{"id":"c853c9a7-d34d-46a4-8162-2c53f0e5c1f7","arxiv_id":"2608.07003","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"HRDiT enables off-the-shelf diffusion transformer text-to-image models to produce coherent high-resolution images with far less computation, without any retraining.","lead":"HRDiT is a training-free method that adapts off-the-shelf diffusion transformer image models such as FLUX and Stable Diffusion 3 to generate 2K, 4K, and 8K images, using position-embedding adjustments and per-head attention pruning. It reports better FID and CLIP scores and 2x to 4x faster generation than existing training-free high-resolution baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"HAP's per-head scope selection is optimized against a loss L (Eq. 7) that is never defined; without L, the pruning component and reported quality gains are not independently checkable, so the central claim stands only conditionally.","rationale":"I read the core claim as conditional on HAP's per-head scopes being derived from a meaningful quality loss. The empirical results are strong and the SPA component has independent ablation support, but HAP is not an optional speed-only extra: it is the mechanism that delivers the 2-4x latency reductions while preserving FID and CLIP. The preparatory stage optimizes Eq. 8 against Iq estimates from Eq. 7, so the correctness of the scope assignment hinges entirely on L. The manuscript does not define L. This is not a disagreement with consensus; it is a missing specification that blocks reproduction and independent verification. It also creates a concrete circularity risk: without a stated L, any metric reported in Tab. 1 could in principle be the very objective used to pick the windows. The ablation against fixed/random windows is helpful evidence that HAP is not vacuous, and the user study and additional ablations suggest the authors did run a real pipeline; I am not contesting that. I agree with the reader's weakest_assumption, and I see no basis to change the CONDITIONAL verdict: the concern is concrete, addressable, and should be settled by code inspection or by running the pipeline with a defined L.","tokens_in":28114,"tokens_out":4871,"duration_ms":52136,"concrete_test":"Inspect the released repository for the definition and implementation of L used in Eq. 7. If it is absent, re-run the HAP preparatory stage on FLUX 4K with a fully specified, reference-free loss (e.g., LPIPS distance between the final image and the same image generated under full attention) and recompute the 'FLUX + Ours' row of Tab. 1. If the resulting FID, CLIP, and latency are materially unchanged, the concern is resolved; if the choice of L changes outcomes, or if the code reveals that L is FID, KID, CLIP, or a metric directly aligned with them, the central quality claim is circular and needs re-evaluation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that L in Eq. 7 of Sec. 3.2 is a meaningful, well-defined measure of generation-quality degradation, and that the single-pass Taylor estimate faithfully predicts the effect of omitting the attention pairs in S_omit. The paper never defines L. Sec. 3.2 says only that L is the loss computed on the image generated from the single model pass; Supplementary B adds that L is computed using the output from the final denoising step at that resolution, but gives no formula, network, reference image, or metric. This matters because Preparatory Step 1 uses Eq. 7 to choose each head's window, and those windows produce the quality/latency numbers in Tables 1-3. If L is one of the reported metrics (FID, KID, CLIP) or a feature-space proxy hand-tuned to correlate with them, the HAP assignment is selected on the evaluation criterion, making the 'quality preserved after pruning' claim circular. If L is instead arbitrary, the derivation of Eq. 7 cannot be checked and the entire HAP component is unfalsifiable from the manuscript alone. The ablations in Tab. 5 show that fixed/random/intermediate-feature variants are worse, which partially supports HAP, but they do not resolve the identity of L.