Jano: Adaptive Diffusion Generation with Early-stage Convergence Awareness

Yuyang Chen; Linqian Zeng; Yijin ZHou; Hengjie Li; Jidong Zhai

cs.CVFeb 28, 2026

Jano: Adaptive Diffusion Generation with Early-stage Convergence Awareness

Authors: Yuyang Chen, Linqian Zeng, Yijin ZHou, Hengjie Li, Jidong Zhai

Organizations: Shanghai Jiaotong University · Tsinghua University

Abstract

Diffusion models have achieved remarkable success in generative AI, yet their computational efficiency remains a significant challenge, particularly for Diffusion Transformers (DiTs) requiring intensive full-attention computation. While existing acceleration approaches focus on content-agnostic uniform optimization strategies, we observe that different regions in generated content exhibit heterogeneous convergence patterns during the denoising process. We present Jano, a training-free framework that leverages this insight for efficient region-aware generation. Jano introduces an early-stage complexity recognition algorithm that accurately identifies regional convergence requirements within initial denoising steps, coupled with an adaptive token scheduling runtime that optimizes computational resource allocation. Through comprehensive evaluation on state-of-the-art models, Jano achieves substantial acceleration (average 2.0 times speedup, up to 2.4 times) while preserving generation quality. Our work challenges conventional uniform processing assumptions and provides a practical solution for accelerating large-scale content generation. The source code of our implementation is available at https://github.com/chen-yy20/Jano.

Figures & tables

Figure 1: Jano foresees regional convergence through early-stage complexity recognition. Then it adaptively allocates computation resources, achieving $2.0\times$ speedup without perceptual loss.

Figure 2: Motivation Example with complexity and convergence pattern analysis ( $r=0.70,\ \rho=0.74$ ).

Figure 3: Velocity differences between both intra-frame similar points (A, B) and inter-frame similar points (A, D) remain approximately constant across timesteps.

Figure 4: Statistical validation and visualization of complexity-driven convergence-level categorization. (a) Convergence score exhibits significant correlation with complexity across the dataset ( $r=0.61,\ \rho=0.69,\ p<0.001$ ), naturally forming three distinct levels. Predicted complexity scores (b) are mapped to three convergence levels (c) through optimized thresholds.

Figure 5: Jano adaptively computes tokens of different convergence levels through an interleaved pipeline

Figure 6: Jano DiT with KV Cache. Cached Key-Value pairs are fetched and concatenates to maintain global sequence dependency in full attention.

Model	Resolution	Latent	Seq Length
Flux-1	1024x1024	[16,1,128,128]	4096
Wan-1.3B	832×480	[16,21,60,104]	32760
Wan-14B	720×1280	[16,21,90,160]	75600

Table 1: Evaluated Model Specifications

Figure 7: Jano advanced the Pareto frontier in the quality-latency tradeoff, achieving a better balance between generation quality and speed.

Method	Efficiency		Visual Quality
Method	Speedup $\uparrow$	Latency (s) $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	ImageQual $\uparrow$	SubConsist $\uparrow$
Flux-1 $(T=50)$	1 $\times$	11.37	-	-	-	-	-
PAB	1.43 $\times$	7.98	0.2866	0.7141	18.99	-	-
TeaCache	1.67 $\times$	6.83	0.1770	0.8014	20.14	-
ToCa	1.68 $\times$	6.78	0.1487	0.8223	21.91	-	-
Jano	1.86 $\times$	6.12	0.1310	0.8533	24.40	-	-

Table 2: Quantitative evaluation of inference efficiency and visual quality in generation models.

Figure 8: Early-stage complexity recognition on Flux and Wan 2.1. (a) Prompt-averaged accuracy versus warm-up step. (b) Accuracy distributions across prompts at early-stage (Flux: 6 steps; Wan: 7 steps).

Figure 9: Jano sensitivity analysis across different dynamic scenarios.

Figure 10: Detailed time breakdown of Jano generation under different token complexity scenarios.

Figure 11: Ablation study comparing random masking versus Jano ’s convergence-aware masking

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 12: Latte video generation. (a) Frames 10 and 11 from the Latte-generated clip; markers A–D indicate four probe locations—the same A–D indices used in panels (b) and (c). (b) Raw DDIM model output at A–D across reverse-diffusion timesteps, before any transformation. (c) The same signals after applying the epsilon-to-FM linear mapping (Eq. ( 16 )).

