{"id":"30a6267e-0fca-4b7b-a516-3ae8c3520f84","arxiv_id":"2508.15774","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"CineScale extends pre-trained diffusion models to 8k image and 4k video generation with mostly tuning-free inference plus a small LoRA adaptation for video.","lead":"This paper describes an inference-time recipe, CineScale, that lets existing image and video diffusion models generate far higher resolutions than they were trained for: 8k images and 4k video. It combines cascade upscaling, frequency-based attention fusion, and small LoRA fine-tuning, aimed at content creation and film production.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 7's frequency split is unvalidated: if repetition artifacts include low-frequency content, Scale Fusion cannot remove them, undermining the core mechanism behind the 8k/4k claims.","rationale":"The paper's headline contributions—8k image and 4k video generation—depend on the method actually suppressing repetition artifacts. Scale Fusion is the component specifically designed for this; the self-cascade and dilated convolution handle structure and receptive field, but the paper claims Scale Fusion 'eliminates all kinds of unexpected object repetitions.' That claim rests entirely on the frequency-separation hypothesis of Eq. 7. I considered alternative concerns: (1) the unusual FID evaluation (comparing low-res references to high-res outputs) is a methodological weakness, but it affects only the quantitative ranking, not the qualitative capability; (2) the lack of code affects reproducibility but not the validity of the idea; (3) the NTK-RoPE transfer is a secondary adaptation. The frequency split is the core novel mechanism and is the least supported. The reader's weakest_assumption identifies the same issue; I agree. The existing ablations (Table 4) show that removing Scale Fusion hurts, but that only establishes correlation, not that the Gaussian frequency decomposition is the reason. A direct spectral test could falsify the assumption. Therefore, I do not change the reader's CONDITIONAL verdict; the paper should either provide such validation or temper the strength of the claims.","tokens_in":20221,"tokens_out":6910,"duration_ms":75829,"concrete_test":"On a repeated-artifact case (e.g., SDXL at 4096² without CineScale), extract h_global and h_local at an intermediate attention layer. Apply the same Gaussian blur G used in Eq. 7 and compute the frequency content of the residual. Specifically, segment the output image into regions with visible repeated objects vs. normal regions, back-project these masks to the feature map, and compare the energy of G(h_local) (low-pass) in repeated vs. normal regions. If the low-pass component contains as much energy in repeated regions as in normal regions, the assumption that repetition is high-frequency is false. Additionally, run the full pipeline with Eq. 7 replaced by (a) a simple average (h_global+h_local)/2 and (b) a hard FFT cutoff split; if (a) performs comparably or (b) collapses, the Gaussian frequency interpretation is not the causal factor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism of CineScale is Eq. 7: h_fusion = (h_global - G(h_global)) + G(h_local). This treats the Gaussian-filtered local attention output G(h_local) as pure 'semantic structure' and the residual (h_global - G(h_global)) as clean high-frequency detail. The paper's premise (§3.4 and Abstract) is that repetition artifacts are 'dispersed high-frequency signals' that global attention gathers correctly. But this is asserted, not demonstrated. Repeated objects (e.g., extra faces/eyes) contain substantial low-frequency content (skin tone, shading, coarse shape), so G(h_local) can propagate them. Moreover, G is a spatial blur on hidden features in latent space, not a spectral decomposition of the decoded image; the VAE latent's frequency organization need not align with visual repetition. There is no spectral analysis, no visualization of G(h_local) vs. the high-pass residual, and no ablation that isolates the frequency-split hypothesis (the Table 4 ablation only removes Scale Fusion as a whole). If the decomposition is wrong, the method's ability to 'eliminate all types of repetition' is not grounded, and the 8k/4k headline claims rest on an unstable assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CineScale, an inference-time framework for generating images and videos at resolutions far above a model's training resolution. For UNet-based models it builds on the authors' prior FreeScale, combining self-cascade upscaling, restrained dilated convolution, and a Scale Fusion mechanism (Eq. 7) that mixes global and local attention outputs through a Gaussian-blur frequency split. For DiT-based models it adds NTK-RoPE, attention temperature scaling, noise shifting, and minimal LoRA fine-tuning. The method is evaluated on SDXL and VideoCrafter2 for