Pith. sign in

REVIEW 5 major objections 5 minor 11 cited by

A three-stage latent-cascaded pipeline outperforms previous UHR video generators on quality and efficiency.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 00:04 UTC pith:SQQI7IUB

load-bearing objection A genuinely new three-stage architecture for UHR video generation, but the evaluation is weaker than the abstract claims and the refinement stage may never have seen its test-time input distribution. the 5 major comments →

arxiv 2602.11564 v2 pith:SQQI7IUB submitted 2026-02-12 cs.CV

LUVE : Latent-Cascaded Ultra-High-Resolution Video Generation with Dual Frequency Experts

Chen Zhao , Jiawei Chen , Hongyu Li , Zhuoliang Kang , Shilin Lu , Xiaoming Wei , Kai Zhang , Jian Yang
show 1 more author
Ying Tai
This is my paper
classification cs.CV
keywords ultra-high-resolution video generationlatent upsamplingimplicit neural representationfrequency-specialized expertsvideo diffusionmotion modelingsemantic coherencetemporal consistency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes LUVE, a three-stage framework for generating 2K/4K video from text prompts. It argues that scaling video diffusion models to ultra-high resolution fails because motion modeling, semantic planning, and detail synthesis interfere with one another, and that separating them into dedicated stages resolves the conflict. The first stage generates low-resolution video latents with a frozen text-to-video model, the second up-samples those latents directly in latent space using a lightweight implicit neural representation, and the third applies two frequency-specialized experts to refine semantics and detail. If the central claims hold, UHR video generation no longer requires expensive end-to-end high-resolution fine-tuning, and the output preserves both genuine motion and semantic coherence. The paper reports top scores on established video benchmarks, on patch-level fidelity, and in a 20-participant human preference study, while also acknowledging that overall computational cost remains a practical bottleneck.

Core claim

On its own terms, LUVE's central discovery is that the high-resolution refinement stage of a cascaded video generator should act as a content completer, not a perceptual enhancer. The paper shows that a low-resolution base model can supply reliable motion priors; a 22-million-parameter latent upsampler then maps those latents directly to higher resolution without decoding to pixels; and two LoRA-based experts, activated during different denoising phases and steered by frequency filtering, repair global semantics and synthesize fine-grained texture. The paper claims that this combination yields superior photorealism and content fidelity in UHR video generation, outperforming existing UHR text

What carries the argument

The load-bearing pieces are two. First, the video latent upsampler (VLUer), an INR-based encoder-upsampler-decoder trained with a latent reconstruction loss plus pixel-space and frame-difference losses, so that low-resolution latents land on the high-resolution latent manifold of the frozen video autoencoder. Second, the dual frequency experts: a low-frequency LoRA branch attached to attention blocks and active in the high-noise denoising phase, aimed at global semantic coherence, and a high-frequency LoRA branch attached to feed-forward blocks and active in the low-noise phase, aimed at fine detail and texture.

Load-bearing premise

The learned latent upsampler must place low-resolution video codes inside the valid region of the high-resolution latent space; if it drifts off that manifold, the refinement stage is left repairing corrupted input rather than adding detail.

What would settle it

Train or evaluate VLUer on a held-out set of low-resolution videos, decode its output with the frozen video autoencoder, and compare its latent codes against the true high-resolution latent encodings using a standard distribution-distance measure. If the latent shift remains large even when decoded-frame PSNR is high, the claimed manifold alignment is not occurring and the later refinement stage is carrying most of the semantic load.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If LUVE's decomposition works, UHR video generation no longer requires end-to-end fine-tuning of a large diffusion model at 2K/4K, reducing GPU memory and inference time.
  • Motion consistency should improve because complex dynamics are planned at low resolution, where temporal modeling is more reliable, then carried through latent upsampling rather than re-synthesized at high resolution.
  • Semantic coherence and texture detail become independently controllable: data filtering targets semantics for the low-frequency expert, while unsharp-mask augmentation supplies explicit high-frequency detail signals for the other expert.
  • Conventional generate-then-upscale pipelines that only sharpen frames should be replaced by refinement stages that add semantic content, since LUVE reports both sharper output and better text-video alignment.
  • The learned latent upsampler enables arbitrary-scale upsampling without repeatedly running the video autoencoder's decoder and encoder, a direct and measurable computational saving.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • One extension the paper leaves implicit is whether VLUer transfers across different video diffusion backbones: because it only touches the frozen autoencoder's latent space, it might be reusable when a stronger base text-to-video model is released, but that reuse is not tested here.
  • The frequency-expert split by denoising phase suggests a general recipe for adapting any diffusion model to higher resolution by routing global-semantic and local-detail work through separate parameter-efficient adapters; the paper demonstrates it only for video.
  • The reconstruction table shows that RGB interpolation actually wins in pixel-space PSNR, so the case for VLUer rests on latent-space fidelity, downstream video quality, and efficiency. A natural follow-up is to measure whether latent-space quality, not pixel PSNR, is what predicts final perceptual quality at different upsampling factors.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes LUVE, a three-stage cascaded framework for ultra-high-resolution (2K/4K) text-to-video generation built on Wan2.1-1.3B. Stage 1 (LMG) generates low-resolution motion priors; Stage 2 (VLU) upsamples latents directly in latent space with a lightweight INR-based Video Latent Upsampler; Stage 3 (HCR) applies low-frequency and high-frequency LoRA experts to refine semantic coherence and fine details. The authors report state-of-the-art results on VBench metrics, a self-defined FIDpatch, an MLLM-based realism/detail/alignment evaluation using Doubao-1.5 Pro, and a 20-participant human study, with ablations on the upsampler, frequency experts, data selection, and skipped steps.

