{"id":"74c6a162-e81c-458b-ad78-f6725f6ce202","arxiv_id":"2412.09626","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A tuning-free scale-fusion method that lets frozen diffusion models generate 8k images and high-res videos by combining global and local attention through frequency filtering.","lead":"FreeScale is an inference-only method that lets pre-trained diffusion models generate images up to 8192x8192 pixels and higher-resolution videos without any fine-tuning. It fuses global and local attention signals in the frequency domain to prevent the repeated-pattern artifacts that normally appear when such models go beyond their training resolution.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 8k-generation claim rests on a few qualitative examples; all quantitative evaluations stop at 4096, so the headline result lacks distribution-level support.","rationale":"The reader's weakest assumption was the frequency decomposition in Eq. 7, which is a legitimate concern about the mechanism. However, the most load-bearing issue for the central claim is that the headline '8k' capability is asserted without quantitative evaluation at that resolution. If FreeScale fails on a broad set of prompts at 8192, the paper's main contribution is not established, regardless of whether the frequency split is correct. The reader's rationale did list the missing 8k evaluation as the first shortcoming, so there is partial agreement. My concern reinforces the CONDITIONAL verdict: acceptance should require evidence that the method generalizes at 8k, not just on a few hand-picked examples. The concrete test directly addresses this gap by measuring distribution-level quality and user preference at the claimed resolution.","tokens_in":15410,"tokens_out":7881,"duration_ms":79117,"concrete_test":"Run FreeScale on 50–100 randomly sampled LAION-Aesthetics prompts at 8192×8192 with the authors' hyperparameters (K=700, α=1, dilation d=8) and compute FIDc/KIDc against 1024×1024 SDXL references using the protocol of Table 1; additionally run a forced-choice user study (≥20 users) comparing 8k outputs against SDXL+Real-ESRGAN. If FIDc degrades by more than ~20% versus the 4096 result or the user study does not show majority preference for FreeScale, the 'unlock 8k' claim would be unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim—'unlocks the 8k-resolution text-to-image generation for the first time'—is not supported by distribution-level evidence. Quantitative experiments in Table 1 are limited to 2048² and 4096²; the 8k results are only a small qualitative gallery (Section C.3, Figures 10 and 12). No FID/KID, user study, or error analysis is reported at 8192², and the method's key free parameters (α in Eq. 4, blur kernel size) are adjusted differently for 8k (α=1 vs α=2), so the 8k setting is a distinct regime that has not been validated. If the method produces structural incoherence or repetition on a nontrivial fraction of prompts at 8192², the central claim fails. The paper also does not compare against recently published tuning-free methods that reach similar resolutions, making the 'first' claim hard to assess.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"FreeScale proposes a tuning-free inference framework that extends pre-trained U-Net diffusion models (SDXL for images, VideoCrafter2 for video) to resolutions far above their training resolution, up to 8192x8192 images and 640x1024 videos. The method combines three components: tailored self-cascade upscaling with a cosine detail-control factor (Eq. 4), restrained dilated convolution applied only to down-blocks and mid-blocks in early timesteps, and scale fusion, which replaces the high-frequency component of local self-attention with the high-frequency component of global self-attention via a Gaussian blur (Eq. 7). Experiments compare against SDXL-DI, ScaleCrafter, DemoFusion, and FouriScale on image metrics (FID/KID/FIDc/KIDc/IS) and video metrics (FVD, VBench), with ablations and user studies. The paper claims to 'unlock 8k-resolution text-to-image generation for the first time.'","tokens_in":15564,"tokens_out":6755,"duration_ms":66755,"significance":"If the central claims hold, FreeScale would be a practically useful, training-free recipe for pushing U-Net diffusion models far beyond their native resolution, with a simple mechanism that also transfers to video and supports flexible aspect ratios and local control. The method is clearly described, the components are individually ablated, and the reported inference times are competitive. However, the headline 8k claim is not supported by quantitative evaluation, and the method's own quantitative comparisons show mixed results on the local-crop metrics that are most relevant to the repetition artifacts the paper targets. The frequency-separable artifact assumption behind scale fusion is untested. As presented, the contribution is a plausible and well-engineered empirical method, but its strongest claims require additional validation before they can be accepted at face value.","major_comments":[{"comment":"The paper's headline claim is that FreeScale 'unlocks the 8k-resolution text-to-image generation for the first time' (Abstract and Introduction), yet all quantitative evaluations stop at 4096x4096 (Table 1), and the 8k results are confined to qualitative figures (Figures 10 and 12). Moreover, Section A states that for 8k generation alpha is reduced from 2 to 1, so the 8k results correspond to a different hyperparameter regime that has not been quantitatively validated. The authors should either provide