text-to-image/text-to-video, and on Wan and ReCamMaster for text-to-video, image-to-video, and video-to-video, with claims of 8k image generation without fine-tuning and 4k video generation with minimal LoRA. Quantitative comparisons use FID/KID/FIDc/KIDc/IS, VBench metrics, user studies, and ablations.","tokens_in":20629,"tokens_out":7121,"duration_ms":81399,"significance":"If the central claims hold, the paper has clear practical significance: it offers a largely tuning-free recipe for extending open diffusion models to resolutions two orders of magnitude above training, and it expands high-resolution generation to I2V and V2V tasks that prior baselines do not cover. The paper's strengths are its breadth (multiple architectures, multiple tasks, several baselines, user studies, and ablations), its transparency about reusing the prior FreeScale method, and the reported time-cost comparisons. However, the quantitative support is weakened by the choice of reference distributions, the absence of quantitative evaluation at the headline 8k/4k resolutions, and the lack of direct validation of the frequency-separation hypothesis behind Eq. 7. The empirical claims are plausible but not yet fully established.","major_comments":[{"comment":"The core mechanism for eliminating repetition is the frequency split h_fusion = (h_global - G(h_global)) + G(h_local), where G is a Gaussian blur. The paper asserts that repetition artifacts are 'dispersed high-frequency signals' and that the low-pass-filtered local attention output contains the semantic structure. This is not demonstrated. Repeated objects (e.g., extra eyes or faces) contain substantial low-frequency content such as skin tone and coarse shape, so the split is not guaranteed to isolate repetition. The Gaussian blur operates on latent-space deep features, not on the decoded image spectrum, so the frequency interpretation in Eq. (7) needs direct evidence. Table 4 ablates Scale Fusion only as a whole and does not isolate the frequency-split hypothesis. Please add a spectral analysis, visualizations of G(h_local) and of the high-pass residual, or a controlled variant that sw","section":"§3.4, Eq. (7)"},{"comment":"The quantitative protocol is non-standard: FID/KID/FVD are computed between the originally generated low-resolution outputs and the corresponding high-resolution outputs, not against a real high-resolution reference distribution. This measures the distributional similarity of high-res outputs to their own low-res counterparts, which rewards methods that preserve the low-res content and penalizes legitimate novel high-frequency detail. Consequently, the scores in Tables 1, 2, and 5 do not support an absolute claim of high-resolution quality. Please report metrics against a real high-resolution image/video dataset or otherwise clearly reframe the tables as measuring consistency with the low-resolution output, and justify why this is the appropriate criterion for 'superiority'.","section":"§5 (Evaluation Metrics), Tables 1, 2, 5"},{"comment":"The headline claim is '8k image generation without any fine-tuning,' yet all image-quantitative tables (Tables 1-4) evaluate at 2048x2048 and 4096x4096. The 8k result appears only as a single qualitative example (Fig. 4). Since 8k is a load-bearing claim, the paper should provide quantitative support at 8k, even if limited (e.g., a small set of prompts with FID/FIDc, or a dedicated user study), or at minimum explicitly state that the 8k claim is qualitative-only and discuss variability across prompts.","section":"Abstract, §5, Tables 1-4"},{"comment":"In Table 1, SDXL-DI achieves the best FIDc and KIDc at both 2048^2 and 4096^2. The text explains this by saying SDXL-DI tends to generate duplicated objects, so its crops may be closer to the reference images. That is a post hoc reinterpretation and undermines the validity of FIDc as a quality metric in this setup: a method that visibly produces repetitive artifacts is rewarded. If FIDc and KIDc are retained, the paper must either justify why these metrics are meaningful when the worst-behaving method wins, or replace/annotate them with a metric that better agrees with the qualitative and user-study outcomes.","section":"Table 1 (SDXL-DI row)"},{"comment":"The 4k text-to-video and image-to-video results are presented only qualitatively (Figs. 18, 19). The quantitative VBench tables for DiT models are at 960x1664 (Tables 9, 10, 12) or 1088x1920; the 1920x3328 table (Table 11) reports only three metrics. The LoRA fine-tuning recipe (about 20,000 videos, rank/alpha, steps) is not ablated, and no comparison is made against full fine-tuning or against a high-resolution-trained model of comparable size. Please add quantitative 4k evaluation (e.g., a subsampled VBench run or a user study at 4k) and ablate the LoRA configuration, or explicitly state that the 4k claim is qualitative-only.","section":"§6.1, Fig. 18, Table 