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

Abstract: 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.

• The paper introduces MrFlow, a training-free staged pipeline that generates global structure at low resolution, applies pixel-space GAN super-resolution, and performs single-step high-resolution refinement.
• MrFlow achieves over 10× end-to-end speedup on FLUX.1-dev and Qwen-Image-20B while keeping OneIG-Bench degradation within 1%, and reaches up to 25× when combined with timestep distillation.
• The method relies on low-strength noise to remove mainly high-frequency super-resolution errors, but its effectiveness depends on GAN-based upscaling and may decline when outputs contain blur or low-frequency deviations.

MrFlow is a training-free acceleration strategy for pretrained flow-matching text-to-image models that replaces the conventional single-resolution denoising trajectory with a staged low-to-high-resolution pipeline. The central observation is that the low-resolution stage determines global structure and semantics while the high-resolution stage only needs to correct local high-frequency residuals introduced by super-resolution. By exploiting this structure–detail decoupling, the method achieves more than $10\times$ end-to-end speedup on FLUX.1-dev and Qwen-Image-20B with OneIG-Bench degradation within $1\%$, and composes orthogonally with timestep distillation to reach up to $25\times$.

Motivation and positioning

The paper situates MrFlow against three families of hardware-agnostic diffusion accelerators. Timestep distillation reduces NFEs from 50–100 to 1–4 but requires costly fine-tuning; feature caching (TeaCache, DB-Taylor) attains roughly $4\times$ training-free speedup but collapses at more aggressive settings; token compression methods such as ToMA yield less than $1.5\times$. Multi-resolution strategies (LSSGen, RALU, SPEED) exceed $5\times$ without training, but they perform upsampling in the latent or frequency domain and often rely on runtime dynamic region identification, which produces blurring and artifacts. MrFlow's distinguishing claim is that pixel-space super-resolution with a GAN-based model, followed by low-strength noise injection, avoids these failure modes entirely without any training or runtime statistics.

The staged pipeline

The pipeline consists of seven operations: low-resolution latent sampling from noise, VAE decoding, pixel-space super-resolution, VAE re-encoding, noise injection, high-resolution latent refinement, and final VAE decoding. The default configuration uses $K_L=12$ Euler steps at low resolution, Real-ESRGAN $\times2$ upsampling, noise strength $\sigma_t\in[0.1,0.15]$, and a single high-resolution denoising step. Each stage is justified by an independent analysis:

Low-resolution structure generation. Acceleration has two sources. Per-step cost scales nearly linearly with token count, giving about $4\times$ speedup when each spatial side is halved (the quadratic self-attention gain is diluted by linear operators). More notably, fewer steps are required: under an equal step budget, a schedule of 10 LR steps plus 1 HR step achieves higher CLIP consistency than 11 direct HR steps at roughly half the latency. The authors attribute this to stronger text-to-image attention mass at low resolution (e.g., $1\%$0 vs. $1\%$1) and to a shorter ODE path — frequency-domain decomposition shows the low-pass component of sampled trajectories accounts for only about $1\%$2 of total path length across resolutions.

Pixel-space GAN super-resolution. Upsampling is performed on decoded pixels rather than latents because natural-image priors exist in the pixel domain, and VAE re-encoding acts as a lightweight regularizer on out-of-distribution high frequencies. The choice of GAN-based SR over interpolation or L2 regression (SwinIR) is motivated by the downstream refine's asymmetric correction capability: structured-noise ablations show that after refinement, kNN distance to the clean HR cloud is essentially unchanged for Gaussian and high-frequency perturbations, but rises to $1\%$3 under low-frequency noise at strength $1\%$4 — a more than $1\%$5 expansion. The refine can therefore resample high-frequency errors but cannot repair diffuse blur or low-frequency deviation, which is precisely the residual profile of L2-based SR.

Low-strength noising as frequency-selective resampling. The theoretical core is a Wiener-gain analysis of single-step denoising under a locally anisotropic Gaussian prior. Along eigendirection $1\%$6, the posterior mean weights the SR estimate by $1\%$7; directions with small clean-data variance (high frequency) revert to the prior when noise dominates. Under a spatial-stationarity approximation, eigenvalues reduce to the radial power spectrum, yielding the sufficient condition $1\%$8. Measured spectra give $1\%$9 for the high-frequency band, consistent with the empirical operating range of $25\times$0. A distribution-level KL bound via Gaussian smoothing is provided as a complementary view, requiring substantially larger $25\times$1 — reinforcing that low-strength noising suffices only because SR errors are confined to high frequency.

