PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling

Xingyu Zheng, Xianglong Liu, Yifu Ding, Weilun Feng, Junqing Lin, Jinyang Guo, Haotong Qin

arXiv:2607.01642v1cs.CV

TL;DR

Scaling diffusion models increases image-generation cost, motivating acceleration methods that preserve quality without specialized systems support. MrFlow uses staged low-to-high-resolution sampling with pixel-space super-resolution and high-resolution refinement, achieving more than 10× speedup with OneIG-Bench loss within 1% of native inference and up to 25× when combined with timestep distillation.

  • Problem

    The rising computational cost of high-resolution diffusion generation creates a need for hardware-agnostic acceleration that maintains generation quality.

  • Method

    MrFlow stages low-resolution structure generation, pixel-space GAN super-resolution, low-strength latent noising, and high-resolution refinement for pretrained flow-matching models.

  • Results

    More than 10× end-to-end speedup keeps OneIG-Bench loss within 1% of native inference, while combining MrFlow with timestep distillation yields 25× speedup.

  • Takeaways & Limitations

    MrFlow provides training-free acceleration without runtime dynamic statistics and can be combined with timestep distillation without additional training.

  • Takeaways & Limitations

    Low-resolution generation can appear blurrier and struggle with fine-grained semantics, while GAN super-resolution may introduce localized artifacts and character-stroke shifts.

Abstract

from arXiv · show

Hardware-agnostic strategies for accelerating text-to-image diffusion, such as timestep distillation and feature caching, can reduce inference time without custom kernels or system-level optimization. Among them, multi-resolution generation strategies have recently received broad attention, attaining more than 5x speedup without any training. However, the design of performing upsampling in the latent space, together with the selective modification of partial regions, causes these methods to exhibit noticeable blurring or artifacts. To this end, we propose MrFlow, a training-free multi-resolution acceleration strategy for pretrained flow-matching models built upon a staged low-to-high-resolution pipeline. MrFlow first rapidly generates the main structure at low resolution, then performs super-resolution in the pixel space using a lightweight pretrained GAN-based model, subsequently injects low-strength noise to enable high-frequency resampling, and finally refines the details at high resolution. Quantitative and qualitative results on FLUX.1-dev and Qwen-Image show that MrFlow exploits the quadratic token reduction and reduced step requirement of low-resolution sampling to achieve 10x end-to-end acceleration while keeping OneIG within a 1% gap relative to that before acceleration, significantly surpassing other training-free acceleration strategies, and requiring no training or runtime dynamic identification whatsoever. MrFlow can further be directly combined orthogonally with pre-trained timestep distillation strategies, achieving even higher generation acceleration of up to 25x.

1 INTRODUCTION

MrFlow is a training-free multi-resolution sampling strategy that combines low-resolution structure generation, pixel-space super-resolution, noise-enabled high-frequency resampling, and high-resolution detail refinement. It achieves over 10× acceleration while keeping generation quality within a 1% gap, and combines with timestep distillation for over 25× acceleration.

  • Motivation: Scaling diffusion models increases computational cost, with Qwen-Image-20B requiring up to 47s for 1024 × 1024 text-to-image diffusion sampling on an Nvidia A100.
  • Method: MrFlow generates image structure in the low-resolution latent space, applies pixel-space super-resolution, injects low-strength noise, and refines details through high-resolution latent-space sampling.The pipeline includes VAE decoding and encoding between latent-space sampling and pixel-space super-resolution.
  • Results: Over 10× speedup keeps generation results within a 1% gap of native trajectory-flow generation, significantly surpassing other training-free strategies in quality and speed.Experiments evaluate metrics including OneIG-Bench and real sample generation, without training or runtime dynamic statistics and identification.
  • Results: Up to more than 25× acceleration is achieved by combining MrFlow orthogonally with pretrained timestep distillation strategies.

2 RELATED WORK

Related work establishes flow matching as the velocity-field formulation adopted by modern diffusion models, then situates acceleration among timestep, feature, token, and multi-resolution methods. It also distinguishes resolution expansion from native-resolution acceleration and connects pixel-space super-resolution with image repainting.