10"}],"minor_comments":[{"comment":"There are several presentation issues: 'destroy their signa [64]' should read 'signal'; the symbol 'λβ' is introduced but never defined; and Eq. (9) is malformed in the text ('softmax( qTmkn t p |D| )'). These need correction.","section":"§4.2"},{"comment":"The captions contain sentence fragments: 'Although all variants can generate rough results. Our full method performs the best.' Please rewrite as complete sentences.","section":"Figures 12, 15, 16 captions"},{"comment":"The user study reports 23 participants for images and an unspecified number for video, with no confidence intervals or statistical significance tests. Given the large claimed margins, adding per-aspect vote counts and a basic significance test would strengthen the conclusions.","section":"§5.4, Tables 6, 7"},{"comment":"Several hyperparameters are not specified: the attention temperature t in Eq. (9), the noise shift schedule, the LoRA rank/alpha and number of training steps, and the exact data-filtering procedure for the 20,000 Pexels videos. Please include these for reproducibility.","section":"§4.2, §4.3"},{"comment":"The statement that 'only three metrics in VBench can be measured due to memory limitations' is vague. Specify which metrics, why they could not be computed, and the subsampling procedure used for the reported metrics.","section":"Table 11"}],"recommendation":"major_revision","confidential_remarks":"The paper is transparent about building on the authors' prior FreeScale, but the incremental contribution—the DiT extension and LoRA-based 4k video—should be weighed carefully in the editorial decision. The 'free lunch' framing overstates the cost: 4k video requires LoRA fine-tuning on ~20k videos, and the 8k image claim is not quantitatively evaluated. The quantitative evaluation protocol is a substantive issue that would need to be addressed for a journal-level acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is essentially a second part to the authors' FreeScale paper, now packaged as CineScale. The genuinely new material is the DiT-side adaptation — NTK-RoPE, attention temperature, noise shifting, and the LoRA fine-tune on 2K videos — plus the I2V and V2V evaluations. If you want to push Wan or SDXL beyond training resolution without retraining, the recipe seems to work: the 4k video stills are striking and the ablated variants behave as expected. I give the authors credit for being upfront that the UNet half is FreeScale, not a new contribution.\n\nThe weak spot is the paper's signature: Eq. 7 splits attention outputs into high- and low-frequency via Gaussian blur in latent space, then replaces the local high frequencies with global ones. This is asserted, not demonstrated. Repeated objects (extra faces, second noses) carry plenty of low-frequency structure — skin tone, shading, coarse shape — so a spatial blur on hidden features isn't obviously separating the artifact from the semantics. The ablation in Table 4 removes Scale Fusion as a unit, giving no evidence that the frequency-split hypothesis is the active ingredient. That doesn't mean the method fails; it means the stated mechanism is under-supported.\n\nThe quantitative evaluation has real caveats. FID/KID are computed against low-res references, and the paper itself explains SDXL-DI's winning FIDc as a duplicate-object artifact — a post hoc reinterpretation. There are no error bars, and the user study is only 23 raters. The paper also ships no code, and the method has a long list of free parameters (alpha, K, patch sizes, blur sigma, temperature, NTK base scale, LoRA hyper-parameters) that are lightly specified.\n\nNone of this is fatal. The experiments are broader than the prior work and the qualitative comparisons are convincing. But the headline claims — 'free lunch,' 8k without tuning, 4k with minimal LoRA — overstate the evidentiary support. I'd send this to peer review, not desk-reject it: the prompt engineer's guide to high-res generation is useful to the community and the manuscript is coherent. A revision should release code/checkpoints, add variance/repetition stats, and run a targeted test of the frequency-split assumption (e.g., swap the Gaussian for different low-pass filters, or visualize the decomposed latents).","headline":"A useful engineering recipe for high-res image/video generation, but the frequency-split mechanism is asserted, not validated, and the quantitative claims need more support.","tokens_in":21089,"tokens_out":3281,"would_cite":true,"duration_ms":34407,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CineScale claims pretrained diffusion models can generate 8k images with no fine-tuning and 4k video with minimal LoRA, by fusing global and local attention through a frequency split.","keywords":["diffusion models","high-resolution generation","tuning-free inference","scale fusion","frequency separation","DiT","LoRA fine-tuning","text-to-video