One-sentence Summary

Beihang University, Nanyang Technological University, and Chinese Academy of Sciences, among others, propose MrFlow, a training-free multi-resolution acceleration strategy for pretrained flow-matching models that first generates main structure at low resolution, then applies pixel-space super-resolution via a lightweight GAN, injects low-strength noise for high-frequency resampling, and finally refines details at high resolution; evaluated on FLUX.1-dev and Qwen-Image, MrFlow achieves a $10\times$10× end-to-end speedup with a OneIG gap $<1\%$<1% relative to unaccelerated sampling, and when combined with timestep distillation, reaches up to $25\times$25× acceleration, significantly outperforming existing training-free approaches.

Key Contributions

• MrFlow is a training-free multi-resolution acceleration pipeline for pretrained flow-matching models that generates an image structure at low resolution, upsamples in pixel space via a lightweight GAN-based super-resolution model, injects low-strength noise, and performs a single high-resolution refinement step.
• On FLUX.1-dev and Qwen-Image, MrFlow achieves over 10× end-to-end speedup while keeping the OneIG-Bench loss within a 1% gap relative to native inference, outperforming other training-free acceleration strategies without any training or runtime dynamic statistics.
• The method allows flexible configuration for speedup–quality tradeoffs and can be directly combined with pretrained timestep distillation models to yield an orthogonal compounded speedup of up to 25×, with no additional training.

Introduction

Modern diffusion models built on Transformers and flow matching have become the standard for high-quality image generation, but this quality comes with substantial computation. For example, a leading 20-billion-parameter model can take 47 seconds to generate a single 1024×1024 image on an NVIDIA A100. Researchers have pursued several acceleration strategies: timestep distillation slashes the number of sampling steps to just 1–4 but requires expensive model retraining; feature caching and token pruning offer training-free speedups but rarely exceed 4×. Training-free multi-resolution methods that exploit spatial redundancy can surpass 5× speedup, yet they typically depend on runtime dynamic identification of image regions, and performing upsampling directly in the latent space often introduces blurring or artifacts. The authors leverage this multi-resolution insight but address its shortcomings. They propose MrFlow, a training-free staged pipeline that first generates the image’s coarse structure at low resolution in the latent space, then uses a lightweight GAN-based super-resolution network in pixel space to upscale while preserving structure, applies low-strength noise to weaken any erroneous high-frequency details, and finally refines the result with a fast, few-step high-resolution sampling. This design exploits the fact that structure emerges quickly at low resolution and that the flow trajectory is straighter near clean images, enabling a total acceleration of more than 10× while keeping generation quality within 1% of native inference. MrFlow requires no training, no dynamic statistics, and can be directly paired with existing timestep-distillation models to reach a combined speedup of over 25×.

Method

The authors propose MrFlow, a multi-resolution generation pipeline that accelerates text-to-image diffusion by decoupling the generation of global structure from detail refinement. The key insight is that the low-resolution stage determines the overall layout and semantic composition of the image, while the high-resolution stage need only adjust local high-frequency details inherited from an intermediate super-resolution step. By performing the bulk of computation at low resolution and limiting the high-resolution phase to one denoising step with minimal noise, MrFlow achieves substantial speedups without sacrificing output quality.

The pipeline consists of five sequential stages: low-resolution latent-space sampling, VAE decoding, pixel-space super-resolution, VAE encoding with low-strength noise injection, and a final high-resolution latent-space sampling followed by VAE decoding.

Low-resolution structure generation. The initial stage generates a clean latent at a reduced spatial size. Starting from Gaussian noise $\mathbf{z}_1^{\mathrm{LR}} \in \mathbb{R}^{C \times H_L \times W_L}$z1LR​∈RC×HL​×WL​, a pretrained flow matching velocity network $\mathbf{v}_\theta$vθ​ is integrated using $K_L$KL​-step Euler discretization of the rectified flow ODE, yielding $\mathbf{z}_0^{\mathrm{LR}} = \Phi_{\mathbf{v}_\theta,\mathbf{c}}^{K_L}(\mathbf{z}_1^{\mathrm{LR}})$z0LR​=Φvθ​,cKL​​(z1LR​), where $\mathbf{c}$c is the text-condition embedding. The authors set $K_L = 12$KL​=12 by default. Decoding through the pretrained VAE produces a low-resolution pixel-space image $\mathbf{x}_{\mathrm{LR}}$xLR​. This stage already captures the complete global structure and semantic content of the prompt, albeit with the softness typical of a low-resolution render.