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HRDiT, a training-free framework for adapting off-the-shelf Diffusion Transformer (DiT) text-to-image models such as FLUX and Stable Diffusion 3 to high-resolution generation. It identifies two challenges: spatial disorder and long generation time. To address them, HRDiT introduces Spatial Position Alignment (SPA), which replaces token indices with bundle/slide indices and averages over N positional mappings before attention, and Head-adaptive Attention Pruning (HAP), which assigns each attention head a per-head window chosen by optimizing an estimated quality-loss degradation under a computational-cost constraint. Experiments at 2K, 4K, and 8K report that HRDiT outperforms all compared training-free baselines on FID, FIDp, KID, KIDp, and CLIP score while reducing latency by roughly 2x--4x, with ablations showing both components contribute.","tokens_in":28426,"tokens_out":4388,"duration_ms":48000,"significance":"If the claims hold, HRDiT fills an important gap: existing training-free high-resolution methods target U-Net backbones, while the paper tackles the increasingly dominant DiT architecture. The empirical work is extensive, spanning two backbones, three resolutions, five quality metrics, a user study, comparison against a super-resolution baseline, and ablations that separate the contributions of SPA and HAP. The release of code is a further strength. However, the HAP component is optimized against a loss L that is never defined (Eq. 7), which makes the pruning-quality trade-off uncheckable and risks circularity if L coincides with one of the reported metrics. In addition, the theoretical support for SPA establishes per-token uniqueness of bundle-index tuples, not the pairwise distinguishability that Eq. 6 and the h(T)=|S_pe| argument require. These are load-bearing issues, but they are local and fixable, and the empirical foundation is strong enough that the paper merits a major revision rather than rejection.","major_comments":[{"comment":"The loss L used to define I_q(n_head, n_scope) is never defined in the main paper or the supplementary material. Eq. (7) is a Taylor expansion involving derivatives dL/dA(u,v), but no formula, network, reference image, or metric is given for L. Supplementary B merely states that L is computed from the output of the final denoising step at that resolution. This is load-bearing because the per-head scope assignment in HAP is selected by minimizing the I_q estimate in Eq. (7), and the resulting assignment produces the headline quality/latency numbers in Tables 1 and 3. If L is one of the reported metrics (FID, KID, CLIP, or a feature-space proxy tuned to them), the quality-preservation claim becomes circular; if L is arbitrary, the derivation cannot be checked. The authors should define L explicitly and either prove Eq. (7) or provide the derivation, including how the softmax and the denominator 1-A(u,v) are handled when A(u,v)=1.","section":"Sec. 3.2, Eq. (7)"},{"comment":"The theoretical justification for SPA is incomplete. Proposition 1 bounds h(T), the number of distinguishable pairwise positional signals under the original function g with f_pe(i,j). The SPA operation in Eq. (6) replaces c_{i,j} with an average over N different bundle-index mappings, so the function whose distinguishability matters is a new averaged function, not the g analyzed in Proposition 1. The proof in Supplementary J shows that each token index i has a unique N-tuple of bundle indices, but it does not show that distinct pairs (i,j) and (i',j') produce averaged outputs that remain separated under the distance dis_g used in Eq. (4). The paper should either provide a formal pairwise-distinguishability statement for the SPA-averaged attention contribution or soften the claim that SPA restores h(T)=|S_pe|.","section":"Sec. 3.1, Eq. (6) and Proposition 1"},{"comment":"Several HAP and SPA hyperparameters are selected by sweeping on the same quality metrics that appear in the headline tables. In particular, N_scope (Tab. 8), the bundle size N (Tab. 9), the cost ratio r_c (Tab. 10), and the number of prompts for I_q estimation (Tab. 11) are all chosen by comparing FID, FIDp, KID, KIDp, and CLIP values. The 1,000 evaluation prompts are sampled from LAION-5B, and although the 30 prompts used for I_q estimation are explicitly disjoint from the evaluation prompts, it is not stated whether the hyperparameter sweeps in the supplementary used the same evaluation set. If they did, part of the reported gains reflects tuning to the test set. The authors should clarify the selection protocol and, if needed, re-run the main comparisons with hyperparameters fixed on a held-out validation set.","section":"Sec. 4.1 and Supplementary Tables 8--11"}],"minor_comments":[{"comment":"The indexing is inconsistent: the text defines token vectors as {q_t}_{t=0}^{T-1}, but the softmax sum in Eq. (2) is written as running from t=1 to T. Please unify the indexing.","section":"Sec. 3.1, Eq. (2)"},{"comment":"The text says