generation"],"falsifier":"Generate a 4096×4096 image of a periodic texture (e.g., a brick wall or herringbone fabric) with a model trained at 1024×1024 under CineScale, and inspect the spectrum of the intermediate self-attention outputs at the highest scale. If the Gaussian blurred low-frequency component already contains the duplicated-object artifacts, or if removing the high-frequency residual fails to eliminate repetition, then Eq. (7)'s premise is violated. A simpler version: compare CineScale to a variant that replaces the Gaussian blur with an ideal low-pass filter at the same cut-off; large quality differences","tokens_in":20161,"feed_emoji":"🎬","tokens_out":8594,"duration_ms":82469,"temperature":0.7,"pith_summary":"CineScale claims that pre-trained diffusion models can generate far beyond their training resolution without retraining, if the inference recipe compensates for the extra high-frequency information that appears when the model sees more pixels. The paper proposes an inference-time pipeline—self-cascade upscaling, restrained dilated convolution, and a Scale Fusion step that combines global and local attention through a frequency split—and reports that it removes the repetitive-object artifacts that plague naive upscaling. On image models it claims 8192×8192 generation with no fine-tuning, and on transformer-based video models it claims 4K (2176×3840) generation after only a small LoRA adaptation. The paper also claims the recipe extends the task scope beyond text-to-image and text-to-video to image-to-video and video-to-video, where prior tuning-free methods did not operate.","feed_headline":"Tuning-free diffusion reaches 8k images; small LoRA reaches 4k video","feed_subtitle":"A frequency split in attention kills repetition, letting pretrained diffusion models exceed training resolution.","key_machinery":"The load-bearing object is Scale Fusion, the identity h_fusion = (h_global − G(h_global)) + G(h_local), where G is a Gaussian low-pass filter applied inside the self-attention layer. It fuses a global self-attention pass, which gathers object semantics across the whole latent, with shifted local self-attention patches, which preserve local detail, by keeping the global high frequencies and the local low frequencies. The paper also uses a tailored self-cascade upscaling schedule that re-noises and denoises progressively, a restrained dilated convolution that only widens receptive fields in down/mid blocks and early timesteps, and—for the DiT variant—NTK-RoPE positional encoding, attention tem","core_discovery":"At the center of the paper is a claim about why high-resolution generation fails and how to fix it without training. The paper argues that when a diffusion model generates at a resolution beyond its training data, the latent representations carry an excess of high-frequency information, and that this excess accumulates into repetitive patterns. The proposed remedy, Scale Fusion, edits self-attention outputs with the identity h_fusion = high_freq(h_global) + low_freq(h_local): it takes the global attention output, removes its low-frequency content with a Gaussian blur to keep the high-frequency signal, and re-adds the Gaussian-blurred local attention output so the low-frequency semantic struc","pith_inferences":["An implication the paper leaves implicit: the frequency-separation hypothesis predicts that Scale Fusion should be most effective when repetition artifacts are spectrally localized; a user could test this by generating a high-resolution image of a scene with a fine, periodic texture and checking whether the Gaussian-blur cut-off frequency needs to scale with the upsampling factor.","The success of NTK-RoPE and attention scaling in this setting suggests that long-context adaptation techniques from language models may transfer to any attention-based generative model with positional encodings, so the same recipe could apply to future non-DiT transformer generators, not merely the Wan backbone used here.","Because the cascade regenerates ambiguous low-resolution regions from the model's learned prior, CineScale is a generative upsampler rather than a faithful restorer; for applications that require pixel-level fidelity to an input frame, the paper's own comparisons with super-resolution methods indicate a user would still need an SR model."],"forward_implications":["Pre-trained image diffusion models trained at 1024×1024 can generate 8192×8192 outputs with no training, at quality the paper reports as better than prior tuning-free baselines on FID, KID, and IS.","Pre-trained video models can be pushed to 4K (2176×3840) with only a minimal LoRA adaptation, and the same pipeline extends to image-to-video and video-to-video, not just text-conditioned generation.","The method works across architectural families: UNet-based models (SDXL, VideoCrafter2) and DiT-based models (Wan) need different fixes, and the paper