Skip to content

ScrollScape: Unlocking 32K Image Generation With Video Diffusion Priors

Conference: ECCV 2026
Paper: ECCV Official Link
Area: Image Generation
Keywords: extreme aspect ratio generation, video diffusion priors, rotary positional embedding, latent super-resolution, panoramic image synthesis

TL;DR

ScrollScape reformulates the synthesis of ultra-high-resolution images with extreme aspect ratios (e.g., 8:1) as a sequential video panning task, leveraging the temporal consistency of pre-trained video diffusion models as a global structural constraint, combined with Scanning Positional Encoding (ScanPE) and Scrolling Super-Resolution (ScrollSR) to seamlessly scale generation up to 32K resolution while completely eliminating localized repetition and tiling seams.

Background & Motivation

Text-to-image diffusion models have achieved remarkable photorealism across conventional canvas dimensions, yet pushing them to generate imagery at extreme aspect ratios (EAR, such as 8:1 or 1:8)—like ultra-wide landscape panoramas or traditional cultural scrolls—invariably triggers catastrophic structural failures. Because standard foundational models are predominantly trained on moderate resolutions and conventional aspect ratios, they internalize static, frame-centric spatial priors and lack the long-range dependency modeling necessary to orchestrate coherence over expansive canvases spanning tens of thousands of pixels.

To synthesize beyond native canvas dimensions, existing solutions typically resort to two compromise paradigms: window-based tiling methods (such as SyncDiffusion, MultiDiffusion, and Tiled Diffusion) that partition the target canvas into overlapping patches, which fundamentally lack a holistic spatial prior and inevitably yield disjointed compositions and seam artifacts; and training-free extrapolation approaches (like ScaleCrafter and DyPE) that manipulate internal representations via dilated convolutions or position embedding interpolation. While these modifications widen the receptive field, perturbing the original coordinate distribution severely destabilizes generation, frequently producing semantic looping and duplicate objects like multiple suns or repeated mountain clusters.

The core tension lies in the fact that static image diffusion models rely intrinsically on stationary viewport correlations, rendering them incapable of autonomously maintaining continuous spatial narratives across extreme aspect ratios without explicit spatial motion constraints. The angle of attack in this paper transcends the static image paradigm altogether: the physical expansion of a continuous panoramic scroll is perceptually equivalent to a camera executing a smooth panning trajectory, whereas video diffusion models (such as Wan2.1) have already internalized powerful spatiotemporal continuity and physical constraints through pre-training on massive dynamic sequences. Core idea: reformulate extreme aspect ratio image generation as a continuous video panning process, utilizing video temporal consistency as a macro spatial regularizer and pairing dynamic global coordinate encoding with latent-space super-resolution to synthesize seamless scrolls up to 32K resolution.

Method

Overall Architecture

ScrollScape translates ultra-high-resolution EAR image synthesis into a structured pipeline of latent-space temporal panning, coordinate-guided diffusion, and video super-resolution. Given an expansive target canvas, the framework first decomposes the latent space into an overlapping sequence of local scanning windows, where ScanPE assigns dynamic, absolute global coordinates to every token. A video diffusion Transformer (initialized from Wan2.1-T2V-1.3B) then generates a coherent sequence of low-resolution latents. Subsequently, ScrollSR applies video super-resolution priors directly within the latent domain to enrich high-frequency textures without exceeding GPU memory budgets. Finally, coordinate-aligned 3D VAE decoding and median-consensus ramp blending fuse the sequential tiles into a seamless 32K panorama.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Text Prompt & Scanning Trajectory Configuration"] --> B["Window Trajectory Decomposition & Global Anchoring"]
    B --> C["Scanning Positional Encoding ScanPE<br/>Spatiotemporally Decoupled RoPE Mapping"]
    C --> D["Video Diffusion DiT Generation<br/>Low-Resolution Latent Panning Sequence"]
    D --> E["Scrolling Super-Resolution ScrollSR<br/>Latent-Domain Sequential Refinement"]
    E --> F["Coordinate-Aligned 3D VAE Decoding<br/>Grouped Anchors & Median Feature Consensus"]
    F --> G["Panoramic Frame Fusion<br/>Ramp-Weighted Seamless 32K Output"]