the precompiled assignment can be applied 'in a cost-free manner,' but the associated per-image latency in Tables 1--3 excludes the one-time preparatory stage reported in Table 7 (about 48 minutes for FLUX at 4K). This should be stated clearly wherever latency numbers are discussed, since 'training-free' and 'cost-free' are distinct claims.","section":"Sec. 3.2"},{"comment":"The floor notation in Eq. (5) and in the caption of Fig. 3 is rendered inconsistently (e.g., 'T+N-N_1 over N' appears with different bracket styles). Please use a consistent notation such as \\lceil\\cdot\\rceil or \\lfloor\\cdot\\rfloor throughout, and verify the exact intended grouping.","section":"Sec. 3.1, Fig. 3 and Eq. (5)"},{"comment":"Reference [49] contains a typo: 'Ning, X., , Chen, B.' has an extra comma after 'X.' Please correct.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The empirical contribution is substantial and the ablations give reasonable evidence that both SPA and HAP matter. The main risk is the undefined loss L in Eq. (7), which is exactly the kind of omission that a careful reader will treat as circular or unfalsifiable. If the authors define L concretely (e.g., a perceptual distance to the full-attention output or to a reference image), show the Taylor derivation, and clarify the hyperparameter selection protocol, the paper could become acceptable. The 'training-free' label should also be qualified with the per-resolution one-time preparation cost."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth your time. HRDiT shows real gains: on FLUX and Stable Diffusion 3 at 2K, 4K, and 8K, it beats the compared training-free baselines on FID, KID, CLIP, and cuts latency by 2-4x. The ablations in Tables 2 and 5 and the supplementary are reasonably thorough, and the user study, while small, backs the headline. Code is promised. That is more than many arXiv submissions deliver.\n\nThe new thing is the specific combination: bundle/slide positional index manipulation (SPA) plus per-head attention-window pruning (HAP) for off-the-shelf DiT image models. Both pieces are adapted from prior work (SelfExtend/LongDiff and MoA), and the paper says so. That is a legitimate extension, not a new mechanism. The empirical validation is the contribution, and it is mostly solid.\n\nThe soft spots are real but not fatal. The big one: the loss L in Eq. 7 is never defined. HAP picks each head's window by minimizing an estimated quality degradation that is a Taylor expansion around L. If L is one of the reported metrics (FID, KID, CLIP) or a hand-tuned proxy, then the pruning assignment is selected on the evaluation criterion, and the 'quality preserved' claim becomes circular. If L is arbitrary, the whole HAP component is unfalsifiable from the paper alone. The supplementary says L is computed from the final denoising output at that resolution, but gives no formula, no reference image, no metric. That is a gap the authors can close, but right now it blocks independent checking.\n\nSecond, the SPA theory in Sec. 3.1 proves per-token uniqueness of the slide representation, but Eq. 6 needs a statement about pairwise distinguishability of positional signals after averaging. The uniqueness result is necessary, not sufficient. Minor.\n\nThird, hyperparameters (N, N_scope, r_c, prompt count) are chosen by evaluating the same FID/KID/CLIP metrics used in the headline tables, with no error bars. That is tuning on the test metric, which weakens the strength of the headline numbers. It doesn't overturn them, but it should be acknowledged.\n\nFourth, DiTFastAttnV2 is a close baseline and it is cited but not compared directly. That is an omission, not a fatal flaw.\n\nMy take: the central empirical claim — the method works and is fast — probably holds. But the HAP quality-preservation claim is not independently checkable until L is defined. The authors need a serious referee, not a desk rejection: the paper is substantive, readable, and the issues are fixable. If I were the editor, I'd send it out with a request to define L, report error bars, and add the missing baseline.","headline":"Strong, honest empirical paper on training-free high-resolution DiT generation, but HAP's pruning component rests on an undefined loss that blocks independent verification.","tokens_in":28962,"tokens_out":2071,"would_cite":true,"duration_ms":18839,"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":"HRDiT lets off-the-shelf diffusion transformers draw 2K, 4K, and 8K images without retraining, by sliding token indices apart and pruning