Megapixel visual synthesis with latent diffusion models requires operating far beyond the resolutions at which these models are typically trained. Cascade refinement is a leading strategy for this setting: it starts from a base image at a native resolution and progressively upsamples it to the target resolution through multiple refinement stages. Each stage typically performs a pixel-space resolution transition followed by latent-space denoising. While prior acceleration efforts mainly focus on reducing denoising cost, we identify a complementary and underexplored bottleneck: the repeated movement of representations through the VAE. We call these operations pixel-space transitions. In cascade pipelines, intermediate latents are often decoded to RGB, resized, and re-encoded before refinement, while final latents are decoded patch by patch with dense overlap to suppress boundary artifacts. We propose UltraFlash, a transition-efficient approach that accelerates these VAE-heavy operations. UltraFlash has two components. Latent Hyperloop replaces intermediate RGB round trips with a learned latent-space shortcut that approximates the decode-resize-encode transformation without repeated VAE decoding and encoding. Sparse Patch-based Decoding accelerates final reconstruction by addressing patch seams caused by inconsistent GroupNorm statistics and reusing cached statistics across patches, enabling sparse VAE decoding with much smaller overlap. Integrated into a state-of-the-art cascade baseline, UltraFlash reduces 4K generation latency by approximately 2x while improving image quality. These results show that optimizing pixel-space transitions can improve both efficiency and fidelity, offering an effective route toward fast, high-quality megapixel visual synthesis.
Figures & tables
Figure 2: UltraFlash targets transition costs common to cascade megapixel synthesis. Cascade pipelines often use pixel space twice: an intermediate RGB round trip before refinement and dense patch-wise VAE decoding for final output. UltraFlash accelerates both without changing the denoising/refinement stage. Latent Hyperloop removes repeated VAE decode/encode calls during intermediate transfer, while Sparse Patch-based Decoding reduces final decoding redundancy by enforcing consistent GroupNorm statistics across patches.
Figure 3: Shared GroupNorm statistics enable sparse final RGB decoding. When overlap is small, standard patch-wise decoding shows seam artifacts caused by independently computed GroupNorm statistics. Sparse Patch-based Decoding uses shared cached statistics to align normalization across patches, making adjacent patches decode consistently. The right plot shows that our method maintains consistently low boundary MSE across overlap factors when decoding with sparse overlaps.
Resolution 2K ( 2048×2048 )
Resolution 4K ( 4096×4096 )
Method
FID ↓
FID c ↓
IS ↑
Time (s)
FID ↓
FID c ↓
IS ↑
Time (s)
FouriScale ( Huang et al., 2024 )
72.65
51.90
12.31
87.84
98.97
69.56
8.54
680.57
DemoFusion ( Du et al., 2024 )
68.46
43.39
13.15
75.89
74.75
52.14
12.57
507.74
FreeScale ( Qiu et al., 2025 )
52.87
29.64
13.56
53.20
58.28
34.01
13.35
323.93
PixelRush-SD ( Lai et al., 2026 )
52.56
29.66
13.69
3.17
56.50
33.76
13.42
16.37
PixelRush-XL ( Lai et al., 2026 )
50.13
29.13
14.32
4.06
54.28
33.15
13.71
17.11
Table 1: Quantitative comparison among SDXL-based high-resolution generation methods. We report FID ( ↓ ), FID c ( ↓ ), Inception Score (IS ↑ ), and Inference Time (seconds).
Figure 4: Qualitative comparison at 2K and 4K resolution. We compare high-resolution generation from FouriScale, DemoFusion, FreeScale, PixelRush, and UltraFlash at 2048×2048 (left) and 4096×4096 (right). UltraFlash produces sharper textures, cleaner facial regions, and more coherent fine structures, while achieving substantially faster inference. Runtimes are measured at 4096×4096 .
Method
Backbone
Route
FID ↓
Time (s) ↓
SANA-4K ( Xie et al., 2025 )
SANA
Native 4K
72.13
24.79
DyPE-FLUX ( Issachar et al., 2025 )
FLUX
Direct 4K
59.05
587.13
UltraFlash-SD ∗
SDXL/SD-Turbo
Cascade
52.38
8.04
UltraFlash-XL ∗
SDXL/SDXL-Turbo
Cascade
51.32
9.61
Table 2: Additional evaluations beyond the main SDXL cascade comparison. Left: contextual comparison with alternative 4K generation routes, separated from Table 1 because these methods use different backbones or training assumptions. Right: transfer of the proposed transition operators to a second high-resolution pipeline.
Figure 5: User preference study. Preference share of each method on three criteria. UltraFlash (gold bars) receives the highest share ( ∼ 30–33%) on all three.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
Scale
Dense
Sparse
SDXL-Turbo, 2K
× 4
3
3
SDXL-Turbo, 4K / SD-Turbo, 2K
× 4
7
6
SD-Turbo, 4K
× 8
15
11
Appendix
Table A1: Patching configurations. We report here the patch count per spatial dimension (total patches = count 2 ) in each setting.
Figure A1: Latent Hyperloop ablations and convergence analysis. Left: We report 4K generation ablations and robustness across different base/refinement backbone combinations. Latent Hyperloop improves generation quality and latency, while Sparse Patch-based Decoding further reduces inference time with only a small FID trade-off. Right: SSIM convergence curves. (c) Attention-based objectives accelerate convergence and improve the final reconstruction quality over the ℓ1 baseline and E-LatentLPIPS [ 11 ] ; (d) the Spatial Prior architecture consistently outperforms alternative designs.