Key Designs

1. Spatial Extension as Sequential Scanning: Video Panning over Tiled Montages Traditional tiling frameworks generate patches either independently or with weak local gradient consensus, easily losing the macroscopic structure of expansive canvases. ScrollScape represents the entire target canvas in the latent space as \(z \in \mathbb{R}^{H \times W \times C}\) and decomposes it along the scanning trajectory into a sequence of \(T\) overlapping local windows \(\mathcal{S} = \{z_t\}_{t=1}^T\), each covering an \(h \times l \times C\) region. Each window \(t\) is anchored to a global canvas coordinate \(O_t\): $\(O_t = P_{\text{init}} + \sum_{k=1}^{t-1} \Delta \cdot \mathbf{d}_k\)$ where \(P_{\text{init}}\) denotes the canvas origin, \(\Delta\) represents the scanning stride, and \(\mathbf{d}_k \in \{(\pm 1, 0), (0, \pm 1)\}\) defines the unit scanning direction at step \(k\). For standard horizontal 8:1 panoramas, \(\mathbf{d}_k = (0, 1)\) simplifies the horizontal offset to \((t-1)\Delta\); for non-standard 2D layouts, the direction sequence periodically reverses at canvas borders to form a snake-like trajectory. By mapping spatial extension directly onto the temporal axis, the framework converts static image synthesis into a smooth moving-camera capture.

2. Scanning Positional Encoding (ScanPE): Uncoupling Fixed-Viewport Constraints Standard 3D-RoPE in video Transformers assumes that all frames share an identical, stationary camera coordinate frame, assigning the exact same spatial rotation frequencies to tokens at identical local indices across all temporal steps. Panning under standard 3D-RoPE causes severe positional misalignment and repeating patterns. ScanPE resolves this by dynamically mapping any local token coordinate \(p_{\text{loc}} = (h_{\text{loc}}, w_{\text{loc}})\) within the \(t\)-th window to its global canvas position \(P_g(t, p_{\text{loc}}) = p_{\text{loc}} + O_t = (H_g, W_g)\). The attention head dimension is divided into three distinct orthogonal subspaces for temporal, vertical, and horizontal coordinates, concatenating axis-wise rotary embeddings: $\(\mathcal{R}(t, H_g, W_g) = \Theta_t(t; \theta^t) \oplus \Theta_h(H_g; \theta^h) \oplus \Theta_w(W_g; \theta^w)\)$ Tokens from different scanning steps are thus indexed by their actual, physical positions across the expansive canvas rather than identical frame-centric grids, granting the self-attention layers accurate spatial awareness across tens of thousands of pixels.

3. Scrolling Super-Resolution (ScrollSR) and Coordinate-Aligned Decoding: Bypassing Memory Walls Directly synthesizing 32K pixels in pixel space triggers immediate GPU out-of-memory errors, while running standard 3D VAE decoders across lengthy sequences causes noticeable inter-frame boundary flickers and appearance drift. ScrollSR performs super-resolution directly on the low-resolution latent representations generated by the video panning DiT, leveraging video super-resolution priors (adapted from FlashVSR) to enhance high-frequency latent details frame by frame. Before decoding into pixels, latent blocks are grouped according to their global anchors \(O_t\): anchor blocks provide a consistent, stable decoding context, ensuring that adjacent overlapping windows share identical coordinate references and preventing geometric distortion.

