While diffusion models excel at generating images with conventional dimensions, pushing them to synthesize ultra-high-resolution imagery at extreme aspect ratios (EAR) often triggers catastrophic structural failures, such as object repetition and spatial fragmentation. This limitation fundamentally stems from a lack of robust spatial priors, as static text-to-image models are primarily trained on image distributions with conventional dimensions. To overcome this bottleneck, we present ScrollScape, a novel framework that reformulates EAR image synthesis into a continuous video generation process through two core innovations. By mapping the spatial expansion of a massive canvas to the temporal evolution of video frames, ScrollScape leverages the inherent temporal consistency of video models as a powerful global constraint to ensure long-range structural integrity. Specifically, Scanning Positional Encoding (ScanPE) distributes global coordinates across frames to act as a flexible moving camera, while Scrolling Super-Resolution (ScrollSR) leverages video super-resolution priors to circumvent memory bottlenecks, efficiently scaling outputs to an unprecedented 32K resolution. Fine-tuned on a curated 3K multi-ratio image dataset, ScrollScape effectively aligns pre-trained video priors with the EAR generation task. Extensive evaluations demonstrate that it significantly outperforms existing image-diffusion baselines by eliminating severe localized artifacts. Consequently, our method overcomes inherent structural bottlenecks to ensure exceptional global coherence and visual fidelity across diverse domains at extreme scales.
Figures & tables
Figure 4: Qualitative results on 8:1 vertical scroll paintings. ScrollScape maintains superior global structural coherence and expansive diversity, effectively eliminating the boundary artifacts and semantic repetition found in existing frameworks.
Figure 5: Demonstration of high fidelity 8K imagery generated by ScrollScape. The results showcase the versatility of ScrollScape in producing high fidelity imagery with structural richness and local clarity across various subject matters, ranging from microscopic textures to macroscopic landscapes. Please zoom in for high resolution details.
Method
FID ↓
CLIP ↑
KID ↓(×10−2)
Style-L ↓(×10−3)
GSD ↑ (LPIPS)
GSD ↓ (DINOv2)
FLUX-Krea
333.7
20.9
20.4
6.0
0.174
0.822
FLUX-Krea+Fill
281.4
19.1
4.9
13.5
0.293
0.841
DyPE
248.1
24.6
4.7
5.5
0.569
0.682
MultiDiffusion
261.7
29.7
3.7
5.0
0.658
0.902
SyncDiffusion
245.2
26.5
3.2
4.7
0.618
0.895
Tiled Diffusion
241.2
27.3
3.0
4.5
0.591
0.901
Table 1: Quantitative comparison. ScrollScape surpasses alternative approaches across all metrics. The best results are bolded .
Method Comparison
Structural Coh.
Content Rich.
Image Qual.
Ours vs DyPE
92%
89%
87%
Ours vs MultiDiffusion
83%
76%
82%
Ours vs SyncDiffusion
85%
74%
78%
Ours vs Tiled Diffusion
76%
85%
74%
Table 2: User Preference Study. The percentages denote the fraction of raters who prefer our results over each baseline.
Method
FID ↓
CLIP ↑
KID ↓(×10−2)
Style-L ↓(×10−3)
GSD ↑ (LPIPS)
GSD ↓ (DINOv2)
Wan2.1
246.5
25.9
4.7
2.0
0.303
0.975
w/o Training
272.0
24.6
6.3
7.6
0.569
0.858
w/o TAP
231.6
25.6
3.2
6.0
0.469
0.730
w/o ScrollSR
218.8
26.5
2.4
4.9
0.668
0.671
ScrollScape (Ours)
214.7
30.0
2.0
4.0
0.674
0.670
Table 3: Quantitative ablation study. While Wan2.1 achieves an outstanding Style-L score due to mode repetition, ScrollScape demonstrates superior performance across most metrics, ensuring both high fidelity and global structural diversity.
Figure 6: Qualitative ablation study on the effectiveness of ScanPE. From top to bottom: (a) the vanilla Wan2.1 exhibiting severe content repetition; (b) w/o training, where ScanPE successfully breaks repetitive patterns but yields chaotic textures; (c) w/o TAP showing significant blurring due to decoding limitations; and (d) ScrollScape producing seamless panoramas with structural coherence and rich details.