Significance. If the central claims hold, the paper makes a useful practical contribution: it shows how a 720p/1K T2V model can be adapted to 2K/4K video generation without end-to-end UHR fine-tuning, using a lightweight latent upsampler and parameter-efficient frequency-decomposed experts. The three-stage design is coherent, and the ablations in Tables 5–6 and the human-study confidence intervals in Table 13 provide some internal evidence for the contributions. However, the strongest quantitative claims rest on point estimates without uncertainty, a proprietary MLLM judge, a self-designed metric from the authors' own prior work, and a potentially untested training/inference distribution gap between the VLUer and HCR stages. The paper is currently a promising system paper, but several load-bearing points need to be addressed before the SOTA claim is fully supported.

major comments (5)
  1. [§3.3–3.4, §4.1, Table 12] The HCR stage is trained on real high-resolution latents (Section 4.1: 15K iterations of UHR scaling plus 3K iterations per expert), while at inference it receives latents produced by the VLUer after renoise-and-skip (Section 3.4). The VLUer is validated only by PSNR/MSE reconstruction on 60 UltraWan clips (Table 12) and is not shown to place its outputs on the Wan2.1 HR latent manifold. If VLUer outputs drift off-manifold, HCR—the only stage that can repair semantic or detail errors—has never seen such inputs during training. This is load-bearing because the reported gains could then stem from an unvalidated distributional assumption rather than the dual-expert design. Please add a direct analysis of the VLUer output distribution (e.g., latent-space MMD/FID, or decode-reencode consistency) or, ideally, fine-tune HCR on VLUer-upsampled latents and report whether the results change.
  2. [Tables 1–3, 5–6] All main quantitative comparisons are reported as point estimates with no error bars, confidence intervals, or repeated-seed statistics. Several differences are small relative to likely run-to-run variation—e.g., Table 1 average 84.34 vs. 83.75, and Table 1 SC 95.83 vs. 95.86. Since the paper's central claim is state-of-the-art performance, the lack of uncertainty quantification is not a minor omission. Please report at least three independent generations per configuration (or per checkpoint) with mean±std, or bootstrap CIs, for all headline metrics.
  3. [Table 2, Appendix H] The Realism/Detailness/Alignment scores are obtained from a proprietary commercial MLLM, Doubao-1.5 Pro, using custom rubrics. The paper does not report the model version, API settings, decoding parameters, the number of evaluation videos, or the raw per-video scores. This makes Table 2 effectively irreproducible. Please release all MLLM inputs/outputs and evaluation code, and as a sanity check, also report scores from an open-source multimodal judge (e.g., Qwen2.5-VL or InternVL) with the same rubric.
  4. [Appendix F, Table 2] FIDpatch is a metric proposed in the authors' own prior work (Zhao et al., 2025d), and here it is evaluated against a reference set (UltraHR-eval4k) from the same group. This is acceptable in principle, but the paper provides no validation that FIDpatch correlates with human preference for UHR video, and no evidence that the chosen patch size, number of patches, or reference distribution do not systematically disadvantage the baselines. Please include a patch-level FID for the baselines with the same reference distribution, and report the number of patches per video and the reference set size.
  5. [§4.1, Appendix A, Table 7] Several free parameters are central to the method: the switching timestep tswitch=0.417, the HPS v3 threshold of 6.5, the skipped steps S=5, and the VLUer training scales. Only S is ablated (Table 7). Since the dual-expert design is defined by the split of the noise schedule at tswitch, and the data curation depends on the HPS threshold, the paper should include a sensitivity analysis for these two parameters (e.g., tswitch ∈ {0.3, 0.5, 0.7} and threshold around 6.5). This would substantiate the claim that the chosen values are not overfit to VBench.
minor comments (5)
  1. [Eq. (4)–(5)] The LowPass and HighPass operators are not defined. Please specify the exact filters (e.g., kernel size, frequency cutoffs) and whether they are applied per-channel along the feature dimension.
  2. [Abstract and Section 1] The project link is inconsistent: the abstract says https://github.io/LUVE/ while the text says https://unicornanrocinu.github.io/LUVE_web/. Please unify.
  3. [Table 7 / Table 10] Table 10 is captioned 'Ablation study with different skipped steps' but it actually reports efficiency differences with/without experts. The caption should be corrected.
  4. [Appendix G] The related-work appendix contains several citation typesetting errors, e.g., 'Zhang et al., 2024b;b;' and 'Peng et al.; 2024a'. These should be cleaned up in a final revision.
  5. [Appendix E] The human study reports 20 participants but gives no information about their recruitment, prior experience with AI-generated video, or inter-rater agreement. Given the small sample, a brief description of participant selection would be helpful.