  • Flow matching: Flow matching and rectified flow recast diffusion training as learning a velocity field along linear interpolants between data and noise, with sampling integrating the learned velocity from t=1 to t=0.Modern models since SD3, including FLUX and Qwen-Image, largely adopt this formulation.
  • Acceleration of diffusion generation: Diffusion acceleration includes hardware- or systems-dependent quantization and efficient attention, alongside diffusion-specific timestep reduction, feature caching, and token pruning.The passage characterizes the gains of general hardware- or systems-dependent strategies on diffusion as relatively limited.
  • Multi-resolution generation: Multi-resolution generation uses staged low-to-high-resolution synthesis, with one class extending attainable resolution and another treating multi-resolution as an acceleration strategy at native resolution.The resolution-extension class targets images beyond the model’s native resolution rather than native-resolution acceleration.
  • Super-resolution and image repainting: Pixel-space super-resolution spans regression-based, generative-adversarial-based, and diffusion-based methods, while image repainting adds noise to an existing image before further processing.The cited representatives are SwinIR, Real-ESRGAN, OSEDiff, and SDEdit, respectively.

3 METHOD

MrFlow accelerates flow-matching generation through staged low-to-high-resolution sampling: it builds global structure cheaply at low resolution, applies pixel-space GAN super-resolution, then resamples and refines high-frequency details. The design preserves low-frequency content while using low-strength noise and minimal high-resolution computation to correct artifacts.

  • Overview: MrFlow generates global structure at low resolution, upsamples in pixel space, injects low-strength noise, and performs final high-resolution detail refinement.The low-resolution output nearly matches the prompt’s global structure but remains blurrier and struggles with fine-grained semantics such as text and symbols.
  • Low-Resolution Generation: About 4× speedup comes from shrinking each image side by 2×, because single-step inference cost scales almost linearly with image-token count.Low-resolution sampling also requires fewer steps; the method typically uses K_L = 12 for low-resolution flow sampling.
  • Pixel-Space Super-Resolution: MrFlow uses pixel-space Real-ESRGAN super-resolution to preserve coarse structure while adding sharp, natural-image-like high-frequency details for downstream refinement.Pixel-space upsampling reuses super-resolution priors, while GAN outputs better match the regime where refinement corrects mainly local residuals.
  • High-Frequency Resampling: With σ_t ∈[0.1, 0.15], flow-matching noising preserves low-frequency structure while lowering high-frequency SNR so refinement can resample GAN-induced artifacts.The condition is effective when super-resolution errors are localized in high-frequency directions rather than diffuse blur or low-frequency deviations.
  • High-Resolution Refinement: A single high-resolution Euler step is sufficient by default; at s = 0.1, it reaches CLIP consistency 0.9974 against the 8-step reference.The noised super-resolution latent is already close to the clean endpoint, reducing velocity variation and discretization error; extra steps provide marginal changes.

4 EMPIRICAL RESULTS

MrFlow is evaluated on FLUX.1-dev and Qwen-Image at 1024 × 1024 against training-free and training-dependent acceleration strategies. It consistently outperforms naive timestep configurations across tested models, while high-resolution refinement and super-resolution choices reveal important quality–speed trade-offs.

  • Experimental comparison: The experiments compare training-free and training-dependent acceleration methods on FLUX.1-dev and Qwen-Image using Geneval, DPG-Bench, and OneIG-Bench at 1024 × 1024.Training-dependent comparisons include timestep distillation methods and LSSGen, while training-free comparisons include feature-cache, token-pruning, and multi-resolution strategies.
  • Different step configurations: MrFlow consistently outperforms naive acceleration across models under varied timestep configurations, with trade-offs between GenEval and speedup shown in Figure 5.The Qwen-Image sweep uses 10, 12, 16, or 20 low-resolution steps and 1, 2, or 3 high-resolution steps; Figure 5 covers both FLUX.1-dev and Qwen-Image.
  • Different step configurations: A single high-resolution refinement step is an efficient default because low-strength noising makes the trajectory near the clean-image endpoint sufficiently straight.This conclusion follows the analysis of different MrFlow step configurations.
  • Super-resolution network: Most super-resolution choices remain close on automatic metrics, but OSEDiff is slightly below Real-ESRGAN and high-resolution refinement corrects local high-frequency defects.The comparison covers direct interpolation, SwinIR, OSEDiff, Real-ESRGAN, and super-resolution-only output; visual results use 12 low-resolution steps and 1 high-resolution step.