distribution-level evidence at 8192x8192 (e.g., a user study on a larger prompt set, or meaningful quantitative metrics) or qualify the 8k claim as a qualitative demonstration.","section":"Section 4.1 and Section C.3"},{"comment":"On the cropped metrics FIDc and KIDc, which are intended to capture local quality, SDXL-DI (direct inference) outperforms FreeScale at both 2048^2 (FIDc 31.042 vs 36.276; KIDc 0.004 vs 0.006) and 4096^2 (FIDc 42.383 vs 71.369; KIDc 0.009 vs 0.029). The paper explains that duplicated objects in SDXL-DI make its crops closer to the reference distribution, but it does not substantiate this explanation. Since the central motivation is to eliminate repetitions, the method should not be worse on local-crop metrics than the baseline that exhibits repetitions; the authors should analyze this discrepancy (e.g., by comparing crops side-by-side or decomposing the metric contribution) or reconsider the interpretation of these metrics.","section":"Section 4.1, Table 1"},{"comment":"Scale fusion assumes that repeated-object artifacts are carried by the high-frequency component of the local attention output, and that the global attention output provides correct high-frequency structure. This is the core design hypothesis, but the paper provides no evidence that the artifact mechanism is separable in frequency. The ablation in Table 3 removes the entire scale fusion module, which shows a large drop in quality, but it does not isolate the frequency split. A direct test would be to ablate the fusion formula, e.g., swapping the high- and low-frequency contributions, or using a simple weighted sum of global and local outputs, to verify that the frequency-based replacement, rather than merely having two receptive fields, is what drives the improvement.","section":"Section 3.4, Eq. (7)"},{"comment":"The paper describes FreeScale as 'tuning-free,' but it requires setting several hyperparameters: noise timestep K, detail-control exponent alpha, dilation factor d, Gaussian blur kernel size and standard deviation, and local attention window size and stride. The default values are given, but no sensitivity analysis is provided, and alpha is changed for 8k generation (from 2 to 1). For a method that claims to be a turnkey solution, the robustness of these choices should be demonstrated; at minimum, a sweep over alpha and K, or an explicit discussion of how to select them, is needed. Without such analysis, the 'tuning-free' characterization is misleading and the 8k results rely on unvalidated parameter adjustments.","section":"Section 3.2 and Section A"},{"comment":"The claim of being 'the first' 8k text-to-image generation method is not supported by a comparison with the most recent tuning-free high-resolution generation methods. The related work cites HiDiffusion, ElasticDiffusion, AP-LDM, CutDiffusion, and others, but these are not included in the quantitative or qualitative comparisons. To substantiate the 'first' claim, the authors should compare against these methods at 8192x8192 (or at least at 4096x4096) and demonstrate that they fail or produce inferior results, or they should qualify the novelty claim to reflect the limited comparison set.","section":"Section 2 and Section 4.1"}],"minor_comments":[{"comment":"All quantitative metrics are reported as single values without standard deviations or confidence intervals. Given that the differences from some baselines are modest (e.g., FIDc at 2048^2), the paper could benefit from multiple runs or at least a bootstrap-based confidence interval to establish that the findings are not noise.","section":"Section 4.1, Table 1"},{"comment":"The video evaluation is performed only on VideoCrafter2; the claim that FreeScale generalizes to video would be strengthened by a second video backbone, even if the results are reported in the supplementary material.","section":"Section 4.2, Table 2"},{"comment":"The spatial variant of the detail-control factor (alpha as a 2D tensor) is described but only demonstrated qualitatively; the paper should clarify how the alpha map is derived from the semantic mask and how it interacts with the rest of the pipeline.","section":"Section 3.2, Eq. (4) and Figure 4"},{"comment":"The text contains several typos, including 'higher-solution' where 'higher-resolution' is intended (e.g., Section 4.1 and Section 4.3), and the reference list has formatting artifacts (e.g., '2, 3, 4, 5, 7, 1' at the end of references [14], [20], and [25]). These should be cleaned up.","section":"Throughout"},{"comment":"The comparison with super-resolution post-processing (SDXL+Real-ESRGAN) is useful but somewhat orthogonal to the main contribution; it might be clearer to integrate this into the main paper as a discussion of the difference between upscaling and higher-resolution generation, or to keep it in the supplementary material as is.","section":"Section B.1, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The reader's report and the stress-test note correctly identify the main weaknesses: the 8k claim is not backed by quantitative evidence, and the frequency-separable artifact model is not validated. In my view, these are fixable with additional experiments and careful rewriting, so I recommend major revision rather than rejection. I also note that the comparison set is somewhat narrow; a more thorough comparison with recent tuning-free methods would help the authors support their 'first' claim without overstating it. The paper's empirical contribution is real and the ablation shows that the proposed components matter, but the current presentation overstates the certainty of the central claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"FreeScale is a solid engineering paper. The genuinely new piece is Eq. 7: fuse global and local self-attention by taking the high-frequency part of the global output and the low-frequency part of the local output. That is a plausible fix for repetitive-object artifacts, and the ablations in Table 3 show each component helps. The video extension is a nice bonus, and the limitations section is honest about cost, knowledge limits, and the DiT structure gap.