Circularity Check

0 steps flagged

No derivation-level circularity; central claims are empirical. Minor self-citation caveat from using the authors' own FIDpatch metric and reference set.

full rationale

The paper's load-bearing claims (SOTA UHR video quality, component effectiveness) are empirical comparisons, not consequences of definitions. The VLUer is trained on LR/HR latent pairs from UltraVideo and evaluated with reconstruction metrics (Table 12); the HCR experts are trained on real HR latents and evaluated in the full cascade (Tables 5-6). No fitted parameter is relabeled as a prediction: each reported number comes from a held-out evaluation protocol. The frequency-expert design cites external works (Yi et al., 2024; Zhang et al., 2024c) as motivation, but the implementation (LoRA plus low/high-pass filtering) is not an imported ansatz that predefines the result. No uniqueness theorem is invoked. The one notable self-citation is in the evaluation: FIDpatch and the UltraHR-eval4k reference set come from the authors' own prior work (Zhao et al., 2025d), and FIDpatch is used as a headline metric in Tables 2, 5, and 6. This is a caveat about in-house evaluation, but it is not circular because the same metric and reference protocol are applied uniformly to all baselines, and the conclusions are corroborated by external benchmarks (VBench, MUSIQ/MANIQA/NIQE/DOVER), an independent commercial MLLM (Doubao-1.5 Pro), and a 20-participant human study. The paper's stated limitation (computational efficiency, in the 'Limitations and Future Works' paragraph) is orthogonal to circularity. The VLUer-to-HCR distribution mismatch flagged by the skeptic is a generalization/robustness gap, not a circular step: HCR was trained on real latents and tested on VLUer outputs, so the test-time input distribution may differ, but the reported gains are not identical to the training targets by construction. Score 2 reflects the minor self-citation, not a circular derivation.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

All free parameters are hyperparameters chosen by the authors via heuristic or ablation; none are derived from theory. The axioms are common inductive biases in diffusion-based systems; they are plausible but not independently verified.

free parameters (5)
  • tswitch (noise-level switching timestep) = 0.417
    Splits training intervals of the low-frequency and high-frequency experts in Section 3.4; no sensitivity analysis is reported.
  • Skipped steps S in high-resolution refinement = 5
    Selected via ablation in Appendix A, Table 7; default for all main results.
  • HPS v3 data filtering threshold = 6.5
    Retained only UltraVideo samples above this human-preference score for expert training (Section 3.4, Figure 7); threshold set without sensitivity analysis.
  • Number of VLUer training pairs = ~20,000
    Result of filtering to native resolution >=1440x1440 (Appendix C.1); dataset size is not varied.
  • VLUer downsampling factors = 1.5x, 2.0x, 3.0x
    Used to create LR-HR latent pairs for arbitrary-scale upsampling (Appendix C.1); choice not justified beyond coverage of target scales.
axioms (4)
  • domain assumption The Wan2.1 VAE latent space is sufficiently continuous/structured that a lightweight INR can map low-resolution latents to the high-resolution latent manifold.
    Section 3.3 motivation: the whole VLU stage relies on learnable latent upsampling avoiding manifold deviation; no direct distributional analysis is provided.
  • domain assumption Low-frequency components of intermediate DiT features encode global semantic structure, while high-frequency components encode fine texture.
    Section 3.4 motivation and Figure 6; the design of LFE and HFE and their placement in attention/FFN depends on this frequency-semantics correspondence.
  • domain assumption A frozen Wan2.1-1.3B backbone can be adapted to UHR synthesis with LoRA-style experts without task-specific modifications to attention/temporal modules.
    Used throughout Sections 3.2-3.4; standard practice but not justified empirically beyond final performance.
  • domain assumption The filtered UltraVideo subset is representative and sufficient for training UHR generation at 2K/4K.
    Section 3.4 and Appendix C: data curation by resolution and HPS threshold; no analysis of domain coverage or scaling.

pith-pipeline@v1.3.0-alltime-deepseek · 21834 in / 13528 out tokens · 131861 ms · 2026-08-03T00:04:58.227510+00:00 · methodology

0 comments
read the original abstract

Recent advances in video diffusion models have significantly improved visual quality, yet ultra-high-resolution (UHR) video generation remains a formidable challenge due to the compounded difficulties of motion modeling, semantic planning, and detail synthesis. To address these limitations, we propose \textbf{LUVE}, a \textbf{L}atent-cascaded \textbf{U}HR \textbf{V}ideo generation framework built upon dual frequency \textbf{E}xperts. LUVE employs a three-stage architecture comprising low-resolution motion generation for motion-consistent latent synthesis, video latent upsampling that performs resolution upsampling directly in the latent space to mitigate memory and computational overhead, and high-resolution content refinement that integrates low-frequency and high-frequency experts to jointly enhance semantic coherence and fine-grained detail generation. Extensive experiments demonstrate that our LUVE achieves superior photorealism and content fidelity in UHR video generation, and comprehensive ablation studies further validate the effectiveness of each component. The project is available at \href{https://unicornanrocinu.github.io/LUVE_web/}{https://github.io/LUVE/}.

