Research map / Image generation / Autoregressive and latent generation

FreCaS: Efficient Higher-Resolution Image Generation via Frequency-aware Cascaded Sampling

ICLR 2025 · Image generation › Autoregressive and latent generation · approach: High-resolution diffusion synthesis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FreCaS, a training-free frequency-aware cascaded sampling framework that generates higher-resolution images from pretrained diffusion models by progressively increasing resolution, applying frequency-split classifier-free guidance, and reusing cross-attention maps across stages.

Problem

Training-free generation of images at resolutions above the training size (e.g., 2048x2048 and 4096x4096 with SDXL) is computationally expensive, and existing methods run full-resolution sampling and process all frequencies at once, causing latency and duplicated-object artifacts.

Approach

Decompose sampling into cascaded resolution stages that expand the frequency band, split the denoising score into low- and high-frequency parts with separate CFG strengths (high weight for newly added frequencies), and fuse cross-attention maps from the previous stage to keep layouts consistent.

Key results

• On 2048x2048 SDXL generation, FreCaS is about 2.86x faster than ScaleCrafter and 6.07x faster than DemoFusion per the abstract, with FID_b improvements of 11.6 and 3.7 respectively.
• On x16 SDXL (4096x4096), FreCaS achieves FID_b 42.75 vs 47.80 for DemoFusion at 3.64x speedup over DirectInference (85.87 s latency).
• On SD3 x4 generation, FreCaS reaches FID_b 9.76 vs 15.19 for DemoDiffusion and runs 2.42x faster than DirectInference.

Measured effects reported in the paper

Surprising findings

• A single cascaded sampling process with resolution transitions is both faster and higher quality than methods that run a full diffusion process at each resolution (e.g., DemoFusion).
• Excessively high high-frequency guidance (w_h=45) introduces artifacts, so the FA-CFG weight has a sweet spot.

Limitations

• Relies on hyperparameters (w_h, w_c, stage schedule L and N, gamma) that are tuned per model and resolution.
• Success rate of generated images remains low (below 40% in user study), per the authors' own user study.
• Cannot be applied directly to architectures with convolutions removed without adaptation; SD3 results only at x4 due to GPU memory limits.

Experimental setup

Datasets: LAION5B aesthetic subset

Benchmarks: LAION5B aesthetic subset prompts

Metrics: FID_b FID_p IS IS_p CLIP score latency CLIPIQA NIQE MUSIQ user study

Baselines: DirectInference MultiDiffusion AttnEntropy ScaleCrafter FouriScale HiDiffusion AccDiffusion DemoFusion

Builds on: classifier-free guidance MultiDiffusion DemoFusion ScaleCrafter rectified flow matching

Closest papers by meaning

• FAM Diffusion: Frequency and Attention Modulation for High-Resolution Image Generation with Stable Diffusion (CVPR 2025 · similarity 0.89)
• DiffuseHigh: Training-Free Progressive High-Resolution Image Synthesis Through Structure Guidance (AAAI 2025 · similarity 0.87)
• FreqTS: Frequency-Aware Token Selection for Accelerating Diffusion Models (AAAI 2025 · similarity 0.86)
• SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.86)
• ResMaster: Mastering High-Resolution Image Generation via Structural and Fine-Grained Guidance (AAAI 2025 · similarity 0.86)
• ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models (ICLR 2024 · similarity 0.86)