Two factors enable the efficiency of this stage. First, inference cost scales roughly linearly with the number of image tokens, so halving each spatial dimension yields roughly a $4\times$4× speedup per step. Second, low-resolution generation requires fewer sampling steps to converge, attributed to both higher text-to-image semantic utilization at low resolution and a shorter ODE path corresponding to the low-frequency skeleton.

Pixel-space super-resolution. The low-resolution image is upsampled in the pixel domain using a pretrained Real-ESRGAN network $\mathbf{x}_{\mathrm{SR}} = U(\mathbf{x}_{\mathrm{LR}})$xSR​=U(xLR​). This GAN-based super-resolver faithfully preserves the global layout while injecting plausible high-frequency details. Operating in the pixel space is deliberate: natural-image priors for edges and textures reside in pixels, and subsequent re-encoding by the VAE acts as a lightweight regularizer that attenuates out-of-distribution high-frequency artifacts. GAN-based upscaling is chosen over interpolation or regression-based methods because its outputs are sharper and contain only local high-frequency residuals that are amenable to downstream refinement; interpolation-based methods, by contrast, would require stronger noise to overcome persistent blur, undermining the pipeline’s speed advantage.

Low-strength noise for high-frequency resampling. The super-resolved image is encoded back into the high-resolution latent space $\mathbf{z}_0^{\mathrm{SR}} = \mathcal{E}(\mathbf{x}_{\mathrm{SR}})$z0SR​=E(xSR​). At this point, the latent may contain minor high-frequency artifacts such as character-stroke shifts introduced by the GAN. To correct these while preserving the low-frequency structure, a small amount of flow-matching noise is injected: $\mathbf{z}_t^{\mathrm{HR}} = (1 - \sigma_t)\mathbf{z}_0^{\mathrm{SR}} + \sigma_t \boldsymbol{\epsilon}$ztHR​=(1−σt​)z0SR​+σt​ϵ, with $\sigma_t \in [0.1, 0.15]$σt​∈[0.1,0.15]. The noise level is chosen such that the signal-to-noise ratio in the low-frequency band remains high, while the high-frequency band’s SNR is lowered enough for the subsequent denoising step to resample it according to the flow prior. A theoretical lower bound $\sigma_t^\star = \frac{\sqrt{\lambda_{\mathrm{hf}}}}{1 + \sqrt{\lambda_{\mathrm{hf}}}}$σt⋆​=1+λhf​​λhf​​​ is derived from the high-frequency power $\lambda_{\mathrm{hf}}$λhf​ of the clean latent, confirming that the typical $\sigma_t$σt​ values are sufficient when the super-resolution residual errors are predominantly high-frequency. This insight eliminates the need for heavy noise that would erase useful SR detail and demand many refinement steps.

High-resolution detail refinement. The noised high-resolution latent $\mathbf{z}_t^{\mathrm{HR}}$ztHR​ is denoised via the same velocity network with $K_H$KH​ Euler steps: $\mathbf{z}_0^{\mathrm{HR}} = \Phi_{\mathbf{v}_\theta,\mathbf{c}}^{K_H}(\mathbf{z}_t^{\mathrm{HR}})$z0HR​=Φvθ​,cKH​​(ztHR​). By default, $K_H = 1$KH​=1. Because the noised latent is already very close to the clean manifold, the velocity field magnitude is nearly flat around that point, making a single-step discretization highly accurate. The final image $\mathbf{x}_{\mathrm{HR}} = \mathcal{D}(\mathbf{z}_0^{\mathrm{HR}})$xHR​=D(z0HR​) thus retains the global structure established at low resolution while the high-frequency artifacts of the GAN are corrected. The entire pipeline delivers high-fidelity results at the target resolution with a total of only 12 low-resolution steps, one noise injection, and one high-resolution step, translating to a latency reduction of more than half compared to direct high-resolution generation with comparable visual quality.

