{"id":"4259a3e6-ce65-43f7-9b57-437c4e65730d","arxiv_id":"2511.19778","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Expressing all RoPE positions on the query's grid ('one attention, one scale') plus a small boundary content-exchange step restores mixed-resolution diffusion generation that naive position interpolation destroys.","lead":"This paper shows why mixing high- and low-resolution patches inside one diffusion-transformer attention step collapses generation, and offers a training-free fix. The fix re-expresses every token's position on the query's own grid so distances are always measured consistently, recovering near-full-resolution quality at roughly a quarter of the compute.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central causal claim — that RoPE phase mismatch, not content-fidelity mismatch, drives mixed-resolution failure — is untested: no ablation runs Boundary Expand-and-Replace on a non-CRPA baseline, and Tab. 8 shows the content-exchange module contributes more than phase alignment.","rationale":"The paper's most consequential claim is causal: pretrained DiT mixed-resolution denoising collapses because RoPE phases are sampled at incompatible rates, and CRPA fixes this by enforcing one coherent reference grid per query-key dot product. Sec. 3's derivation of Eq. 6 is a correct identity, and the method's overall gains over PI-LR/PI-HR/YaRN are consistently reported. But the experimental design does not isolate the phase mechanism. CRPA's LR-query branch changes content resolution as well as RoPE indexing, and the boundary module is a pure content-exchange operation that adds +6.9 DOVER over CRPA alone, more than the +5.0 CRPA adds over PI-LR (Tab. 8). Without a boundary-without-CRPA ablation, the observed improvement could be mostly content harmonization rather than phase alignment. The κ(Δ) curves are suggestive of sharp learned phase kernels, but they are measured at a single resolution and do not reproduce the mixed-resolution confound of content and phase. This does not invalidate the method as an engineering contribution, but it leaves the central explanatory claim underdetermined. The reader's CONDITIONAL verdict is appropriate; my stress test identifies the same load-bearing assumption and would keep the verdict unchanged pending the proposed ablation.","tokens_in":24745,"tokens_out":3777,"duration_ms":39558,"concrete_test":"Run a 2x2 ablation on Wan2.1-1.3B with fixed seeds and the paper's 50-step / 15% HR-token schedule: (1) PI-LR only (reproduces 63.39), (2) PI-LR + Boundary Expand-and-Replace, (3) CRPA only (reproduces 68.43), (4) CRPA + Boundary (reproduces 75.34). Also add a mismatched-phase control: apply CRPA's exact content-handling (HR keys subsampled for LR queries, LR keys position-lifted for HR queries) but use naive linear-interpolated RoPE positions instead of phase-aligned ones. If cell (2) reaches ~75 or the mismatched-phase control matches CRPA, phase mismatch is not the dominant failure driver and the central principle in Sec. 3.3 is incomplete; if cell (2) remains near 68 and the control collapses, the causal claim survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 defines CRPA as changing both the RoPE index map and, for LR queries, the content of HR keys: the LR-query branch 'subsamples HR keys in content' (Sec. 4.1, Fig. 4). Section 4.2 introduces Boundary Expand-and-Replace, a content-exchange step that alone raises DOVER overall from 68.43 to 75.34 over the same CRPA base (Tab. 8). The paper never runs Boundary Expand-and-Replace on top of PI-LR or YaRN, and never compares a condition in which HR-key content is subsampled/lifted while RoPE positions are left misaligned. Thus the empirical gain attributed to phase alignment is confounded with content-resolution matching: a pretrained DiT may fail under PI simply because LR queries receive HR-key content sampled at the wrong spatial density, not because phases are aliased. The κ(Δ) analysis (Sec. 3.3) does not break this confound: it measures a kernel at a single sampling rate and uses Δ=0 pairs with identical content, so it cannot predict the mixed-resolution content/phase interaction. Until a non-CRPA baseline with the same content-sampling and boundary exchange is tested, 'the issue is not the interpolation scheme but the lack of a