Figures

Figures reproduced from arXiv: 2602.11564 by Chen Zhao, Hongyu Li, Jian Yang, Jiawei Chen, Kai Zhang, Shilin Lu, Xiaoming Wei, Ying Tai, Zhuoliang Kang.

Figure 1
Figure 1. Figure 1: The base corresponds to the pretrained T2V model used in the first stage of our framework (Wan et al., 2025). As shown, compared with existing VSR methods, our model not only produces videos that are noticeably sharper and richer in fine details, but more importantly, it significantly enhances semantic consistency and plausibility. This demonstrates that UHR generation goes beyond merely enhancing visual s… view at source ↗
Figure 2
Figure 2. Figure 2: Scaling T2V models to UHR scenarios introduces several challenges. In motion modeling, models tend to produce static outputs, failing to capture coherent temporal dynamics. In semantic planning, both global and local repetitions emerge, reflecting insufficient semantic understanding. Finally, in detail synthesis, the generated frames often suffer from motion blur and texture degradation. tion poses a subst… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the LUVE framework. (a) and (b) illustrate the core distinction between existing cascaded high-resolution video generation architectures and our proposed paradigm. While previous methods focus on high-resolution detail refinement, our approach prioritizes high-resolution content and semantic fidelity. (c) Our LUVE, which consists of three collaborative stages: low-resolution motion generation (… view at source ↗
Figure 4
Figure 4. Figure 4: Framework Comparison. (a) Existing latent interpolation framework. (b) Existing RGB interpolation framework. (c) Our framework based on our video latent upsampler (VLUer)). Latent interpolation w/o Decoder w/o ℒ𝓅𝒾𝓍ℯℓ Full VLUer [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visual analysis of the key components in VLUer. These results demonstrate that our decoder effectively alleviates blurri￾ness, while our Lpixel successfully mitigates blocky artifacts. for subsequent UHR synthesis. For this stage, we adopt flow matching-based video foundation model Wan2.1 as the backbone (Wan et al., 2025), which consists of a 3D VAE for spatiotemporal compression (Esser et al., 2021; King… view at source ↗
Figure 6
Figure 6. Figure 6: (a) The architecture of the low-frequency expert. (b) The architecture of the high-frequency expert. (a) 1.29 1.43 8.12 9.10 (b) [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Data Selection and Augmentation. (a) First row: low￾HPS V3 scores; second row: high-HPS V3 scores. (b) First row: original data; second row: Unsharp Masking-enhanced data. reconstructed during the high-noise stage, whereas fine￾grained details (high-frequency components) are progres￾sively synthesized in the later low-noise stage (Yi et al., 2024; Zhang et al., 2024c). The insight motivates us to design tw… view at source ↗
Figure 8
Figure 8. Figure 8: Visual comparison with T2V models. These results demonstrate that our method not only preserves details in UHR scenarios but also maintains strong semantic consistency. Furthermore, it effectively captures complex motions in challenging scenes. Compared with recent SOTA methods, LUVE demonstrates a substantial improvement in generative capability [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visual comparison with VSR models. Recent VSR models are applied to enhance the outputs of the Base model, and the comparison demonstrates that our approach possesses a superior capability to recover fine-grained details and enhance overall realism. w/o LF Expert Ours w/o HF Expert Latent Interpolation RGB Interpolation [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visual analysis for ablation study. 4.3. Ablation Study To comprehensively evaluate the effectiveness of our method, we conduct extensive ablation studies. All experi￾ments are performed on 2K video generation. Ablation with Different Upsampling. We first compare different upsampling strategies, including RGB and latent interpolation. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visual analysis for different skipped steps [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 11 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. MetaPoint: Unlocking Precise Spatial Control in Agentic Visual Generation

    cs.CV 2026-06 unverdicted novelty 7.0

    MetaPoint represents 2D coordinates as special tokens in visual generative models to enable precise spatial control using existing positional encodings without architectural modifications.

  2. GS-STVSR: Ultra-Efficient Continuous Spatio-Temporal Video Super-Resolution via 2D Gaussian Splatting

    cs.CV 2026-04 unverdicted novelty 7.0

    GS-STVSR achieves state-of-the-art continuous spatio-temporal video super-resolution quality with nearly constant inference time at standard scales and over 3x speedup at extreme scales using 2D Gaussian Splatting.

  3. From Zero to Detail: A Progressive Spectral Decoupling Paradigm for UHD Image Restoration with New Benchmark

    cs.CV 2026-04 unverdicted novelty 7.0

    A new framework called ERR decomposes UHD image restoration into three frequency stages with specialized sub-networks and introduces the LSUHDIR benchmark dataset of over 82,000 images.

  4. RefineAnything: Multimodal Region-Specific Refinement for Perfect Local Details

    cs.CV 2026-04 unverdicted novelty 7.0

    RefineAnything is a multimodal diffusion model using Focus-and-Refine crop-and-resize with blended paste-back to achieve high-fidelity local image refinement and near-perfect background preservation.

