, object interaction, water, and reflections
Animals	15	Fur, articulated motion, fast motion, and water
Nature and landscape	15	Water, weather, foliage, and particles
Sports and dance	15	Fast articulated motion, hands, and fine body geometry
Cinematic portraits	12	Faces, skin, hair, hands, and low-light detail
Urban and architecture	12	Structural lines, repeated geometry, and reflections

Appendix

Table 8 : Content groups in Refiner-Bench. The descriptions summarize the main visual challenges represented by each group.

Generator	Configuration	Resolution	VBench $\uparrow$	UniPercept $\uparrow$	Latency $\downarrow$	Overhead $\downarrow$
WAN-5B	Direct high-res	$1280\!\times\!704$	80.15	54.18	87.70	–
	Low-res base	$832\!\times\!480$	79.43	48.14	33.59	–
	Low-res + SoL-Refiner	$1280\!\times\!704$	81.22	54.49	39.75	6.16
WAN-1.3B	Direct high-res	$1280\!\times\!720$	79.73	55.08	272.69	–
	Low-res base	$832\!\times\!480$	80.40	52.42	72.55	–
	Low-res + SoL-Refiner	$1280\!\times\!704$	81.70	56.66	78.71	6.16

Appendix

Table 9 : Detailed quality and latency results across base generators. VBench is reported on a 0–100 scale. Latency and SoL-Refiner overhead are measured in seconds per sample on one H100 GPU. The overhead is the red Stage-2 segment in Figure 4 .

Generator	Configuration	SC $\uparrow$	BC $\uparrow$	MS $\uparrow$	DD $\uparrow$	AQ $\uparrow$	IQ $\uparrow$	AVG $\uparrow$
WAN-5B	Direct high-res	0.92258	0.95235	0.98199	0.71333	0.58522	0.65348	0.80149
	Low-res base	0.90244	0.92967	0.96584	0.82000	0.53374	0.61434	0.79434
	Low-res + SoL-Refiner	0.90250	0.94432	0.96426	0.82000	0.56545	0.67668	0.81220
WAN-1.3B	Direct high-res	0.92994	0.95195	0.97886	0.71333	0.57893	0.63057	0.79727
	Low-res base	0.92382	0.94247	0.97691	0.78000	0.57314	0.62767	0.80400
	Low-res + SoL-Refiner	0.92213	0.95378	0.97583	0.79333	0.59331	0.66346	0.81697

Appendix

Table 10 : Detailed VBench results across base generators. SC, BC, MS, DD, AQ, and IQ denote subject consistency, background consistency, motion smoothness, dynamic degree, aesthetic quality, and imaging quality. All metrics are reported on their original 0–1 scale.

Generator	Configuration	IAA $\uparrow$	IQA $\uparrow$	ISTA $\uparrow$	AVG $\uparrow$
WAN-5B	Direct high-res	60.6266	61.3029	40.5970	54.1755
	Low-res base	52.4973	53.1629	38.7562	48.1388
	Low-res + SoL-Refiner	59.1759	64.2542	40.0342	54.4881
WAN-1.3B	Direct high-res	62.5408	62.5327	40.1709	55.0815
	Low-res base	59.1791	58.7869	39.2797	52.4152
	Low-res + SoL-Refiner	63.1862	66.1531	40.6515	56.6636

Appendix

Table 11 : Detailed UniPercept results across base generators. IAA, IQA, and ISTA denote Image Aesthetics Assessment, Image Quality Assessment, and Image Structure and Texture Assessment, respectively.

Panel of Figure 8 — Figure 8 : Additional qualitative effects of RL post-training. Panels (a) and (b) show selected human and animal examples. Each panel compares the low-quality input, the refiner without RL post-training, and the same refiner with RL post-training at four aligned frames from 0.0 to 5.0 seconds. Each complete frame is followed by a magnified crop from the red box. RL post-training sharpens the side profile and cap edges in panel (a), and improves the eyes, facial contours, and fur texture in panel (b).

Explore similar work

PixelWizard: Towards Efficient High-Fidelity Video Generation at Ultra-Large Spatial Resolution
May 25, 2026Wenxue Li, Jingjing Ren, Peng Zhang +4Video Generation Visual Fidelity
The Hong Kong University of Science and Technology (Guangzhou) · MiLM Plus, Xiaomi Inc · The Hong Kong University of Science and Technology
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
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.

May 25, 2026, cs.CVJ/K move · Enter open · S save

PixelWizard: Towards Efficient High-Fidelity Video Generation at Ultra-Large Spatial Resolution

Wenxue Li, Jingjing Ren, Peng Zhang, Tian Ye+3

The Hong Kong University of Science and Technology (Guangzhou) · MiLM Plus, Xiaomi Inc · The Hong Kong University of Science and Technology

Video Generation Visual Fidelity High-Resolution Motion Coherence Prior Knowledge Integration

Open paper AskPDF arXiv

High-resolution video generation faces a coupled bottleneck of optimization instability and prohibitive computational costs. The massive expansion of the token sequence not only biases optimization toward local textures at the expense of global coherence, leading to structural collapse, but also imposes prohibitive training costs and severe inference latency. To address this, we propose PixelWizard, a framework that hierarchically decouples global structure modeling from fine-grained detail synthesis. PixelWizard first establishes a compact spatiotemporal anchor to concentrate dense structural priors, which then guides fine-grained generation at high resolution. This mitigates the local optimization bias to ensure structural stability without compromising high-frequency details. Leveraging this structural stability, we introduce Noise-Span Aligned Shortcut Training to break the inference bottleneck. By explicitly modeling the step size, this mechanism allows the model to traverse the generation trajectory with large steps. Crucially, we incorporate Exponential Index-Biased Sampling and Adaptive Noise-Span Calibration to align optimization with the shifted noise schedules of high-resolution grids, ensuring robust few-step inference without incurring the heavy overhead of distillation. Extensive experiments demonstrate that PixelWizard achieves superior visual quality while accelerating the generative sampling of native 2K/4K videos by over 10x.

Abstract

Explore similar work

PixelWizard: Towards Efficient High-Fidelity Video Generation at Ultra-Large Spatial Resolution

Dynamic Video Generation: Shaping Video Generation Across Time and Space

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