uniform sampling rate for Δ' (Sec. 3.3) is an assertion, not a demonstrated causal result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies why linear interpolation of RoPE positions fails in mixed-resolution diffusion transformers, and proposes Cross-Resolution Phase-Aligned Attention (CRPA). The central principle is that within each attention computation all query/key positions must be expressed on the query's spatial stride, so that equal physical distances produce equal phase increments. A supplementary derivation (Eq. (6)) shows that the RoPE-modulated attention score is a mixture of sinusoids in the relative offset. The paper measures a phase kernel κ(Δ) on pretrained DiTs, observing a sharp peak near Δ=0, periodicity, and amplification in RoPE-dominant heads. CRPA is a training-free index rescaling, augmented by a Boundary Expand-and-Replace module that exchanges content across resolution boundaries. Experiments on Wan2.1-1.3B and FLUX.1-dev report large gains over PI/NTK/YaRN at roughly 4x lower latency than full-resolution inference.","tokens_in":24902,"tokens_out":3256,"duration_ms":35361,"significance":"If the causal attribution is correct, the paper provides a simple, training-free, drop-in fix for a practically important failure mode in mixed-resolution DiT inference. The strengths are real: the derivation in Supplementary A is a correct trigonometric identity; κ(Δ) is a measurement rather than a fitted relation; CRPA has no tunable parameters beyond the schedule choices; and the method is tested on public benchmarks (DOVER, VBench, MSCOCO) against external baselines. The main limitation is that the empirical evidence does not isolate phase misalignment from content-resolution mismatch, and the reported metric deltas are not accompanied by uncertainty estimates. These issues are addressable and do not invalidate the method itself, but they currently leave the central causal claim only partially supported.","major_comments":[{"comment":"The causal claim that phase mismatch — not content-fidelity mismatch — drives mixed-resolution failure is not isolated by the current ablation. With CRPA alone, DOVER overall is 68.43 over a PI-LR baseline of 63.39; adding Boundary Expand-and-Replace raises it to 75.34. Thus the content-exchange module contributes a larger share of the total gain than phase alignment alone. The paper never runs Boundary Expand-and-Replace on top of PI-LR or YaRN, and never compares a condition in which HR-key content is subsampled/lifted while RoPE positions are left misaligned. As written, the sentence in Sec. 3.3 ('the issue is not the interpolation scheme but the lack of a uniform sampling rate for Δ') is an assertion. Please add the missing controls: content exchange on non-CRPA baselines and a content-matched/phase-mismatched condition.","section":"Sec. 4.2, Tab. 8"},{"comment":"The κ(Δ) measurement is a single-rate kernel: it samples offsets on one grid, uses Δ=0 pairs with identical content, and averages over content. It is consistent with the phase-aliasing story but cannot predict the mixed-resolution interaction between phase mismatch and content-density mismatch. A direct diagnostic — for example, measuring attention entropy or κ on LR-query/HR-key pairs with matched versus mismatched RoPE reference scales — would substantially strengthen the causal attribution.","section":"Sec. 3.3, Fig. 3"},{"comment":"All quantitative tables report single-run metrics without error bars or significance tests. Several headline differences are small: in Tab. 2 FID is 32.04 (Ours) versus 31.50 (HR) and 33.83 (YaRN); in Tab. 3 FID is 32.45 (Ours-18) versus 32.67 (FLUX-50) and 32.91 (RALU-18). Deltas around 0.2–0.5 are typically within sampling noise for FID on 5K pairs. The claims 'close to HR' and 'surpasses RALU' need at least multiple seeds or a confidence interval, otherwise the comparisons are not reliable.","section":"Tabs. 1–3, 5–7"}],"minor_comments":[{"comment":"The abstract introduces the method as 'Phase-Aligned Mixed-Resolution Attention (PMA)', but the rest of the paper calls it 'Cross-Resolution Phase-Aligned Attention (CRPA)'. Please unify the acronym.","section":"Abstract vs. Sec. 4.1"},{"comment":"Typo: 'denosing' should be 'denoising' in both captions.","section":"Table 1, Table 2 captions"},{"comment":"The