  5. AtlasVid: Efficient Ultra-High-Resolution Long Video Generation via Decoupled Global-Local Modeling

    cs.CV 2026-05 unverdicted novelty 6.0

    AtlasVid proposes a decoupled global-local diffusion framework that trains at low resolution with LoRA and generalizes to ultra-high-resolution long video synthesis via semantic proxy guidance and locality-preserving ...

  6. Spiking Pyramid Wavelet Transformation for High-efficient and Low-energy Image Restoration

    cs.CV 2026-06 unverdicted novelty 5.0

    SPWM introduces spiking dual pyramid wavelet blocks to lower computational costs and energy use in image restoration while keeping quality comparable to prior methods.

  7. Temporal and Cross-Modal Alignment for Enhanced Audiovisual Video Captioning

    cs.CV 2026-07 unverdicted novelty 4.0

    TCA-Captioner introduces an Observer-Checker-Corrector refinement loop and TCA-Bench to address modality detachment and temporal incoherence in audiovisual video captioning.

  8. RankVR: Low-Rank Structure Perception and Value Recalibration for Robust Composed Image Retrieval

    cs.CV 2026-06 unverdicted novelty 4.0

    RankVR introduces GSCP and ASVC modules to improve CIR robustness by decoupling clean samples via low-rank structure and dynamically scoring triplet value in noisy datasets.

  9. IMAGINE: Adaptive Schema-Imagery Enhanced Composition for Composed Video Retrieval

    cs.CV 2026-06 unverdicted novelty 4.0

    IMAGINE uses adaptive schema-imagery via dynamic multimodal prototypes to incorporate implicit semantics into composed video retrieval, claiming SOTA results on CVR and CIR benchmarks.

  10. Why Do DiT Editors Drift? Plug-and-Play Low Frequency Alignment in VAE Latent Space

    cs.CV 2026-05 unverdicted novelty 4.0

    VAE-LFA suppresses semantic drift in multi-turn DiT image editing by low-pass filtering latent discrepancies and aligning low-frequency components to an EMA of previous rounds in VAE space.

  11. On the Controllability-Fidelity Frontier in Diffusion Editing

    cs.GR 2026-06 unverdicted novelty 3.0

    A study deriving mathematical formulations and bounds for diffusion editing objectives while empirically comparing methods on fidelity and control metrics and discussing ethical issues.

Reference graph

Works this paper leans on

42 extracted references · 24 linked inside Pith · cited by 11 Pith papers

  1. [1]

    Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127,

    Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y ., English, Z., V oleti, V ., Letts, A., et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127,

    Pith/arXiv arXiv

  2. [4]

    Diffusion model for camouflaged object detection.arXiv preprint arXiv:2308.00303, 2023b

    Chen, Z., Gao, R., Xiang, T.-Z., and Lin, F. Diffusion model for camouflaged object detection.arXiv preprint arXiv:2308.00303, 2023b. Chen, Z., Li, Y ., Wang, H., Chen, Z., Jiang, Z., Li, J., Wang, Q., Yang, J., and Tai, Y . Ragd: Regional-aware diffusion model for text-to-image generation. InProceedings of the IEEE/CVF International Conference on Compute...

    Pith/arXiv arXiv

  3. [5]

    score" (integer 1-10) and

    Detail-to-Motion Stability:Do these fine textures remain stable during movement? Details should not ’vibrate’, disappear, or become a blurry mess when the object or camera moves. II. Scoring Scale (1-10): • 10: Masterpiece— Breathtaking UHR richness. Tex- tures are indistinguishable from 4K/2K real-world footage. Exceptional density of meaningful informa-...

    2026

  4. [6]

    Eraseanything: Enabling concept erasure in rectified flow transformers

    Gao, D., Lu, S., Zhou, W., Chu, J., Zhang, J., Jia, M., Zhang, B., Fan, Z., and Zhang, W. Eraseanything: Enabling concept erasure in rectified flow transformers. InForty- second International Conference on Machine Learning, 2025a. Gao, Y ., Guo, H., Hoang, T., Huang, W., Jiang, L., Kong, F., Li, H., Li, J., Li, L., Li, X., et al. Seedance 1.0: Exploring t...

    Pith/arXiv arXiv

  5. [7]

    Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.arXiv preprint arXiv:2307.04725,

    Guo, Y ., Yang, C., Rao, A., Liang, Z., Wang, Y ., Qiao, Y ., Agrawala, M., Lin, D., and Dai, B. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.arXiv preprint arXiv:2307.04725,

    Pith/arXiv arXiv

  6. [10]

    Cogvideo: Large-scale pretraining for text-to-video gener- ation via transformers.arXiv preprint arXiv:2205.15868,

    Hong, W., Ding, M., Zheng, W., Liu, X., and Tang, J. Cogvideo: Large-scale pretraining for text-to-video gener- ation via transformers.arXiv preprint arXiv:2205.15868,

    Pith/arXiv arXiv

  7. [11]

    Kingma, D. P. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

    Pith/arXiv arXiv

  8. [12]

    As shown the Table, direct latent inter- polation performs poorly across all metrics, significantly underperforming our VLUer. This clearly indicates that the video latent space is highly non-linear, and that simple interpolation leads to severe information loss and structural distortion, thereby necessitating a learnable mapping such as VLUer. In contras...