5 CONCLUSION · A BACKGROUND

MrFlow is a training-free, staged multi-resolution strategy that accelerates pretrained flow-matching models through low-resolution generation, pixel-space super-resolution, low-strength noising, and high-resolution refinement. The paper situates this approach among hardware-agnostic diffusion acceleration, multi-resolution generation, super-resolution, and image-repainting methods.

  • 5 CONCLUSION: More than 10× end-to-end speedup keeps OneIG-Bench loss within 1% of native inference without training or runtime dynamic statistics.MrFlow combines fast low-resolution structure generation, lightweight GAN-based pixel-space super-resolution, low-strength latent noising, and one-step high-resolution refinement.
  • 5 CONCLUSION: MrFlow is a training-free multi-resolution strategy for accelerating pretrained flow-matching models.
  • A BACKGROUND: Hardware-agnostic diffusion acceleration includes timestep reduction, feature caching, and token pruning, alongside broader quantization and efficient-attention strategies.Quantization and efficient attention generally target broad structures and commonly depend on hardware or systems for acceleration.
  • A BACKGROUND: Multi-resolution generation exploits information distributed across image resolutions and commonly uses staged low-to-high-resolution generation.The literature is divided into classes including methods that target extending attainable resolution, with examples such as SR3, Cascaded Diffusion Models, and Imagen.
  • A BACKGROUND: Pixel-space super-resolution methods include regression-based, generative-adversarial-based, and diffusion-based approaches.Examples include SwinIR for reconstruction-based super-resolution and GAN-based methods within the broader taxonomy.
  • A BACKGROUND: Unlike image-editing methods that use noising-denoising to preserve or modify existing images, MrFlow uses low-strength noise as a frequency-selective mechanism.Image editing approaches discussed include SDEdit, InstructPix2Pix, Imagic, and Prompt-to-Prompt.

B DETAILED EXPERIMENTAL SETUP

The experiments evaluate MrFlow primarily on FLUX.1-dev and Qwen-Image-20B, with extended tests on additional flow-matching models and comparisons against representative acceleration strategies. Quantitative evaluation uses 1024 × 1024 images and three complementary text-to-image quality benchmarks.

  • Experimental Models and Baselines: Experiments primarily use FLUX.1-dev and Qwen-Image-20B, with extended experiments on FLUX.2 Klein, Z-Image, and Z-Image-Turbo.The comparison includes representative acceleration strategies such as feature-cache methods TeaCache and DB-Taylor.
  • Evaluation Protocol: Quantitative evaluation fixes resolution at 1024 × 1024 and measures compositional semantics, dense alignment, and bilingual multi-dimensional attributes using Geneval, DPG-Bench, and OneIG-Bench.The attribute evaluation includes character rendering and style consistency under Chinese and English prompts.

C STAGE-WISE ANALYSIS

This appendix analyzes MrFlow stage by stage, providing experimental evidence and theoretical characterization for the design choices introduced in Section 3. It also separately examines how low-resolution sampling determines structure while high-resolution sampling refines details.

  • C STAGE-WISE ANALYSIS: Sections C.1–C.4 mirror Sections 3.1–3.4, supplementing each MrFlow stage’s design choices with complete experimental evidence and theoretical characterization.The appendix follows the stages in the same order as the main-text presentation.
  • C STAGE-WISE ANALYSIS: Section C.5 separately analyzes the division of labor between low-resolution structure determination and high-resolution detail refinement.

C.1 SOURCES OF ACCELERATION AT THE LOW-RESOLUTION STAGE · C.2 TWO DESIGN CHOICES OF PIXEL-SPACE SUPER-RESOLUTION

MrFlow’s low-resolution stage accelerates sampling through cheaper steps and faster structural convergence, while pixel-space GAN super-resolution avoids latent-space artifacts and leaves refinement to correct mainly high-frequency variation.