Context pixels p
0
4
8 (ours)
12
16
FID ↓
59.18
55.86
51.99
51.67
51.71
IS ↑
13.12
13.48
13.89
13.82
13.89
Appendix
Table A3: Ablation on the context size p of Latent Hyperloop. Same 4K setting as Fig. A1 a (SD-Turbo refiner, before adding Sparse Patch-based Decoding); p=0 corresponds to the variant without context expansion.
Figure A2: Scalability Analysis. Comparison of inference latency across increasing resolutions (log scale). UltraFlash (blue) offers significantly better scaling properties than PixelRush (orange).
Figure A3: Latent Hyperloop vs. image bilinear interpolation. Our Latent Hyperloop produces outputs that are visually indistinguishable from the expensive pixel-space round-trip. Best viewed zoomed-in .
Figure A4: Effect of training objectives on Latent Hyperloop reconstruction. From left to right: reconstructions trained with ℓ1 only, ℓ1 + E-LatentLPIPS, and ℓ1 + Attention Alignment, followed by the ground-truth reconstruction. All variants use the same latent upsampling backbone and are trained under the same setting. While ℓ1 preserves the overall structure, it produces noticeably smooth reconstructions, and E-LatentLPIPS provides only marginal visual improvement. The proposed Attention Alignment objective better preserves local structure and fine facial details, yielding reconstructions that are visually closest to the ground truth. Best viewed zoomed in.
Figure A5: Uncurated UltraFlash outputs at 4096×4096 . Generated with FLUX base model and UltraFlash refinement pipeline.
Figure A6: Uncurated UltraFlash outputs at 4096×4096 . Generated with FLUX base model and UltraFlash refinement pipeline.
Figure A7: Uncurated UltraFlash outputs at 4096×4096 . Generated with SDXL base model and UltraFlash refinement pipeline.
Figure A8: Uncurated UltraFlash outputs at 4096×4096 . Generated with SDXL base model and UltraFlash refinement pipeline.
Latent diffusion models have enabled high-quality video synthesis, yet their inference remains costly and time-consuming. As diffusion transformers become increasingly efficient, the latency bottleneck inevitably shifts to VAE decoders. To reduce their latency while maintaining quality, we propose a universal acceleration framework for VAE decoders that preserves full alignment with the original latent distribution. Specifically, we propose (1) an independence-aware channel pruning method to effectively mitigate severe channel redundancy, and (2) a stage-wise dominant operator optimization strategy to address the high inference cost of the widely used causal 3D convolutions in VAE decoders. Based on these innovations, we construct a Flash-VAED family. Moreover, we design a three-phase dynamic distillation framework that efficiently transfers the capabilities of the original VAE decoder to Flash-VAED. Extensive experiments on Wan and LTX-Video VAE decoders demonstrate that our method outperforms baselines in both quality and speed, achieving approximately a 6× speedup while maintaining the reconstruction performance up to 96.9%. Notably, Flash-VAED accelerates the end-to-end generation pipeline by up to 36% with negligible quality drops on VBench-2.0. Our code is available at https://github.com/Aoko955/Flash-VAED.
Lunjie Zhu, Yushi Huang, Xingtong Ge +5
iComAI Lab, Hong Kong University of Science and Technology, Hong Kong SAR, China · School of Data Science, Lingnan University, Hong Kong SAR, China
High-resolution synthesis with Latent Diffusion Models (LDMs) faces a fundamental trade-off between computational cost and generative fidelity. This trade off is largely governed by the latent compression factor f (e.g. f8 denotes f=8 and f32 f=32). While aggressive compression (f32) significantly improves efficiency, it suffers from severe information loss and ambiguity compared to standard f8 models. We propose UltraDiff, a framework that reformulates this challenge as joint latent diffusion learning followed by marginal score distillation. We introduce MixTeacher, a dual branch DiT that models the joint f8-f32 distribution, utilizing shared attention to resolve the ambiguity of the compressed space with high fidelity f8 priors. Crucially, we propose vector field distillation in the native f32 space. Theoretically, this compels the student to learn the conditional expectation of the teacher trajectory, effectively filtering out compression artifacts. Across the evaluated 1K to 4K settings, UltraDiff approaches f8 level quality while preserving f32 efficiency. With 20 sampling steps on a single A100 GPU, it generates a 2K image in 6.2 seconds over 40 times faster than our f8 baseline.
Jingjing Ren, Haitian Zheng, Haoyu Chen +5
The Hong Kong University of Science and Technology (Guangzhou) · Adobe Research
Real-time video generation demands fast decoding as much as fast denoising, yet current latent video diffusion models rely on 3D convolutional decoders that are slow and memory-intensive at high resolutions or for long video. We introduce FlashDecoder, a fast, memory-efficient pure-Transformer video decoder that decodes latents to pixels frame by frame. At each step, the current frame attends only to a fixed-size window of past frames through a rolling KV cache. The fixed temporal window keeps decoding fast and memory bounded regardless of video length, enabling constant-latency streaming. Because frames are processed sequentially, temporal causality is enforced without explicit attention masks, enabling training at resolutions up to 1080p and matching the reconstruction quality of convolutional decoders. On the Wan2.1 and Wan2.2 latent spaces, FlashDecoder matches each convolutional decoder in reconstruction quality (e.g., 41.55dB vs. 41.49dB PSNR at 1080p) while decoding 3.6x-4.7x faster with up to 11x less memory on a single H100 GPU. With architecture-aware inference optimizations, the speedup widens to 12x.