    2026

  9. [13]

    Open-sora plan: Open-source large video generation model.arXiv preprint arXiv:2412.00131,

    Lin, B., Ge, Y ., Cheng, X., Li, Z., Zhu, B., Wang, S., He, X., Ye, Y ., Yuan, S., Chen, L., et al. Open-sora plan: Open-source large video generation model.arXiv preprint arXiv:2412.00131,

    Pith/arXiv arXiv

  10. [14]

    Hiprompt: Tuning- free higher-resolution generation with hierarchical mllm prompts.arXiv preprint arXiv:2409.02919,

    Liu, X., He, Y ., Guo, L., Li, X., Jin, B., Li, P., Li, Y ., Chan, C.-M., Chen, Q., Xue, W., et al. Hiprompt: Tuning- free higher-resolution generation with hierarchical mllm prompts.arXiv preprint arXiv:2409.02919,

    Pith/arXiv arXiv

  11. [15]

    Lu, S., Wang, Z., Li, L., Liu, Y ., and Kong, A. W.-K. Mace: Mass concept erasure in diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 6430–6440, 2024a. Lu, S., Zhou, Z., Lu, J., Zhu, Y ., and Kong, A. W.-K. Ro- bust watermarking using generative priors against image editing: From benchmarking to ad...

    Pith/arXiv arXiv

  12. [16]

    Towards realistic data generation for real- world super-resolution

    10 Submission 2026 Peng, L., Li, W., Pei, R., Ren, J., Xu, J., Wang, Y ., Cao, Y ., and Zha, Z.-J. Towards realistic data generation for real- world super-resolution. InThe Thirteenth International Conference on Learning Representations. Peng, L., Cao, Y ., Pei, R., Li, W., Guo, J., Fu, X., Wang, Y ., and Zha, Z.-J. Efficient real-world image super- resol...

    Pith/arXiv arXiv 2026

  13. [18]

    L., Cai, X., Huang, Q., Kang, Z., Li, H., Liang, S., Ma, L., Ren, S., Wei, X., Xie, R., et al

    Team, M. L., Cai, X., Huang, Q., Kang, Z., Li, H., Liang, S., Ma, L., Ren, S., Wei, X., Xie, R., et al. Longcat-video technical report.arXiv preprint arXiv:2510.22200,

    arXiv

  14. [19]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314,

    Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.-W., Chen, D., Yu, F., Zhao, H., Yang, J., et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314,

    Pith/arXiv arXiv

  15. [20]

    Ex- posing and defending the achilles’ heel of video mixture- of-experts.arXiv preprint arXiv:2602.01369,

    Wang, S., Liu, Q., Lyu, Y ., Li, N., He, Z., and Shan, C. Ex- posing and defending the achilles’ heel of video mixture- of-experts.arXiv preprint arXiv:2602.01369,

    arXiv

  16. [21]

    Robustgs: Unified boosting of feedforward 3d gaussian splatting under low-quality conditions.arXiv preprint arXiv:2508.03077,

    Wu, A., Peng, L., Di, X., Dai, X., Wu, C., Wang, Y ., Fu, X., Cao, Y ., and Zha, Z.-J. Robustgs: Unified boosting of feedforward 3d gaussian splatting under low-quality conditions.arXiv preprint arXiv:2508.03077,

    Pith/arXiv arXiv

  17. [22]

    Addsr: Accelerating diffusion-based blind super-resolution with adversarial diffusion distillation

    Xie, R., Zhao, C., Zhang, K., Zhang, Z., Zhou, J., Yang, J., and Tai, Y . Addsr: Accelerating diffusion-based blind super-resolution with adversarial diffusion distillation. arXiv preprint arXiv:2404.01717,

    Pith/arXiv arXiv

  18. [23]

    Star: Spatial-temporal augmentation with text-to-video mod- els for real-world video super-resolution.arXiv preprint arXiv:2501.02976,

    Xie, R., Liu, Y ., Zhou, P., Zhao, C., Zhou, J., Zhang, K., Zhang, Z., Yang, J., Yang, Z., and Tai, Y . Star: Spatial-temporal augmentation with text-to-video mod- els for real-world video super-resolution.arXiv preprint arXiv:2501.02976,

    Pith/arXiv arXiv

  19. [24]

    Camel: Energy-aware llm infer- ence on resource-constrained devices.arXiv preprint arXiv:2508.09173,

    Xu, H., Peng, L., Song, S., Liu, X., Jun, M., Li, S., Yu, J., and Mao, X. Camel: Energy-aware llm infer- ence on resource-constrained devices.arXiv preprint arXiv:2508.09173,

    Pith/arXiv arXiv

  20. [25]