Figure 7: Qualitative ablation study on the effectiveness of ScrollSR. The baseline w/o ScrollSR (top) suffers from perceptual blurriness in the lotus petals. In contrast, our full framework w/ ScrollSR (bottom) successfully restores fine grained textures and structural definition, as evidenced by the sharp details and high resolution clarity in the zoomed in patches.
Recent diffusion-based video generators have achieved remarkable visual fidelity and prompt controllability, yet scaling them to ultra-high-resolution (UHR) long videos remains prohibitively expensive. The difficulty is especially pronounced for long single-shot generation where a continuous scene must preserve global temporal coherence, and fine-grained spatial details without relying on clip transitions or autoregressive shot stitching. In this work, we revisit this challenge from the perspective of decoupled modeling. We argue that existing video diffusion models already encode strong local visual priors, while the main bottleneck lies in efficiently extending global spatiotemporal modeling as resolution and duration increase. Based on this insight, we propose AtlaVid, a decoupled global-local framework for efficient UHR long video generation. AtlaVid first generates a low-resolution and low-FPS global semantic proxy via temporally scaled RoPE, thereby extending the temporal horizon without increasing the training token count. Guided by this proxy, a high-resolution detail branch performs joint denoising with hierarchical locality-preserving attention. Reordered spatiotemporal windows preserve geometric locality and asymmetric global-local attention injects aligned semantic guidance and preserves the model's pretrained ability. This design enables resolution-agnostic training: the model is trained only at 720P with lightweight LoRA adaptation, yet generalizes directly to 4K and beyond for longer (>10s) video synthesis. Experiments show that AtlaVid substantially improves the efficiency of ultra-high-resolution long video generation, achieving high-quality UHR long video generation with 60.9x speed up and significantly less training cost and even better performance than native 4K video generators.
Video diffusion models have achieved remarkable progress in recent years, yet generating high-resolution videos remain a fundamental challenge. Most video generators are trained at limited spatial resolutions due to the scarcity of high-resolution 4K video data and the prohibitive computational cost of large-scale training on such data. Most video diffusion models are trained on 720p videos and are therefore effectively limited to generating videos at similar resolutions during inference. To address this gap, we propose CineScale. CineScale, to the best of our knowledge, is the first tuning-free inference framework enabling pretrained video diffusion models to generate high-quality videos at resolutions far beyond those seen during training. Our key observation is that generation quality degrades at higher resolutions because positional encodings shift beyond their training distribution, producing blurred details and structurally incoherent videos. To address this gap, we introduce Adaptively Rectified RoPE. Our extensive experiments show that CineScale enables pretrained diffusion models, despite never being trained on high-resolution data, to generate high-fidelity 4K video without any fine-tuning, improving local detail and sharpness while preserving temporal coherence. This demonstrates that high-resolution generation capabilities can be unlocked purely at inference time.
Gordon Chen, Haonan Qiu, Ning Yu +3
Nanyang Technological University · Netflix Eyeline Studios
Diffusion-based image-to-video (I2V) models are increasingly effective, yet they struggle to scale to ultra-high-resolution inputs (e.g., 4K). Generating videos at the model's native resolution often loses fine-grained structure, whereas high-resolution tiled denoising preserves local detail but breaks global layout consistency. This failure mode is particularly severe in the fresco animation setting: monumental artworks containing many distinct characters, objects, and semantically different sub-scenes that must remain spatially coherent over time. We introduce FrescoDiffusion, a training-free method for coherent large-format I2V generation from a single complex image. The key idea is to augment tiled denoising with a precomputed latent prior: we first generate a low-resolution video at the underlying model resolution and upsample its latent trajectory to obtain a global reference that captures long-range temporal and spatial structure. For 4K generation, we compute per-tile noise predictions and fuse them with this reference at every diffusion timestep by minimizing a single weighted least-squares objective in model-output space. The objective combines a standard tile-merging criterion with our regularization term, yielding a closed-form fusion update that strengthens global coherence while retaining fine detail. We additionally provide a spatial regularization variable that enables region-level control over where motion is allowed. Experiments on the VBench-I2V dataset and our proposed fresco I2V dataset show improved global consistency and fidelity over tiled baselines, while being computationally efficient. Our regularization enables explicit controllability of the trade-off between creativity and consistency.
Hugo Caselles-Dupré, Mathis Koroglu, Guillaume Jeanneret +2
Obvious Research, Paris, France · Sorbonne Université, CNRS, ISIR, F-75005 Paris, France · Valeo.ai, France