Experiment

The evaluation setup on FLUX.1-dev and Qwen-Image-20B at 1024x1024 resolution compares MrFlow with training-free and training-dependent acceleration methods. MrFlow maintains image quality at aggressive speedups of 8x to 10x where feature-caching and other multi-resolution approaches fail, and a single high-resolution denoising step suffices for detail refinement. Ablations confirm that more low-resolution steps improve metrics while one refinement step is sufficient, and Real-ESRGAN yields the best visual fidelity. MrFlow matches training-dependent distillation accuracy and synergizes with distilled models for up to 25x speedup without extra training.

Training-free acceleration methods on FLUX.1-dev show that token pruning (ToMA) yields negligible speedup and silently collapses image quality despite stable metrics. Techniques like Teacache and DB-Taylor achieve around 4.5× speedup with only minor drops in Geneval and DPG, while further speedups from RALU and SPEED incur larger metric declines, illustrating a clear tension between efficiency and fidelity. ToMA attains at most 1.13× speedup and its OneIG-En score plunges from 0.44 to 0.21, reflecting severe image collapse not captured by Geneval. Teacache and DB-Taylor provide speedups near 4.5× while retaining Geneval scores within 0.03 of the baseline and DPG above 82.6, offering a favorable quality–speed trade-off.

On FLUX.1-dev, the training-free MrFlow model with 12+1 steps attains an 8.25× speedup while preserving quality close to the 4-step training-dependent SenseFlow method. When combined with the Pi-Flow distillation model using only 4+1 steps, MrFlow† reaches an 11.3× speedup, outperforming standalone Pi-Flow in both speedup and Geneval score. The training-required LSSGen method exhibits a noticeable generation quality drop at 3.93× speedup despite similar automatic metric scores. MrFlow (12+1 steps) achieves an 8.25× speedup without any training, with a Geneval score comparable to the 4-step SenseFlow method. MrFlow† (4+1 steps, reusing pre-trained distillation weights) delivers a higher 11.3× speedup, surpassing the native Pi-Flow (4 steps) in both speedup and Geneval. LSSGen at 3.93× speedup shows little metric degradation but the generated images collapse significantly according to the paper. Pi-Flow maintains the best DPG and OneIG-En scores among all methods, while MrFlow† reaches the highest Geneval and the fastest speedup.

On Qwen-Image at 1024×1024, automatic metrics for interpolation, SwinIR, OSEDiff, and Real-ESRGAN differ only slightly, but visual quality reveals clear ranking: interpolation and SwinIR remain blurry, OSEDiff adds character artifacts, while GAN-based Real-ESRGAN provides the best blend of sharpness, semantic accuracy, and efficiency. A fast Real-ESRGAN variant using a 12+2 step configuration reaches 14× speedup with metric scores comparable to or better than the 12+1-step setting. All evaluated super-resolution methods achieve tightly clustered Geneval, DPG, and OneIG-Bench scores. Real-ESRGAN with a 12+2-step configuration yields a 14× speedup while maintaining a Geneval of 0.87 and slightly higher DPG compared to its 12+1-step counterpart. Interpolation and SwinIR produce blurry enlargements, whereas OSEDiff introduces character inaccuracies; Real-ESRGAN best balances clarity and semantic fidelity.

The evaluations first examine training-free acceleration on FLUX.1‑dev, where caching techniques such as Teacache and DB-Taylor achieve around 4.5× speedup while preserving image quality, whereas token pruning yields negligible speedup and causes severe visual collapse that automatic metrics fail to capture. A subsequent experiment shows that the training-free MrFlow model attains an 8.25× speedup and, when integrated with a distilled model, reaches 11.3× speedup with the highest Geneval score, surpassing even some training-dependent methods. Finally, super-resolution benchmarks on Qwen‑Image reveal that despite similar metric scores, interpolation and SwinIR remain blurry and diffusion-based methods introduce artifacts, while a fast GAN-based Real‑ESRGAN variant provides the best balance of sharpness, semantic accuracy, and efficiency with a 14× speedup.