UltraFlash: Accelerating Megapixel Visual Synthesis

• Models trained around 512 × 512 or 1024 × 1024 often produce repeated objects, distorted layouts, or unstable textures on larger canvases.High-resolution fine-tuning is difficult to scale because diverse megapixel datasets are limited, training is expensive, and adapted models often specialize to a fixed resolution [35].Source: §1 · p. 2
• Cascade pipelines spend substantial runtime on VAE-heavy transitions as well as denoising, through intermediate RGB round trips and dense patch-wise final decoding.The intermediate round trip is avoidable, whereas final RGB decoding is necessary but can be made sparse without introducing seams.Source: §1 · p. 2
• UltraFlash replaces intermediate RGB round trips with Latent Hyperloop and accelerates final reconstruction with Sparse Patch-based Decoding.The learned latent shortcut approximates decode–resize–encode; the decoder shares cached GroupNorm statistics across patches to enable sparse decoding with smaller overlap.Source: §1 · p. 2
• UltraFlash reduces 4K generation latency by 2× while improving image quality when integrated into the current state-of-the-art cascade baseline [13].The result highlights VAE transitions as an optimization opportunity even in highly optimized synthesis systems.Source: §1 · p. 2

• Cascade and region-based refinement form one route to megapixel synthesis, alongside high-resolution training and direct extrapolation of pretrained models.High-resolution approaches require data, adaptation, or dedicated training; direct extrapolation can face memory cost, structural instability, or unnatural textures at extreme resolutions.Source: §2 · p. 3
• UltraFlash uses PixelRush as its experimental setting because its denoising speed is already optimized, making it a test of remaining VAE transition overhead.It optimizes intermediate RGB round trips and final patch-wise VAE decoding rather than introducing another refinement strategy.Source: §2 · p. 3
• Latent Hyperloop approximates the decode–resize–encode transformation patch-wise, unlike ordinary latent interpolation, and leaves the denoiser unchanged.FAM Diffusion instead uses whole-latent frequency and attention modulation and changes the denoiser.Source: §2 · p. 3
• Sparse Patch-based Decoding addresses patch-boundary artifacts by sharing cached GroupNorm statistics, enabling reconstruction with much smaller overlap.Large overlaps repeat computation, while small overlaps can reveal boundary artifacts.Source: §2 · p. 3
• Megapixel synthesis differs from super-resolution restoration because it expands a generated intermediate toward coherent synthesis rather than recovering a unique high-resolution target.The two settings therefore require different evaluation criteria.Source: §2 · p. 3

• Latent diffusion encodes an image as z0 = E(x), gradually corrupts that latent, and trains a denoising network to reverse the process.At inference, a Gaussian latent is denoised and the final latent is decoded with the VAE decoder.Source: §3 · p. 3
• Cascade refinement starts at native resolution and progressively enlarges the image, typically initializing each larger latent by routing the current latent through RGB.The standard transition decodes, resizes the RGB image by 2×, and re-encodes it before refinement.Source: §3 · pp. 3–4
• Cascade pipelines incur repeated intermediate VAE round trips and final patch-wise VAE decoding, which UltraFlash targets with a learned latent shortcut and sparse, consistent reconstruction.Full-image decoding is memory-intensive at megapixel scale, while dense patch overlap suppresses seams but repeats computation.Source: §3 · p. 4, Figure 2 · p. 4

4 Accelerating Megapixel Visual Synthesis

• UltraFlash handles intermediate RGB round trips with Latent Hyperloop and unavoidable final RGB reconstruction with Sparse Patch-based Decoding.The two transitions arise before refinement and at final decoding, respectively.Source: §4 · p. 4

4.1 Latent Hyperloop: Bypassing Intermediate RGB Round Trips

• Latent Hyperloop learns a direct latent mapping that approximates decode–resize–encode, avoiding repeated VAE decoding and encoding.Ordinary latent interpolation is not equivalent because the VAE encoder is nonlinear.Source: §4.1 · pp. 4–5
• Context-expanded patches supply neighboring information, while a pretrained VAE decoding block provides a spatial upsampling prior for the lightweight network.The method uses 80 × 80 inputs with 8 pixels of context to produce 128 × 128 target patches.Source: §4.1 · p. 5
• Latent Hyperloop combines latent reconstruction with refiner-attention alignment to preserve fine structure beyond what numerical latent distance alone captures.The attention objective penalizes structured mismatches that reconstruction loss may underweight.Source: §4.1 · pp. 5–6

4.2 Sparse Patch-based Decoding: Accelerating Final RGB Decoding

• Sparse patch decoding addresses seams caused by independently computed GroupNorm statistics by reusing shared cached statistics across patches.Consistent normalization makes identical latent values decode consistently across overlapping regions.Source: §4.2 · pp. 6–7
• Up to 8× smaller overlap reduces decoded patches by up to 30–40% in the main 4K setting, while boundary MSE stays low across overlap factors.Smaller overlap reduces patch count and VAE decoding time.Source: §4.2 · p. 7, Figure 3 · p. 7