\n\nThe main soft spot is the 8k claim. All quantitative results stop at 4096²; the 8k evidence is a small gallery with no FID/KID, no user study, no error analysis. The hyperparameters also change at 8k (α=1 instead of 2), so that is a different regime. The paper acknowledges BeyondScene reached 8k with pose input, but does not compare with LinFusion (16k) or other recent methods. Missing comparisons are also a real gap at lower resolutions.\n\nAlso: no seeds, no error bars, no code or prompt lists, and SDXL-DI still holds the best FIDc/KIDc at both resolutions. The paper explains why, reasonably, but it is a caveat. These are normal weaknesses for this type of work and are minor next to the 8k support gap.\n\nOverall, the core idea is well-motivated, the experiments at 2048/4096 are decent, and the writing is clear. It deserves a serious referee, who should ask for 8k quantitative evaluation (even on a subset), seeds/error bars, and comparisons with LinFusion and BeyondScene. I would send it to review.","headline":"A solid engineering paper with a genuinely new frequency-based fusion rule; the 8k resolution claim is not supported by distribution-level evidence.","tokens_in":16175,"tokens_out":2790,"would_cite":true,"duration_ms":27337,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Tuning-free scale fusion lets pre-trained diffusion models generate 8K images.","keywords":["diffusion models","high-resolution image generation","tuning-free inference","scale fusion","frequency decomposition","self-attention","text-to-video generation","8K generation"],"falsifier":"Generate a test image at 16x resolution with FreeScale and inspect the intermediate signals: if a diagnostic shows repeated objects surviving in the low-frequency band of the local attention output $G(h_{\\mathrm{local}})$, or structural collapse in the high-frequency band of the global output $h_{\\mathrm{global}} - G(h_{\\mathrm{global}})$, the frequency-separability premise fails. A simple variant is to run the fusion with the split inverted (high frequency from local, low frequency from global) and check whether repetitions reappear, which the paper's mechanism predicts they would.","tokens_in":15136,"feed_emoji":"🖼️","tokens_out":7842,"duration_ms":62827,"temperature":0.7,"pith_summary":"FreeScale is a tuning-free inference method that lets a pre-trained diffusion model produce images and videos at resolutions far above its training resolution. The paper's claim is that repetitive-object artifacts arise because going beyond training resolution increases high-frequency information, and that the fix is to fuse two receptive scales: keep global attention's high-frequency structure and local attention's low-frequency content, exchanging them through a Gaussian-blur split. On a frozen text-to-image model the method yields coherent 8192x8192 images, and on a text-to-video model it yields 640x1024 video, all without retraining. The authors present this as the first tuning-free method to reach 8K-resolution text-to-image generation.","feed_headline":"Scale fusion lets diffusion models generate 8K images","feed_subtitle":"Frequency-split attention extends pretrained models to 8K images and 640x1024 videos without retraining.","key_machinery":"The central object is the scale-fusion identity in Equation 7, applied to the self-attention output of each transformer block. It combines a global self-attention pass over the full feature map with a local self-attention pass over shifted crops, using Gaussian blur $G$ as the low-pass filter: the final output keeps the high-frequency component of the global pass and the low-frequency component of the local pass. A tailored self-cascade upsampling seeds the high-resolution denoising with a noised, upscaled version of the low-resolution result, and restrained dilated convolution is applied only to down-blocks and mid-blocks. The machinery's job is to balance local and global receptive scales so that neither local repetition nor globally scattered small objects survive.","core_discovery":"The central discovery is a frequency-domain recombination inside the self-attention layer. When a latent diffusion model is asked to denoise at a resolution higher than its training resolution, the self-attention output accumulates dispersed high-frequency signals that read as repeated objects. FreeScale computes self-attention twice, once over the full feature map and once over shifted local crops, then forms its output as $h_{\\mathrm{fusion}} = h_{\\mathrm{global}} - G(h_{\\mathrm{global}}) + G(h_{\\mathrm{local}})$, where $G$ is a Gaussian low-pass filter. This keeps the high-frequency part of the global attention output and the low-frequency part of the local attention output. The paper reports that this eliminates both local repetition and globally scattered small objects, allowing a frozen SDXL model to generate 8192x8192 images and VideoCrafter2 to generate 640x1024 videos, with flexible local detail control and flexible aspect ratios as additional capabilities.","pith_inferences":["Beyond