Single-step high-resolution refinement. Because the noised latent sits near the clean endpoint where the velocity field is flattest (measured velocity change and turn angle increase monotonically with strength), one Euler step suffices: at $25\times$2, single-step refinement reaches CLIP similarity $25\times$3 against an 8-step reference, versus $25\times$4 for 5 steps. An end-dense step allocation saves about $25\times$5 of trajectory displacement relative to front-dense scheduling, which the authors connect to the SD3 timestep-shift principle taken to its extreme.

Seed-sensitivity experiments directly confirm the structure–detail decoupling premise: varying the low-resolution seed changes low-frequency MSE by roughly $25\times$6 more than varying the high-resolution seed, and the HR-stage perturbation concentrates in high frequency (ratio $25\times$7 vs. $25\times$8).

Empirical results

At $25\times$9 resolution on Geneval, DPG-Bench, and OneIG-Bench, MrFlow dominates the training-free baselines at aggressive speedups. On Qwen-Image with the $4\times$0 configuration it achieves $4\times$1 end-to-end speedup with Geneval $4\times$2 (native: $4\times$3), whereas TeaCache collapses to $4\times$4 and DB-Taylor to $4\times$5 near $4\times$6; RALU and SPEED degrade to $4\times$7 and $4\times$8 respectively. On FLUX.1-dev, MrFlow at $4\times$9 retains Geneval $1.5\times$0, close to the 4-step distilled SenseFlow ($1.5\times$1, $1.5\times$2). Combined with Pi-Flow weights without additional training, MrFlow$1.5\times$3 reaches $1.5\times$4 on Qwen-Image with OneIG loss within $1.5\times$5, and combined with FLUX-schnell it slightly exceeds native schnell on DPG despite comparable speedup. Extended experiments on FLUX.2 Klein and Z-Image show the design transfers across backbones, reaching up to $1.5\times$6 on distilled Klein-9B relative to its base model. All reported speedups are end-to-end, including text encoding, VAE transforms, and SR overhead, which total only about $1.5\times$7s per image in the $1.5\times$8 Qwen-Image configuration (4.77s vs. 49.32s native).

Ablations indicate quality depends almost entirely on the number of low-resolution steps, with high-resolution steps beyond one providing negligible benefit — consistent with the trajectory-straightness analysis. Among SR choices, automatic metrics are largely insensitive, but visual inspection favors Real-ESRGAN: interpolation and SwinIR remain blurry, and OSEDiff introduces character inaccuracies.

Limitations and open questions

The paper concedes several boundaries explicitly. The noise-level lower bound is a sufficient, not necessary, condition and presumes that SR residuals are localized in high frequency; if the SR output contains diffuse blur or low-frequency deviations, the premise fails and much stronger noising would be needed, washing away useful detail and increasing step count. This also delimits the applicable regime of the pipeline to GAN-class SR whose bias matches the correctable band. The formal analysis relies on a local anisotropic Gaussian approximation of the latent prior and a spatial-stationarity assumption that reduces covariance spectra to power spectra — approximations whose validity in other domains is not established. Evaluation is confined to text-to-image at around $1.5\times$9; behavior at larger upsampling factors than $5\times$0, or on non-image modalities, remains untested. Finally, the few-step convergence phenomenon at low resolution is explained through two complementary mechanisms (text-condition utilization and shorter ODE paths) but not derived from first principles, leaving open whether the observed $5\times$1 low-frequency path-length ratio generalizes across architectures and schedulers.

Conclusion

MrFlow demonstrates that a carefully staged multi-resolution pipeline — low-resolution structure generation, pixel-space GAN super-resolution, low-strength noise injection grounded in a Wiener-type spectral analysis, and single-step high-resolution refinement — can deliver training-free acceleration exceeding $5\times$2 with quality within $5\times$3 of native inference, and compounds with timestep distillation to $5\times$4. Its main contribution is less any single component than the identification of the frequency-selective conditions under which each stage operates, together with empirical evidence that this division of labor holds robustly across modern flow-matching backbones.