4. Median Feature Consensus Frame Fusion: Eliminating Transient Seams When reconstructing the final panorama from decoded tiles, overlapping boundaries can exhibit subtle decoding discrepancies. For any spatial window \(t\) producing \(F\) candidate decoded tiles \(\{I_{t,i}\}_{i=1}^F\) across local contexts, ScrollScape computes the channel-wise median feature vector \(\mathbf{m}_t = \operatorname{Median}_{k=1}^F f(I_{t,k})\), where \(f(\cdot)\) extracts feature statistics, and selects the representative tile closest to this consensus: $\(i_t^* = \arg\min_{i \in \{1,\ldots,F\}} \|f(I_{t,i}) - \mathbf{m}_t\|_2, \quad I_t = I_{t, i_t^*}\)$ This selection effectively filters out transient decoding outliers while retaining sharp structural edges. Finally, pixel-level distance-based ramp blending masks \(M_t\) smoothly fuse adjacent tiles into the final 32K panorama without visible seams.

Loss & Training

While pre-trained video diffusion models encapsulate robust dynamic priors, a domain gap remains between dynamic video distributions and ultra-high-resolution static canvases. ScrollScape employs conditional flow matching for lightweight task adaptation. Defining a linear interpolation path between noise \(\epsilon \sim \mathcal{N}(0, I)\) and data latent \(z^{\text{data}}\): $\(z_\tau = (1 - \tau)\epsilon + \tau z^{\text{data}}, \quad \tau \in [0, 1]\)$ the model is optimized to predict the vector field velocity: $\(\mathcal{L}_{\text{FM}} = \mathbb{E}_{z^{\text{data}}, \epsilon, \tau} \left\| v_\theta(z_\tau, \tau, c, \mathcal{R}) - (z^{\text{data}} - \epsilon) \right\|_2^2\)$ where \(c\) is the text prompt and \(\mathcal{R}\) is the ScanPE embedding. Initialized from Wan2.1-T2V-1.3B, the model requires fine-tuning on only 3,000 curated high-resolution panoramic images (2,000 natural landscapes at \(\ge 6:1\) and 1,000 Chinese ink paintings at 6:1) on 2 \(\times\) A100 GPUs for 10,000 iterations (batch size 4, AdamW learning rate \(1 \times 10^{-5}\)), demonstrating exceptional parameter and sample efficiency.

Key Experimental Results

Main Results

To evaluate extreme aspect ratios without distortion from non-uniform downsampling, the authors partition 8:1 panoramas (with a short-side resolution of 1024 pixels) into non-overlapping 1:1 square patches. Quantitative metrics include patch-averaged FID and KID (evaluated against Aesthetic Eval), text-image alignment via CLIP, and Style-L measuring stylistic consistency across adjacent windows. Crucially, Global Structural Diversity (GSD) is evaluated through LPIPS (measuring perceptual texture divergence, higher is better) and DINOv2 cosine similarity (detecting repetitive semantic patterns, lower is better). The results from Table 1 of the paper are reported below.

Method FID ↓ CLIP ↑ KID ↓ \((\times 10^{-2})\) Style-L ↓ \((\times 10^{-3})\) 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
ScrollScape (Ours) 214.7 30.0 2.0 4.0 0.674 0.670

Furthermore, an anonymized user study across 100 prompts and 20 independent raters (Table 2 of the paper) reveals that ScrollScape is overwhelmingly preferred over all competitive baselines across structural coherence, content richness, and overall image quality (e.g., scoring 92% structural coherence, 89% content richness, and 87% image quality against DyPE).

Ablation Study

The ablation experiments isolate the contributions of ScanPE, alignment fine-tuning, coordinate-aligned latent decoding, and ScrollSR. Numerical results from Table 3 of the paper are detailed below.

Config FID ↓ CLIP ↑ KID ↓ \((\times 10^{-2})\) Style-L ↓ \((\times 10^{-3})\) GSD ↑ (LPIPS) GSD ↓ (DINOv2) Note
Wan2.1 (vanilla) 246.5 25.9 4.7 2.0 0.303 0.975 Very low Style-L solely due to extreme semantic self-duplication
w/o Training 272.0 24.6 6.3 7.6 0.569 0.858 ScanPE breaks duplication, but outputs chaotic textures and high noise
Vanilla latents 231.6 25.6 3.2 6.0 0.469 0.730 Unaligned VAE decoding causes blurring and boundary degradation
w/o ScrollSR 218.8 26.5 2.4 4.9 0.668 0.671 Maintains global coherence but exhibits visibly soft textures
ScrollScape (full model) 214.7 30.0 2.0 4.0 0.674 0.670 Optimal performance across all metrics, balancing fidelity and diversity