    Ultravideo: High- quality uhd video dataset with comprehensive captions

    11 Submission 2026 Xue, Z., Zhang, J., Hu, T., He, H., Chen, Y ., Cai, Y ., Wang, Y ., Wang, C., Liu, Y ., Li, X., et al. Ultravideo: High- quality uhd video dataset with comprehensive captions. arXiv preprint arXiv:2506.13691,

    Pith/arXiv arXiv 2026

  21. [26]

    Cogvideox: Text-to-video diffusion models with an ex- pert transformer.arXiv preprint arXiv:2408.06072,

    Yang, Z., Teng, J., Zheng, W., Ding, M., Huang, S., Xu, J., Yang, Y ., Hong, W., Zhang, X., Feng, G., et al. Cogvideox: Text-to-video diffusion models with an ex- pert transformer.arXiv preprint arXiv:2408.06072,

    Pith/arXiv arXiv

  22. [27]

    Supergen: An effi- cient ultra-high-resolution video generation system with sketching and tiling.arXiv preprint arXiv:2508.17756,

    Ye, F., Zhao, Z., Mu, Y ., Shen, J., Li, R., Wang, K., Sun, D., Agarwal, S., Lee, M., Cao, T., et al. Supergen: An effi- cient ultra-high-resolution video generation system with sketching and tiling.arXiv preprint arXiv:2508.17756,

    arXiv

  23. [28]

    V o- capter: V oting-based pose tracking for category-level articulated object via inter-frame priors

    Zhang, L., Han, Z., Zhong, Y ., Yu, Q., Wu, X., et al. V o- capter: V oting-based pose tracking for category-level articulated object via inter-frame priors. InACM Multi- media 2024, 2024a. Zhang, L., Zhong, Y ., Wang, J., Min, Z., Liu, L., et al. Rethinking 3d convolution in ℓp-norm space. InThe Thirty-eighth Annual Conference on Neural Information Proce...

    Pith/arXiv arXiv 2024

  24. [29]

    Flashvideo: Flow- ing fidelity to detail for efficient high-resolution video generation.arXiv preprint arXiv:2502.05179, 2025d

    Zhang, S., Li, W., Chen, S., Ge, C., Sun, P., Zhang, Y ., Jiang, Y ., Yuan, Z., Peng, B., and Luo, P. Flashvideo: Flow- ing fidelity to detail for efficient high-resolution video generation.arXiv preprint arXiv:2502.05179, 2025d. Zhang, Y ., Yang, H., Zhang, Y ., Hu, Y ., Zhu, F., Lin, C., Mei, X., Jiang, Y ., Peng, B., and Yuan, Z. Waver: Wave your way t...

    arXiv 2024

  25. [31]

    Pyramid diffusion mod- els for low-light image enhancement.arXiv preprint arXiv:2305.10028,

    Zhou, D., Yang, Z., and Yang, Y . Pyramid diffusion mod- els for low-light image enhancement.arXiv preprint arXiv:2305.10028,

    Pith/arXiv arXiv

  26. [32]

    Migc++: Advanced multi-instance generation controller for image synthesis.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024a

    Zhou, D., Li, Y ., Ma, F., Yang, Z., and Yang, Y . Migc++: Advanced multi-instance generation controller for image synthesis.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024a. Zhou, D., Li, Y ., Ma, F., Zhang, X., and Yang, Y . Migc: Multi-instance generation controller for text-to-image synthesis. InProceedings of the IEEE/CVF confere...

    arXiv

  27. [33]

    Zhou, Z., Lu, S., Leng, S., Zhang, S., Lian, Z., Yu, X., and Kong, A. W.-K. Dragflow: Unleashing dit priors with region based supervision for drag editing.arXiv preprint arXiv:2510.02253, 2025c. Zhuang, J., Guo, S., Cai, X., Li, X., Liu, Y ., Yuan, C., and Xue, T. Flashvsr: Towards real-time diffusion- based streaming video super-resolution.arXiv preprint...

    arXiv

  28. [34]

    Effect of Different Skipped Steps We evaluate the effect of different skipped steps S during high-resolution content refinement

    13 Submission 2026 A. Effect of Different Skipped Steps We evaluate the effect of different skipped steps S during high-resolution content refinement. As shown in Table 7, using S= 5 achieves the best overall performance and is therefore adopted as our default setting. A too-small S re- stricts the extraction of reliable motion priors, whereas an excessiv...

    2026

  29. [35]

    After a preliminary filtering process, the videos are cropped and resized to a fixed size to ensure consistency for model training

    as the source for our training videos. After a preliminary filtering process, the videos are cropped and resized to a fixed size to ensure consistency for model training. To enable the model to handle upsampling across various scales, we apply multiple downscaling factors to the HR videos to generate the corresponding LR versions. These videos are then en...

    2025

  30. [36]

    as the source for training the VLUer. The data curation pipeline is as follows: Filtering.We filter the dataset to retain only videos with a native resolution of at least 1440 ×1440, resulting in a subset of approximately 20,000 high-quality videos. Preprocessing.To construct the High-Resolution (HR) latents, we resize the short edge of the videos to 1440...