attention heads to their native windows.","keywords":["text-to-image generation","high-resolution image synthesis","diffusion transformer","training-free adaptation","positional embedding extrapolation","attention pruning","FLUX","Stable Diffusion 3"],"falsifier":"Run the same 4K FLUX experiment twice: once with the authors' scope-selection loss L and once using an explicitly defined held-out perceptual metric, for example FID on a separate prompt set, as the selection loss. If the two selected per-head windows differ markedly, or if using the reported FID as L reproduces the paper's tables exactly, the single-pass quality estimate's role in the result is unsupported.","tokens_in":27896,"feed_emoji":"🖼️","tokens_out":8136,"duration_ms":74470,"temperature":0.7,"pith_summary":"HRDiT argues that off-the-shelf diffusion transformer image generators such as FLUX and Stable Diffusion 3, trained at around 1024×1024, can be pushed to 2K, 4K, and 8K synthesis with no retraining. It attributes the two main blockers to positional embedding collapse and quadratic attention cost. Its response is SPA, which feeds the positional function sliding bundle indices instead of raw token indices, and HAP, which gives each attention head a per-head window chosen under a compute budget. The reported results across 2K, 4K, and 8K show better FID, KID, and CLIP scores and lower latency than every compared training-free baseline, making high-resolution synthesis a plug-in fix rather than a retraining project.","feed_headline":"DiT image models reach 8K quality without retraining","feed_subtitle":"Sliding positional indices and per-head attention windows yield higher quality 2K–8K images, 2–4x faster.","key_machinery":"The load-bearing object is the modified attention contribution $c^{\\mathrm{SPA}}_{i,j} = \\frac{1}{N}\\sum_{n=1}^{N} g(x_i, x_j, f_{\\mathrm{pe}}(\\phi_{\\mathrm{bundle}}^{(N_1=n)}(i), \\phi_{\\mathrm{bundle}}^{(N_1=n)}(j)))$, in which token indices $i,j$ are mapped to bundle indices by $\\phi_{\\mathrm{bundle}}$; SPA runs $N$ variants with first-bundle sizes $N_1=1,\\ldots,N$ so the bundle boundaries slide, and each token gets a unique tuple of bundle indices. This directly shrinks the number of distinct pairwise positional signals $|S_{\\mathrm{pe}}|$ that the positional encoder must separate, which is what Proposition 1 identifies as the bottleneck. The other load-bearing object is the HAP quality estimate $I_q(n_{\\mathrm{head}}, n_{\\mathrm{scope}})\\approx \\sum_{(u,v)\\in S^{(n_{\\mathrm{scope}})}_{\\mathrm{omit}}}[\\frac{\\partial L}{\\partial A(u,v)}(-A(u,v)) + \\sum_{w\\ne v}\\frac{\\partial L}{\\partial A(u,w)}\\frac{A(u,v)A(u,w)}{1-A(u,v)}]$, a one-forward-pass Taylor estimate of the quality cost of dropping the attention pairs outside each candidate window; a linear-solver pass over Eq.~8 then assigns each head a window under a total-cost ratio $r_c$. Together these two mechanisms convert resolution extrapolation into a change of input indices and a per-head pruning plan, with no training.","core_discovery":"The central claim is that training-free high-resolution generation with off-the-shelf DiT models is not only possible but can beat existing training-free pipelines on quality and speed. Spatial disorder is traced to limited expressiveness of the positional embedding mechanism at large token counts: a pseudo-dimension bound (Proposition 1) shows the number of distinguishable pairwise positional signals cannot keep pace with the growth of the token grid. SPA restores distinguishability by replacing token indices with bundle indices and averaging over N sliding bundle offsets, yielding a unique positional representation for every token. Long generation time is traced to quadratic multi-head attention, and HAP prunes it by estimating, in a single forward pass, the quality degradation caused by omitting each candidate set of attention pairs, then solving an integer program for the per-head window sizes under a cost ratio r_c. Experiments on FLUX and SD3 at 2K, 4K, and 8K report better FID, KID, and CLIP scores and roughly 2–4x lower latency than the compared training-free baselines.","pith_inferences":["The same bundle/slide trick could be probed on non-generative token-index transformers, such as video diffusion transformers or long-context vision encoders, that also fail at length extrapolation; the paper only demonstrates image generation.","HAP's single-pass estimate could be recomputed per prompt or per denoising step to make the speed-quality trade-off adaptive; the paper fixes one plan