CineScale: Open-Sourcing Tuning-Free High-Resolution Video Generation

Gordon Chen^1,*, Haonan Qiu^1,2,*, Ning Yu^2,✉, Ziqi Huang¹, Paul Debevec², Ziwei Liu^1,✉

¹S-Lab, Nanyang Technological University ²Netflix Eyeline Studios

^*Equal contribution ^✉Corresponding authors

Watch in 4K

CineScale is now integrated into Wan 🎉

TL;DR

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-fidelity videos at resolutions far beyond those encountered during training, without any fine-tuning.

CineScale framework with tiled self-attention and Adaptively Rectified RoPE — Framework overview: tiled self-attention with Adaptively Rectified RoPE.

Method

CineScale enables pretrained video diffusion models to generate far beyond their native training resolution without retraining. It partitions query tokens into spatial tiles while retaining access to all global keys and values, then uses Adaptively Rectified RoPE to preserve precise local geometry and keep long-range positional offsets within the range supported by the pretrained model.

Tiled Self-Attention

We partition the query tokens into M non-overlapping tiles {T_m}^M_m=1, with each tile covering the same spatial region across every video frame. If I_m contains the token indices assigned to tile m, its queries are:

Q_m = Q[I_m, :]

Each query tile attends to the complete set of keys and values from the video. Tiling therefore reduces the query-side working set without restricting global context:

O_m = Softmax(Q̃_m(K̃_all^(m))^T√d_h) V_all

Here, Q̃_m and K̃_all^(m) are the positionally encoded queries and global keys, V_all contains every value token, and d_h is the attention-head dimension. With standard RoPE, processing the query tiles independently and concatenating their outputs is mathematically equivalent to conventional global self-attention. The tile outputs are scattered back to their original locations as O[I_m, :] = O_m. CineScale changes only how the global keys are positionally encoded for each tile.

Adaptively Rectified RoPE

At high resolutions, standard RoPE encounters relative spatial offsets far outside its training distribution. For every query tile, Adaptively Rectified RoPE leaves nearby key coordinates unchanged and linearly compresses more distant coordinates toward the tile.

Consider a vertical key coordinate p_y in [0, L_y - 1]. Let [ℓ_y^(m), u_y^(m)] be the vertical extent of tile m, D_y^(m) = u_y^(m) - ℓ_y^(m) its height, and R_y the largest relative offset supported by the pretrained model. The remaining external positional budget is:

B_y^(m) = R_y - D_y^(m)

Given a local preservation threshold τ_y, the unchanged margins on either side of the tile are:

δ_y,-^(m) = min(τ_y, ℓ_y^(m), B_y^(m)),
δ_y,+^(m) = min(τ_y, L_y - 1 - u_y^(m), B_y^(m))

These margins define the interval [κ_y,-^(m), κ_y,+^(m)] whose coordinates remain unchanged:

κ_y,-^(m) = ℓ_y^(m) - δ_y,-^(m),
κ_y,+^(m) = u_y^(m) + δ_y,+^(m)

Coordinates outside that interval are compressed toward the target endpoints γ_y,-^(m) and γ_y,+^(m):

γ_y,-^(m) = max(0, u_y^(m) - R_y),
γ_y,+^(m) = min(L_y - 1, ℓ_y^(m) + R_y)

The corresponding compression slopes are:

λ_y,-^(m) = κ_y,-^(m) - γ_y,-^(m)κ_y,-^(m), λ_y,+^(m) = γ_y,+^(m) - κ_y,+^(m)L_y - 1 - κ_y,+^(m)

With every term now defined, the vertical coordinate mapping is:

ρ_y^(m)(p_y) = γ_y,-^(m) + λ_y,-^(m)p_y, if p_y < κ_y,-^(m) p_y, if κ_y,-^(m) ≤ p_y ≤ κ_y,+^(m) κ_y,+^(m) + λ_y,+^(m)(p_y - κ_y,+^(m)), if p_y > κ_y,+^(m)

The same mapping is applied along the horizontal axis. Query features and their original temporal-spatial coordinates remain unchanged; only the coordinates assigned to keys are rectified relative to the current query tile:

Q̃_m = RoPE(t, y, x) Q_m

K̃_all^(m) = RoPE(t, ρ_y^(m)(y), ρ_x^(m)(x)) K_all

This tile-conditioned encoding preserves local spatial relationships while bounding long-range relative offsets, allowing stable global interactions and coherent detail at resolutions up to 4K.

Results and Ablations

Visual ablation comparing RoPE, NTK-RoPE, and Adaptively Rectified RoPE — RoPE ablation: standard RoPE, NTK-RoPE, and Adaptively Rectified RoPE.

4K video generation comparison across Kling 3.0, LTX 2.0, Wan 2.2, and CineScale — 4K comparison across recent video generation systems.

4K Video Gallery

Hover to Zoom In

Citation

If you find CineScale useful in your research or projects, consider citing our paper:

@article{qiu2025cinescale,
  title={CineScale: Free Lunch in High-Resolution Cinematic Visual Generation},
  author={Haonan Qiu and Ning Yu and Ziqi Huang and Paul Debevec and Ziwei Liu},
  journal={arXiv preprint arXiv:2508.15774},
  year={2025}
}