the paper: if the frequency-separability assumption holds across architectures, the same recombination rule could be adapted to DiT-based diffusion models, which the paper notes face blur rather than repetition at high resolution and may need the inverse frequency split.","Beyond the paper: because the Gaussian blur kernel is sized in proportion to the resolution multiplier, the optimal blur scale should track the spatial-frequency cutoff where the model's training data ends; varying it could yield a diagnostic for each model's practical resolution ceiling.","Beyond the paper: keeping the global high-frequency structure means FreeScale acts as a prior-driven upsampler, which suggests a direct extension to controlled image-to-image and video upscaling beyond the demonstrated FLUX-seeded 8K example."],"forward_implications":["A frozen text-to-image model can generate 8192x8192 images on a single A800 GPU, with quality metrics better than prior tuning-free baselines and competitive with super-resolution post-processing.","A frozen text-to-video model can generate 640x1024 videos with better FVD, dynamic degree, and aesthetic quality than prior tuning-free methods, at lower inference time.","Users can control the detail level spatially through a semantic mask and perform local semantic edits, such as changing hair color or facial features, during high-resolution denoising.","The method supports flexible aspect ratios that are multiples of 512 and can combine with distilled fast models such as SDXL-Turbo to produce 2048x2048 images in as few as two denoising steps."],"supporting_citations":[{"why":"Supplies the frozen text-to-image model whose 1024x1024 training resolution FreeScale extends; all image experiments use it.","marker":"[40]"},{"why":"Supplies the frozen text-to-video model (VideoCrafter2) at 320x512 that FreeScale extends in the video experiments.","marker":"[10]"},{"why":"Provides the self-cascade upscaling and global-local fusion strategy that FreeScale builds on and compares against as a baseline.","marker":"[14]"},{"why":"Introduces the dilated-convolution diagnosis of object repetition that FreeScale adopts with restrictions; also a baseline.","marker":"[20]"},{"why":"Supplies the frequency-domain analysis of repetition and the baseline that FreeScale competes with on frequency editing.","marker":"[25]"},{"why":"Supplies the self-cascade framework behind Equation 3 for progressive resolution upscaling.","marker":"[16]"},{"why":"Introduces the local-patch fusion mechanism that underlies the local self-attention pass in scale fusion.","marker":"[2]"},{"why":"Provides the shifted-crop local attention practice that FreeScale follows for local receptive-scale information extraction.","marker":"[41]"}],"fun_headline_variants":["FreeScale fuses attention scales to hit 8K diffusion images","Tuning-free scale fusion yields 8K diffusion generation","Frequency-split attention unlocks 8K image and video synthesis","Pretrained diffusers reach 8K via scale fusion alone","Scale fusion removes repetition for 8K diffusion output"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The fusion rule assumes that all high-frequency content produced by local attention is harmful repetition while the low-frequency content of global attention is semantically correct; the paper does not analyze whether that frequency split actually separates artifacts from structure.","fun_headline_variants_meta":{"raw":{"variants":["FreeScale fuses attention scales to hit 8K diffusion images","Tuning-free scale fusion yields 8K diffusion generation","Frequency-split attention unlocks 8K image and video synthesis","Pretrained diffusers reach 8K via scale fusion alone","Scale fusion removes repetition for 8K diffusion output"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000347,"raw_usage":{"total_tokens":1890,"prompt_tokens":924,"completion_tokens":966,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":882}},"tokens_in":540,"tokens_out":966,"duration_ms":8734,"temperature":1.0,"reasoning_tokens":882,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:50:58.509579+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a test image at 16x resolution with FreeScale and inspect the intermediate signals: if a diagnostic shows repeated objects surviving in the low-frequency band of the local attention output $G(h_{\\mathrm{local}})$, or structural collapse in the high-frequency band of the global output $h_{\\mathrm{global}} - G(h_{\\mathrm{global}})$, the frequency-separability premise fails. A simple variant is to run the fusion with the split inverted (high frequency from local, low frequency from global) and check whether repetitions reappear, which the paper's mechanism predicts they would.","supporting_citations":[{"cited_title":"Videocrafter2: Overcoming data limitations for high-quality video diffusion models, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen text-to-video model (VideoCrafter2) at 320x512 that FreeScale extends in the video experiments."},{"cited_title":"Demofusion: Democratising high- resolution image generation with no $$$","cited_arxiv_id":null,"evidence_quote":"Provides the self-cascade upscaling and global-local fusion strategy that FreeScale builds on and compares against as a baseline."},{"cited_title":"Scalecrafter: Tuning-free higher- resolution visual generation with diffusion models","cited_arxiv_id":null,"evidence_quote":"Introduces the dilated-convolution diagnosis of object repetition that FreeScale adopts with restrictions; also a baseline."}],"review_version":1}