    2026

  31. [37]

    Gradient checkpointing is enabled for both attention and feed-forward modules to reduce memory consumption

    All experiments are imple- mented in PyTorch. Gradient checkpointing is enabled for both attention and feed-forward modules to reduce memory consumption. It is worth noting that, due to GPU mem- ory constraints, we do not compute Lpixel on full frames; instead, we calculate the loss using cropped patches. 15 Submission 2026 Table 12.Quantitative compariso...

    2026

  32. [39]

    andHiD- iffusion(Zhang et al., 2023), typically employ dilated con- volutions or shifted window mechanisms to mitigate the ”object repetition” issue caused by the limited receptive fields of pre-trained UNets. While achieving remarkable computational efficiency and preserving global structural consistency, they often produce over-smoothed textures and lac...

    2023

  33. [40]

    However, the enormous GPU mem- ory requirements and the scarcity of diverse 4K/8K train- ing data remain significant bottlenecks for these methods

    explore scaling laws in DiT architectures, demonstrating that high-resolution visual quality can be significantly improved by fine-tuning on mas- sive curated datasets. However, the enormous GPU mem- ory requirements and the scarcity of diverse 4K/8K train- ing data remain significant bottlenecks for these methods. Super-resolution (SR) and Video Super-re...

    2025

  34. [41]

    Semantic Fidelity & Complexity:Does the UHR content (e.g., skin pores, fabric textures, distant 17 Submission 2026 scenery) look genuine or like ’abstract noise’? High resolution should mean more meaningful content, not just sharper artifacts. •

    2026

  35. [2013]

    Hunyuan- video: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603,

    Kong, W., Tian, Q., Zhang, Z., Min, R., Dai, Z., Zhou, J., Xiong, J., Li, X., Wu, B., Zhang, J., et al. Hunyuan- video: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603,

    Pith/arXiv arXiv

  36. [2020]

    Ultrapixel: Advancing ultra- high-resolution image synthesis to new peaks.arXiv preprint arXiv:2407.02158,

    Ren, J., Li, W., Chen, H., Pei, R., Shao, B., Guo, Y ., Peng, L., Song, F., and Zhu, L. Ultrapixel: Advancing ultra- high-resolution image synthesis to new peaks.arXiv preprint arXiv:2407.02158,

    Pith/arXiv arXiv

  37. [2021]

    Videoinr: Learning video implicit neural representation for continuous space-time super- resolution

    Chen, Z., Chen, Y ., Liu, J., Xu, X., Goel, V ., Wang, Z., Shi, H., and Wang, X. Videoinr: Learning video implicit neural representation for continuous space-time super- resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2047– 2057,

    2047

  38. [2022]

    Videocrafter1: Open diffusion models for high-quality video generation

    Chen, H., Xia, M., He, Y ., Zhang, Y ., Cun, X., Yang, S., Xing, J., Liu, Y ., Chen, Q., Wang, X., et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023a. Chen, J., Ge, C., Xie, E., Wu, Y ., Yao, L., Ren, X., Wang, Z., Luo, P., Lu, H., and Li, Z. Pixart-σ: Weak-to-strong training of diffusion ...

    Pith/arXiv arXiv

  39. [2023]

    Ltx-video: Realtime video latent diffusion.arXiv preprint arXiv:2501.00103,

    9 Submission 2026 HaCohen, Y ., Chiprut, N., Brazowski, B., Shalem, D., Moshe, D., Richardson, E., Levin, E., Shiran, G., Zabari, N., Gordon, O., et al. Ltx-video: Realtime video latent diffusion.arXiv preprint arXiv:2501.00103,

    Pith/arXiv arXiv 2026

  40. [2024]

    Venhancer: Generative space- time enhancement for video generation.arXiv preprint arXiv:2407.07667,

    He, J., Xue, T., Liu, D., Lin, X., Gao, P., Lin, D., Qiao, Y ., Ouyang, W., and Liu, Z. Venhancer: Generative space- time enhancement for video generation.arXiv preprint arXiv:2407.07667,

    Pith/arXiv arXiv

  41. [2025]

    Textcrafter: Accurately rendering multiple texts in complex visual scenes.arXiv preprint arXiv:2503.23461,

    Du, N., Chen, Z., Gao, S., Chen, Z., Chen, X., Jiang, Z., Yang, J., and Tai, Y . Textcrafter: Accurately rendering multiple texts in complex visual scenes.arXiv preprint arXiv:2503.23461,

    arXiv

  42. [2026]

    Zhao, X., Zhao, C., Hu, X., Zhang, H., Tai, Y ., and Yang, J

    doi: 10.1109/ TGRS.2026.3660483. Zhao, X., Zhao, C., Hu, X., Zhang, H., Tai, Y ., and Yang, J. Learning multi-scale spatial-frequency features for image denoising.arXiv preprint arXiv:2506.16307, 2025e. Zheng, Z., Peng, X., Yang, T., Shen, C., Li, S., Liu, H., Zhou, Y ., Li, T., and You, Y . Open-sora: Democratiz- ing efficient video production for all.ar...

    arXiv 2026