per resolution from the final denoising step and shows stability only in ablations.","If the loss L used in Eq. 7 is left undefined in the released code, the paper's central quality-preservation claim cannot be independently reproduced; defining L as an explicit perceptual metric would turn the method into a self-contained speed-quality optimizer."],"forward_implications":["A released FLUX or Stable Diffusion 3 checkpoint can generate 2K, 4K, and 8K images without any weight update; the only additions are index remapping and a precomputed per-head scope plan.","Reported 8K latency drops from 1,708s to 827s for FLUX and from 822s to 454s for SD3, with better FID/KID/CLIP scores than the compared baselines, so resolution scaling becomes a runtime choice rather than a retraining project.","Because SPA acts only on token indices before the positional function, the same fix should transfer to any diffusion transformer whose attention consumes token-index positional embeddings.","The one-time scope-preparation stage, about 48 minutes for 4K FLUX, is amortized over many images, after which the pruning plan is applied at near-zero runtime cost."],"supporting_citations":[{"why":"Supplies the pseudo-dimension bound (Lemma 1) on which Proposition 1 and the claim of limited positional expressiveness rest.","marker":"[11]"},{"why":"Source of the bundle/slide positional manipulation that SPA adapts to high-resolution DiT image generation.","marker":"[26]"},{"why":"Source of the single-pass Taylor estimate and integer-program scope assignment used by HAP.","marker":"[49]"},{"why":"The DiT-tailored training-free baseline (I-Max) that HRDiT must beat on FLUX.","marker":"[7]"},{"why":"Supplies the cascaded generation strategy and one of the strongest comparison baselines (HiFlow).","marker":"[3]"},{"why":"Defines the evaluation protocol with FID, KID, FIDp, KIDp, and CLIP score that the main tables follow.","marker":"[20]"},{"why":"The Stable Diffusion 3 backbone to which HRDiT is applied and evaluated.","marker":"[8]"},{"why":"The FLUX backbone to which HRDiT is applied and evaluated.","marker":"[22]"},{"why":"The U-Net-based training-free baseline (FreeScale) used for cross-backbone comparison.","marker":"[34]"}],"fun_headline_variants":["DiT models hit 8K without training: faster and sharper","Training-free 8K images from off-the-shelf DiTs","Sliding windows and attention pruning speed up DiT upscaling","Off-the-shelf DiTs generate 8K images 4x faster","DiT upscaling: 8K quality without retraining"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the paper's unstated loss L, used by Eq. 7 to pick per-head windows, really measures generation-quality loss; if L is unrelated to image quality or is itself one of the reported FID/KID numbers, the pruning-preserves-quality claim is circular.","fun_headline_variants_meta":{"raw":{"variants":["DiT models hit 8K without training: faster and sharper","Training-free 8K images from off-the-shelf DiTs","Sliding windows and attention pruning speed up DiT upscaling","Off-the-shelf DiTs generate 8K images 4x faster","DiT upscaling: 8K quality without retraining"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000251,"raw_usage":{"total_tokens":1537,"prompt_tokens":904,"completion_tokens":633,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":540}},"tokens_in":520,"tokens_out":633,"duration_ms":6025,"temperature":1.0,"reasoning_tokens":540,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:48:09.405842+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 4K FLUX experiment twice: once with the authors' scope-selection loss L and once using an explicitly defined held-out perceptual metric, for example FID on a separate prompt set, as the selection loss. If the two selected per-head windows differ markedly, or if using the reported FID as L reproduces the paper's tables exactly, the single-pass quality estimate's role in the result is unsupported.","supporting_citations":[{"cited_title":"In: Proceedings of the Computer Vision and Pattern Recognition Conference","cited_arxiv_id":null,"evidence_quote":"Source of the bundle/slide positional manipulation that SPA adapts to high-resolution DiT image generation."},{"cited_title":"In: Proceedings of the AAAI conference on artificial intelligence","cited_arxiv_id":null,"evidence_quote":"Defines the evaluation protocol with FID, KID, FIDp, KIDp, and CLIP score that the main tables follow."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The FLUX backbone to which HRDiT is applied and evaluated."}],"review_version":1}