5.1 Experimental Settings

• The evaluation uses 1,000 LAION-Aesthetics prompts, reports FID, FIDc, IS, and wall-clock time, and compares SDXL-based cascades including PixelRush.Latency is measured on one NVIDIA A100-40GB unless stated otherwise; PixelRush uses one-step Turbo refinement, unlike the multi-step refinement baselines.Source: §5.1 · pp. 7–8

5.2 Main Results on Fast Cascade Inference

• At 4K, UltraFlash-SD cuts latency from 16.37s to 8.04s and FID from 56.50 to 52.38; UltraFlash-XL cuts latency from 17.11s to 9.61s and FID from 54.28 to 51.32.The comparison is against the corresponding PixelRush configuration; Table 1 reports FID, FIDc, IS, and inference time.Source: Table 1 · p. 8, §5.2 · p. 8

5.3 Qualitative Comparison

• At 2K and 4K, UltraFlash shows sharper textures and finer details than the compared methods, with stronger structural coherence than FouriScale and DemoFusion.PixelRush and UltraFlash maintain robust global layout, while UltraFlash mitigates PixelRush’s subtle smoothing artifacts.Source: §5.3 · p. 8, Figure 4 · p. 9

5.4 Comparison with Non-Cascade 4K Routes

• In the contextual 4K comparison, UltraFlash-SD outperforms SANA-4K and DyPE-FLUX in FID and latency.The paper treats this as contextual because the methods differ in backbone, training data, and inference assumptions.Source: §5.4 · p. 9, Table 2 · p. 9

5.5 Plug-in Transfer

• Applying the transition operators to FreeScale improves FID from 58.28 to 54.42 and IS from 13.35 to 13.66, with runtime changing from 323.93s to 313.40s.The modest runtime change is attributed to FreeScale remaining dominated by multi-step denoising.Source: §5.5 · p. 9, Table 2 · p. 9

5.6 User Study

• In a study of 137 participants on 25 prompts, UltraFlash receives the highest preference share on Text Alignment, Image Quality, and Visual Structure.The comparisons total 3,425; the figure reports preference shares of approximately 30–33% for UltraFlash across the three criteria.Source: §5.6 · p. 10, Figure 5 · p. 10

6 Conclusion

• UltraFlash replaces the intermediate decode–resize–encode round trip with a learned latent shortcut and accelerates final RGB reconstruction using cached GroupNorm statistics.Both components target VAE-heavy transitions in cascade pipelines.Source: §6 · p. 10
• Integrated into a state-of-the-art cascade baseline, UltraFlash reduces 4K latency by 2× while improving image quality.The result supports pixel-space transitions as an effective target for improving the speed–quality trade-off.Source: §6 · p. 10

A.2 Pseudo Algorithm

• The pipeline upsamples latents by 2× at each cascade level and invokes the VAE only once, for final decoding.PixelRush invokes the VAE 2 × s times across s cascade levels.Source: §A.2 · p. 13, §A.1 · p. 13

A.3 Latent Hyperloop

• Latent Hyperloop uses a SwinIR-based architecture with VAE-decoder structural priors and generates 128 × 128 latent patches from 80 × 80 inputs.The adapter is trained from scratch for 20K steps using AdamW.Source: §A.3 · p. 14

A.4 Sparse VAE Decoding

• Sparse VAE decoding replaces decoder GroupNorm layers with modules that use cached reference statistics, enabling smaller overlaps across patches.The first batch supplies the cached statistics without a separate reference pass.Source: §A.4 · pp. 14, 16

B.1 Hybrid Noise Inversion

• Hybrid inversion combines DDIM-inverted noise with InvSR-predicted noise using spherical interpolation, with α = 0.8 weighting DDIM.The authors use Slerp to preserve the norm and angular distribution of the interpolated noise.Source: §B.1 · p. 16
• At 4K, hybrid inversion lowers FID from 52.38 to 51.34, at the cost of one additional InvSR forward pass per patch.The passage reports that the quality gains may be worthwhile when quality is prioritized over speed.Source: §B.1 · p. 16