Model	Static		Moderate		Speedup $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$
Model	threshold	interval	threshold	interval	Speedup $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$
Flux-1	0.1	8	0.5	5	1.86	0.1310	0.8533	24.40
	0.2	10	0.5	3	2.00	0.1728	0.8182	23.11
	0.3	10	0.6	3	2.07	0.1768	0.8153	23.08
Wan-1.3B	0.1	3	0.3	2	1.45	0.0588	0.8799	24.31
	0.15	6	0.4	4	2.09	0.1001	0.8325	21.71

Appendix

Table 3: Jano parameter settings and performance.

Figure 13: Performance comparison between Jano and RAS on Stable Diffusion 3, showing generated images and latency-quality analysis. Test prompt: ”A photorealistic cute cat wearing a simple blue shirt, standing against a clear sky background.”

Figure 14: Comparison of JANO complexity maps and RAS token-importance maps at different diffusion timesteps.

Model	Base Memory	With KV Cache
Flux-1	32.47 GB	34.94 GB
Wan-1.3B	7.57 GB	28.18 GB
Wan-14B	65.02 GB	269.02 GB

Appendix

Table 4: Memory consumption comparison with different caching strategies.

Figure 15: Flux visualization comparison between original images (top) and Jano results (bottom).

Figure 16: Wan-1.3B visualization comparison between original videos (top) and Jano results (bottom).

Figure 17: Wan-14B visualization comparison between original videos (top) and Jano results (bottom).

Explore similar work

Just-in-Time: Training-Free Spatial Acceleration for Diffusion Transformers
Mar 11, 2026Wenhao Sun, Ji Li, Zhaoqiang LiuDiffusion Model Inference Acceleration Diffusion Models
University of Electronic Science and Technology of China · Capital Normal University
Predict to Skip: Linear Multistep Feature Forecasting for Efficient Diffusion Transformers
Feb 20, 2026Hanshuai Cui, Zhiqing Tang, Qianli Ma +3Diffusion Model Acceleration Diffusion Sampling
School of Artificial Intelligence, Beijing Normal University, Beijing, China · Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai, China · Shenzhen University of Advanced Technology, Shenzhen, China
DDiT: Dynamic Patch Scheduling for Efficient Diffusion Transformers
Feb 19, 2026Dahye Kim, Deepti Ghadiyaram, Raghudeep GaddeDiffusion Model Inference Acceleration Diffusion-Based Image Generation
Boston University · Amazon

Mar 11, 2026J/K move · Enter open · S save

Just-in-Time: Training-Free Spatial Acceleration for Diffusion Transformers

Wenhao Sun, Ji Li, Zhaoqiang Liu

University of Electronic Science and Technology of China · Capital Normal University

Diffusion Model Inference Acceleration Diffusion Models Diffusion Transformer

Open paper AskPDF arXiv

Diffusion Transformers have established a new state-of-the-art in image synthesis, but the high computational cost of iterative sampling severely hampers their practical deployment. While existing acceleration methods often focus on the temporal domain, they overlook the substantial spatial redundancy inherent in the generative process, where global structures emerge long before fine-grained details are formed. The uniform computational treatment of all spatial regions represents a critical inefficiency. In this paper, we introduce Just-in-Time (JiT), a novel training-free framework that addresses this challenge by acceleration in the spatial domain. JiT formulates a spatially approximated generative ordinary differential equation (ODE) that drives the full latent state evolution based on computations from a dynamically selected, sparse subset of anchor tokens. To ensure seamless transitions as new tokens are incorporated to expand the dimensions of the latent state, we propose a deterministic micro-flow, a simple and effective finite-time ODE that maintains both structural coherence and statistical correctness. Extensive experiments on the state-of-the-art FLUX.1-dev model demonstrate that JiT achieves up to a 7x speedup with nearly lossless performance, significantly outperforming existing acceleration methods and establishing a new and superior trade-off between inference speed and generation fidelity.

Abstract

Explore similar work

Just-in-Time: Training-Free Spatial Acceleration for Diffusion Transformers

Predict to Skip: Linear Multistep Feature Forecasting for Efficient Diffusion Transformers

DDiT: Dynamic Patch Scheduling for Efficient Diffusion Transformers