  • C.1 SOURCES OF ACCELERATION AT THE LOW-RESOLUTION STAGE: The low-resolution stage is the main acceleration source because it combines cheaper sampling steps with fewer steps needed to establish an indistinguishable global structure.These two benefits are examined separately on Qwen-Image.
  • C.1 SOURCES OF ACCELERATION AT THE LOW-RESOLUTION STAGE: 4× fewer image tokens from 2× spatial shrinking yields about 4× measured single-step speedup, while self-attention’s theoretical upper bound reaches 16×.The overall speedup is diluted because text tokens and non-attention computation contribute to runtime.
  • C.1 SOURCES OF ACCELERATION AT THE LOW-RESOLUTION STAGE: Low resolution converges in fewer steps because it uses the text condition more fully and traverses a shorter ODE path to establish the global skeleton.These explanations are complementary rather than mutually exclusive.
  • C.1 SOURCES OF ACCELERATION AT THE LOW-RESOLUTION STAGE: Low-resolution attention assigns more mass to text keys at measured layers, and the LR10 HR1 schedule achieves a higher CLIP score than HR11 under equal step counts.The attention measurements average 3 prompts and 20 trajectories; the CLIP comparison uses 3 prompts × 3 seeds.
  • C.1 SOURCES OF ACCELERATION AT THE LOW-RESOLUTION STAGE: Only 58% of the decomposed trajectory displacement is low-frequency, supporting the view that low-resolution inference reaches the global skeleton through a shorter ODE path.The low-resolution target is treated as bandwidth-equivalent to applying a low-pass operator to the high-resolution distribution.
  • C.2 TWO DESIGN CHOICES OF PIXEL-SPACE SUPER-RESOLUTION: Pixel-space upsampling is preferred because latent-space enlargement disrupts VAE-decoder spatial statistics and produces regular grid-like artifacts, visible in the MrFlow–LSSGen comparison.Pixel space preserves the domain in which super-resolution models operate on image statistics.
  • C.2 TWO DESIGN CHOICES OF PIXEL-SPACE SUPER-RESOLUTION: After refinement, Gaussian and high-frequency noise remain near the default SR starting point, whereas low-frequency noise reaches 0.498 at strength 0.3, increasing by over 200%.This indicates that refinement mainly resamples high-frequency directions and cannot effectively correct low-frequency bias, motivating a GAN-class super-resolution model.

C.3 FORMAL CHARACTERIZATION OF HIGH-FREQUENCY RESAMPLING

MrFlow’s high-frequency resampling is characterized by an anisotropic posterior-mean decomposition: low-frequency directions preserve the super-resolution estimate, while high-frequency directions revert toward the prior as noise increases. This yields a sufficient noise lower bound for prior takeover, with stronger overall-alignment conditions possible under distribution-level analysis.

  • Weighted decomposition of the posterior mean: The posterior mean decomposes refinement direction-wise into weighted SR content, prior mean, and injected noise under the Tweedie/MMSE characterization.The clean latent prior is locally modeled as an anisotropic Gaussian because natural-image variance differs substantially across spatial frequencies.
  • Weighted decomposition of the posterior mean: κi = SNRi/(1 + SNRi) preserves low-frequency structure as κi →1 but resamples high-frequency detail as κi →0.The Wiener gain depends on clean-data variance and injected noise, while the bound assumes the SR output is already locally plausible rather than arbitrarily erroneous.
  • The noise-level lower bound: SNRhf ≤1 and κhf ≤1 hold when injected high-frequency noise is at least as large as the clean high-frequency signal.This condition makes the prior at least as influential as the SR estimate in the targeted high-frequency band.
  • The noise-level lower bound: σt ≥√λhf/(1 + √λhf) = σ⋆ is a sufficient, not necessary, bound determined by measurable clean-data high-frequency variance.When SR high-frequency content is reliable, smaller σt can preserve more SR information; the bound instead guarantees that the prior can cover the high-frequency band.
  • The noise-level lower bound: σt ≈0.075 supports the main-text choice σt ∈[0.1, 0.15], whereas mid- and low-frequency residuals raise the measured bounds to 0.16 and 0.24.A distribution-level Gaussian-smoothing argument provides a stronger sufficient overall-alignment condition, σt/(1 −σt) ≳W2(ˆp0, p0), than the direction-wise bound.