Key Findings

  • Video priors break mode repetition when paired with ScanPE: Vanilla Wan2.1 suffers from severe content duplication across extreme aspect ratios (DINOv2 similarity of 0.975). ScanPE successfully anchors tokens onto the physical canvas coordinate space, driving DINOv2 redundancy down to 0.670 while elevating perceptual diversity LPIPS to 0.674.
  • Lightweight flow matching aligns static realism: Generating EAR canvases zero-shot without training produces noisy, disordered visual artifacts (FID degraded to 272.0). Fine-tuning on 3,000 curated panoramic samples effectively aligns dynamic video representations into clean static visual structures.
  • Latent super-resolution circumvents memory ceilings: ScrollSR enables scaling to 32K without GPU out-of-memory failure, improving FID from 218.8 to 214.7 and recovering crisp microscopic details such as beetle shells, ice crystals, and delicate lotus petals.

Highlights & Insights

  • Spatial extension reformulated as temporal scanning: Translating the 2D spatial coherence problem of ultra-wide canvases into a 1D continuous video progression repurposes pre-trained temporal self-attention as a macro-level spatial regularizer, sidestepping the severe tiling artifacts of image models.
  • Elegant, zero-overhead ScanPE design: By decomposing RoPE into independent temporal, vertical, and horizontal rotation operators and superimposing global scanning offsets, ScanPE seamlessly transforms a stationary video camera into a dynamic scanning viewer without architectural changes or extra parameters.
  • Modular latent-to-pixel pipeline for extreme scale: Decoupling low-resolution latent trajectory generation, latent-domain video super-resolution, and coordinate-aligned median-consensus frame blending provides a scalable blueprint for generating ultra-high-resolution images on consumer-grade hardware.

Limitations & Future Work

  • Global prompt constraints over evolving narratives: The framework relies primarily on a single global text prompt, making it challenging to steer spatially evolving narrative sequences where different scroll sections depict distinct semantic themes (e.g., transitioning from quiet rural landscapes to bustling marketplaces).
  • Trajectory flexibility primarily along straight axes: While snake-like scanning covers 2D grids, arbitrary curvilinear trajectories or complex omnidirectional non-Euclidean camera motions remain to be explored.
  • Future directions: Integrating multimodal LLMs (MLLMs) to dynamically plan localized sub-prompts across canvas segments, guiding region-specific cross-attention to produce rich, multi-chapter visual epics at 32K scale.
  • vs SyncDiffusion / MultiDiffusion / Tiled Diffusion: Tiling methods rely on overlapping blending or gradient synchronization across isolated windows, which fundamentally lack a holistic generative prior and frequently produce seam artifacts and semantic duplication at 8:1 aspect ratios. ScrollScape's continuous panning video prior natively guarantees smooth macroscopic transitions.
  • vs DyPE / ScaleCrafter: Training-free position extrapolation and dilated convolutions perturb the learned spatial attention patterns of static image models, causing duplicated subjects and structural disintegration. ScrollScape preserves the native attention mechanics of video DiT, introducing only physical coordinate shifts via ScanPE and achieving superior stability with minimal fine-tuning.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Transforming extreme aspect ratio generation into a continuous video panning task is highly creative and effectively unlocks 32K scale.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluations across 8:1 horizontal and vertical scrolls, rigorous patch-based metrics including GSD, detailed ablations, and a blind user study.
  • Writing Quality: ⭐⭐⭐⭐⭐ Clear motivation, rigorous mathematical formulation across all stages, and well-structured qualitative and quantitative demonstrations.
  • Value: ⭐⭐⭐⭐⭐ Offers a transformative framework for ultra-high-resolution image synthesis, digital art preservation, and wide-format graphic generation.