specifies which component addresses which failure mode—repetition for UNet, blur for DiT.","Users can control detail level spatially via a mask and can perform local semantic editing at high resolution by editing prompts during the cascade, which the paper demonstrates for both images and videos.","The recipe is compatible with distilled fast models such as SDXL-Turbo, so it is not restricted to slow full-step samplers."],"supporting_citations":[{"why":"Establishes the dilated-convolution remedy for repetitive high-resolution output and identifies the receptive-field limit that CineScale builds on and restrains.","marker":"[8]"},{"why":"Supplies the local-patch fusion mechanism for coherent high-resolution sampling that Scale Fusion adapts.","marker":"[9]"},{"why":"Provides the global-plus-local patch fusion framework and the self-cascade idea that CineScale's Scale Fusion extends with explicit frequency filtering.","marker":"[10]"},{"why":"Gives the frequency-domain explanation of repetition artifacts (excess high-frequency signal) that motivates the central Eq. (7) separation.","marker":"[11]"},{"why":"Is the UNet-based text-to-video backbone used to evaluate the FreeScale video variant.","marker":"[4]"},{"why":"Provides NTK-aware RoPE and attention reweighting, borrowed by the DiT extension to handle token counts and positional encoding at high resolution.","marker":"[62]"},{"why":"Is the open DiT-based video model used as the backbone for the DiT experiments and 4K video generation.","marker":"[38]"},{"why":"Supplies the parameter-efficient low-rank adaptation used for the minimal fine-tuning step to 4K video.","marker":"[15]"},{"why":"Is the earlier FreeScale paper whose self-cascade, dilated convolution, and Scale Fusion components CineScale extends to DiT and to I2V/V2V tasks.","marker":"[16]"}],"fun_headline_variants":["No training, just a frequency split: diffusion hits 8k images, 4k video","8k images and 4k video from pretrained diffusion — just a trick in attention","Diffusion goes 8K/4K without retraining—just split high and low frequencies","No fine-tuning: CineScale lifts diffusion to 8K images and 4K video","Frequency-split attention unlocks beyond-training resolutions in diffusion"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that a Gaussian low-pass filter on attention outputs in latent space separates repetition-causing high-frequency signals from semantic structure; if repetition artifacts leak into low-frequency channels—or the blur destroys object-level structure—Scale Fusion loses its justification, and the paper reports no quantitative test of that frequency separation.","fun_headline_variants_meta":{"raw":{"variants":["No training, just a frequency split: diffusion hits 8k images, 4k video","8k images and 4k video from pretrained diffusion — just a trick in attention","Diffusion goes 8K/4K without retraining—just split high and low frequencies","No fine-tuning: CineScale lifts diffusion to 8K images and 4K video","Frequency-split attention unlocks beyond-training resolutions in diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00075,"raw_usage":{"total_tokens":3195,"prompt_tokens":783,"completion_tokens":2412,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":2301}},"tokens_in":527,"tokens_out":2412,"duration_ms":20099,"temperature":1.0,"reasoning_tokens":2301,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:40:59.057516+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a 4096×4096 image of a periodic texture (e.g., a brick wall or herringbone fabric) with a model trained at 1024×1024 under CineScale, and inspect the spectrum of the intermediate self-attention outputs at the highest scale. If the Gaussian blurred low-frequency component already contains the duplicated-object artifacts, or if removing the high-frequency residual fails to eliminate repetition, then Eq. (7)'s premise is violated. A simpler version: compare CineScale to a variant that replaces the Gaussian blur with an ideal low-pass filter at the same cut-off; large quality differences","supporting_citations":[{"cited_title":"Scalecrafter: Tuning-free higher- resolution visual generation with diffusion models,","cited_arxiv_id":null,"evidence_quote":"Establishes the dilated-convolution remedy for repetitive high-resolution output and identifies the receptive-field limit that CineScale builds on and restrains."},{"cited_title":"Demo- fusion: Democratising high-resolution image generation with no $$$,","cited_arxiv_id":null,"evidence_quote":"Provides the global-plus-local patch fusion framework and the self-cascade idea that CineScale's Scale Fusion extends with explicit frequency filtering."},{"cited_title":"Videocrafter2: Overcoming data limitations for high- quality video diffusion models,","cited_arxiv_id":null,"evidence_quote":"Is the UNet-based text-to-video backbone used to evaluate the FreeScale video variant."}],"review_version":1}