C.4 SINGLE-STEP SUFFICIENCY OF HIGH-RESOLUTION DETAIL REFINEMENT

Under the recommended low-strength regime, high-resolution refinement is effectively sufficient with one denoising step because trajectories are straighter near t = 0. This also makes allocating limited steps toward the schedule’s low-noise end most cost-effective.

  • Numerical single-step sufficiency: 0.9974 CLIP similarity from one step is only 0.0025 below the 0.9999 achieved by five steps at s = 0.1, whereas s = 0.3 requires more steps.The comparison uses the same super-resolution starting point, refinement noise seed, and 8-step reference; s = 0.3 serves as a larger-noise diagnostic setting.
  • Time-varying characteristics of the velocity field: All three velocity-field statistics increase monotonically with strength, so smaller strength produces flatter trajectories, weaker bending, and lower single-step Euler error.The measured quantities are velocity magnitude, adjacent velocity change, and turn angle during Qwen-Image high-resolution refinement.
  • Step-budget allocation: Because trajectories straighten closer to t = 0, the limited step budget should concentrate steps in the low-noise end rather than distribute them uniformly or front-load them.The schedule comparison fixes KH = 6 and contrasts uniform, front-dense, and end-dense allocations from the same super-resolution starting point.

C.5 STRUCTURE-DETAIL DECOUPLING

MrFlow’s low-resolution stage determines the image’s global structure, while high-resolution refinement primarily adds localized high-frequency detail. Seed perturbation experiments show this separation through large low-frequency sensitivity to low-resolution randomness and high-frequency concentration of high-resolution variation.

  • Experimental setup: The experiment varies low- or high-resolution seeds separately across four seeds, computes pairwise MSE over three prompts, and evaluates full-, low-, and high-frequency components.Images are decomposed with a Gaussian low-pass filter, and the high-frequency/low-frequency MSE ratio identifies where each perturbation is concentrated.
  • Structure-detail decoupling: Changing the low-resolution seed alters low-frequency content about 8000× more and full-image content about 1900× more than changing the high-resolution seed.This asymmetry indicates that final structural content is almost entirely determined by the low-resolution stage.
  • Structure-detail decoupling: High-resolution refinement produces more high-frequency-focused changes, with a high-frequency/low-frequency MSE ratio of 19.8 versus 7.4 for low-resolution variation.The low-resolution seed controls global structure, whereas high-resolution randomness mainly introduces local detail on the fixed low-resolution starting point.

D MRFLOW ON RECENT OPEN MODELS

MrFlow transfers to FLUX.2 Klein and Z-Image, preserving quality near native generation while delivering substantial speedups across base, aggressive, and distilled configurations. Combined with reduced-step models, it reaches up to 26.92× effective acceleration.

  • FLUX.2 Klein: 8.79× speedup is achieved on FLUX.2 Klein Base 9B with the aggressive 12, 1 configuration, while 20, 1 reaches 5.40× with a very small OneIG-En gap.The corresponding 4B variants reach 8.03× and 5.16×, respectively; the aggressive setting incurs a moderate metric drop.
  • FLUX.2 Klein: 26.92× speedup is achieved on FLUX.2 Klein 9B distilled with one high-resolution refinement step, while the corresponding 4B variant reaches 20.10× relative to its base model.These results use the MrFlow‡ configuration on distilled Klein variants.
  • Z-Image: 21.0× effective speedup is achieved on Z-Image-Turbo with the 8, 1 effective DiT-forward setting.This extends MrFlow to a reduced-step Z-Image model.
  • Z-Image: 10.8× speedup is reached on base Z-Image with the 12, 1 setting while maintaining usable metric performance.The less aggressive 20, 1 setting achieves 6.66× while keeping OneIG-En at 0.55 and OneIG-Zh at 0.51, close to the native model.