B.2 Ablation Study

• Together, the components cut inference time to 8.04s from PixelRush’s 16.37s (2×) while improving visual quality.The full Latent Hyperloop also improves FID to 51.99 while reducing latency about 28% relative to the baseline.Source: Figure A1 · p. 17, §B.2 · p. 17
• UltraFlash consistently produces strong FID and IS scores across different base-model and refinement-module combinations.The robustness analysis varies both the base model and refinement module.Source: §B.2 · p. 17, Figure A1 · p. 17
• Attention Alignment improves convergence and decoded-image fidelity over the ℓ1 baseline and E-LatentLPIPS [11], while Spatial Prior further improves SSIM.The Spatial Prior reuses spatial expansion behavior already present in the VAE decoder.Source: §B.2 · p. 17, Figure A1 · p. 17
• UltraFlash offers significantly better latency scaling than PixelRush as resolution increases.Figure A2 compares inference latency across resolutions on a log scale.Source: Figure A2 · p. 18
• At p = 8, Latent Hyperloop’s FID falls from 59.18 at p = 0 to 51.99; increasing context to p = 12 or 16 changes FID by less than 0.35.The authors select p = 8 as a quality–cost trade-off.Source: Table A3 · p. 18, §B.2 · p. 18

C.1 Latent Hyperloop vs. Image Bilinear Interpolation

• Latent Hyperloop closely matches the pixel-space decode–bilinear-interpolation–encode output across diverse content types.The comparison decodes both the adapter output and the standard pixel-space transformation to image space.Source: §C.1 · p. 18

C.2 Effect of Training Objectives on Latent Hyperloop Reconstruction

• Attention Alignment yields sharper, more faithful reconstructions, especially in fine-grained facial and fur details, than ℓ1 alone or ℓ1 plus E-LatentLPIPS.The compared variants use the same latent upsampling backbone and training setting.Source: §C.2 · p. 18, Figure A4 · p. 19

C.3 Additional UltraFlash Outputs

• UltraFlash generates additional 4096 × 4096 outputs using FLUX and SDXL base models on a single A100 GPU.The examples demonstrate generalization across different base architectures.Source: §C.3 · p. 18
• Latent Hyperloop outputs are visually indistinguishable from the expensive pixel-space round trip in the comparison.The figure compares the adapter with image bilinear interpolation.Source: Figure A3 · p. 19

D Limitations

• UltraFlash is developed and evaluated for cascade-based megapixel image synthesis; video generation’s temporal and architectural considerations are outside this work’s scope.The authors identify transition-aware acceleration beyond image cascades as a natural direction for future work.Source: D Limitations · pp. 19–20

E Broader Impact

• The paper frames improved efficiency as potentially making megapixel generation more practical, while recommending the safeguards used for the underlying generative models.These include content filtering, provenance or watermarking mechanisms when available, and responsible-use policies.Source: E Broader Impact · p. 20
• The appendix includes uncurated 4096 × 4096 outputs from the UltraFlash refinement pipeline with FLUX and SDXL base models.The FLUX examples are in Figures A5–A6, and the SDXL examples are in Figures A7–A8.Source: Figure A5 · p. 21, Figure A6 · p. 22, Figure A7 · p. 23

NeurIPS Paper Checklist

• The checklist says the abstract and introduction frame the work as transition-efficient megapixel synthesis, with comparisons to native 4K and FLUX-based methods treated as contextual because of different backbones or training assumptions.It states that the claims are supported by the Section 5 experiments.Source: NeurIPS Paper Checklist · p. 25

Theory assumptions and proofs

• The paper presents no formal theoretical results or theorems; its mathematical expressions define cascade transition operators and training objectives.Source: §2 · p. 25

Experimental result reproducibility

• The checklist says the paper describes the models, prompts, metrics, resolutions, hardware, and baseline settings needed to reproduce the main results, with method and supplement details for both proposed components.Source: §2 · p. 25

Open access to data and code

• The current submission includes no public code or released checkpoints, with code and trained adapters planned for release upon acceptance.Source: §2 · p. 25

Experiment statistical significance

• The paper reports no error bars because each full 4K evaluation is computationally expensive and uses 1,000 prompts; comparisons share prompts, metrics, hardware, and protocol to reduce variance.The checklist answers that statistical significance is not reported.Source: §2 · p. 26

Experiments compute resources

• Inference experiments use a single NVIDIA A100-40GB GPU, and the paper reports wall-clock runtime for each method.It also reports Latent Hyperloop model size and inference time for reproducing the speed claims.Source: §2 · p. 26

Societal impacts and safeguards

• The paper uses existing pretrained image-generation models and public datasets or prompts, reports no sensitive personal-data collection or deployment, and recommends inherited safeguards such as content filtering and provenance mechanisms when available.It states that UltraFlash does not bypass the underlying models’ safety behavior.Source: §2 · p. 26

Crowdsourcing and research with human subjects

• The current submission omits full user-study interface screenshots and does not describe IRB or equivalent review; the authors say they will add participant-risk, consent, anonymity, and institutional details.The study is described as an unpaid, voluntary, low-risk anonymous preference evaluation.Source: §2 · p. 27

Declaration of LLM usage

• The core method does not use LLMs as an algorithmic component; any language-model assistance was limited to writing, editing, or formatting.Source: §2 · p. 27