E EFFICIENCY ANALYSIS · F MORE GENERATION EXAMPLES

MrFlow’s 12+1 Qwen-Image configuration has a measured average end-to-end latency of 4.7, with runtime dominated by low-resolution sampling and high-resolution refinement. Additional figures compare generation quality across acceleration methods on FLUX.1-dev, while noting that PDF compression can reduce apparent visual sharpness.

  • E EFFICIENCY ANALYSIS: 4.7 is the reported average end-to-end latency for MrFlow’s 12+1 configuration on Qwen-Image.The measurement includes text encoding, noise generation, denoising, VAE operations, pixel-space super-resolution, and high-resolution noising.
  • E EFFICIENCY ANALYSIS: 3.24s and 1.03s are consumed by low-resolution sampling and high-resolution refinement, respectively.These are identified as MrFlow’s dominant runtime costs in Figure 6.
  • E EFFICIENCY ANALYSIS: About 0.30s is spent on fixed overhead from initial noise generation, intermediate VAE decoding, Real-ESRGAN super-resolution, and VAE encoding.The breakdown indicates that these non-denoising stages are small relative to the dominant sampling and refinement costs.
  • F MORE GENERATION EXAMPLES: PDF rasterized figures use compression above 90%, which may make fine-grained details appear less sharp in densely arranged comparisons.The passage states that this is a document-preparation artifact rather than a limitation of the generated images or MrFlow.
  • F MORE GENERATION EXAMPLES: Figure 7 compares pretrained, training-free, and timestep-distillation-based strategies on FLUX.1-dev.Dashed lines separate the three strategy groups.
  • F MORE GENERATION EXAMPLES: Figure 8 provides a detail comparison of multi-resolution acceleration methods on FLUX.1-dev.The supplied passage identifies the comparison scope but does not report individual visual outcomes.

F.1 COMPARISON WITH VARIOUS SOTA STRATEGIES · F.2 MORE EXAMPLES OF MRFLOW · F.3 PROMPTS OF THE IMAGES IN THE PAPER

MrFlow shows stronger qualitative generation than existing training-free acceleration strategies across FLUX.1-dev and Qwen-Image, while maintaining 8–10× or higher speedups across varied resolutions and aspect ratios. The appendix also documents the prompts used for reproducibility and inspection of the paper’s image examples.

  • F.1 COMPARISON WITH VARIOUS SOTA STRATEGIES: MrFlow surpasses existing training-free strategies in generation quality and efficiency on FLUX.1-dev DPG-Bench examples, while MrFlow† achieves higher speedup through timestep distillation.The comparison is presented in Figure 7, and MrFlow† requires no extra training.
  • F.1 COMPARISON WITH VARIOUS SOTA STRATEGIES: Compared with LSSGen, RALU, and SPEED, MrFlow shows a clear perceptual-fidelity advantage despite SPEED partially exceeding it on some FLUX.1-dev metrics.LSSGen exhibits dense artifacts at moderate speedup, whereas RALU and SPEED retain competitive quantitative scores above 8×.
  • F.1 COMPARISON WITH VARIOUS SOTA STRATEGIES: On diverse Qwen-Image prompts, MrFlow outperforms other advanced training-free strategies and combines directly with timestep distillation for higher speedup.This observation is consistent with the FLUX.1-dev examples.
  • F.2 MORE EXAMPLES OF MRFLOW: At 8–10× speedup, MrFlow produces excellent results across arbitrary aspect ratios and resolutions using 12 low-resolution steps and 1 high-resolution step.Figures 9 and 10 show these generation effects on FLUX.1-dev and Qwen-Image, respectively.
  • F.2 MORE EXAMPLES OF MRFLOW: On Qwen-Image, Figure 9 reports MrFlow examples with 12 low-resolution steps, 1 high-resolution step, and speedups above 10×.The figure caption identifies the configuration and speedup range.
  • F.3 PROMPTS OF THE IMAGES IN THE PAPER: The listed prompts span realistic photography, science-fiction scenes, food and craft settings, landscapes, wildlife, posters, and text-constrained compositions.Prompts marked † were originally written in Chinese and are provided as English translations or descriptions for pdfLaTeX compatibility.
Loading 2607.01642v1…