text says 'a coarse stage at 480p followed by a mixed-resolution stage', but Fig. 6 includes a '480p-then-960p' reference (LR+HR). Clarify whether this reference is the same two-stage schedule or a separate full-resolution upper bound; the description is currently ambiguous.","section":"Sec. 5.1, video implementation"},{"comment":"The threshold 'rds > 0.085' appears only in the caption and is not defined in the main text or supplementary; please state how it was chosen.","section":"Fig. 3 caption"}],"recommendation":"major_revision","confidential_remarks":"The skeptic's concern is legitimate and load-bearing. The paper's method may well be correct, but the missing control — boundary/content exchange on non-CRPA baselines — directly affects the main causal claim, not just a peripheral comparison. I would ask for that experiment, plus error bars on the key FID/DOVER comparisons, before recommending acceptance. The paper is otherwise well within scope and the core idea is publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper identifies a real, practically painful failure mode — linear RoPE interpolation collapses attention in mixed-resolution DiT inference — and offers a simple training-free fix that delivers close to full-resolution quality at ~4x lower latency. That part is solid and worth taking seriously. But the paper's central claim, that phase mismatch is the driver, is not actually isolated. The fix changes two things at once, and the ablation says the second thing matters more.\n\nWhat's genuinely new: the observation that pretrained DiT attention heads behave as sharply peaked, periodic phase kernels over RoPE offset Δ, and the clean 'one attention, one scale' principle behind CRPA. Eq. (6) is a correct trigonometric identity (derivation in the appendix checks out). The empirical wins are consistent across video (Wan2.1-1.3B) and image (FLUX.1-dev): near-HR DOVER/VBench/FID/MUSIQ at roughly 4x speedup, and it composes with caching methods like TeaCache/MagCache without degradation. That is practically valuable.\n\nThe soft spot is in the causal attribution, and it's load-bearing. CRPA's LR-query branch doesn't just re-index RoPE positions; it subsamples HR key content to the LR grid. Meanwhile, the Boundary Expand-and-Replace module exchanges content across the boundary. Tab. 8 shows the boundary module accounts for most of the gain over PI-LR (68.43 → 75.34 DOVER overall, vs 63.39 → 68.43 for CRPA alone). The paper never runs boundary exchange on PI-LR or YaRN, and never holds content resolution fixed while toggling only phase alignment. So the phrase 'the issue is not the interpolation scheme but the lack of a uniform sampling rate for Δ' is an assertion, not a demonstrated result. The κ(Δ) analysis is suggestive but doesn't break the confound either — it's a single-sampling-rate measurement, not a mixed-resolution interaction experiment.\n\nSmaller issues: every table is single-run with no error bars, and some 'beats the full-res baseline' claims (e.g., FID 32.45 vs 32.67; VBench 0.808 vs 0.801) are noise-scale. No code is released; the learned resizer training is not specified; the content-subsampling branch is described verbally. Those are fixable.\n\nWho should read it: anyone working on efficient diffusion inference or RoPE behavior in vision DiTs. It deserves a serious referee; I would send it out but with a clear request for the missing boundary-without-CRPA ablation, per-seed statistics, and artifacts before acceptance.","headline":"Useful fix with a causal story that outruns the evidence — referee it, but make them untangle phase alignment from content exchange.","tokens_in":25619,"tokens_out":3401,"would_cite":true,"duration_ms":32653,"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":"RoPE phase mismatch, not content, breaks mixed-resolution diffusion; rescaling keys to the query grid restores quality.","keywords":["mixed-resolution denoising","rotary positional embeddings","diffusion transformers","phase alignment","position interpolation","training-free acceleration","image generation","video generation"],"falsifier":"A controlled experiment that turns off phase alignment while keeping the boundary content-exchange module: if the resulting DOVER/FID matches the full method, the phase-mismatch account is falsified. Conversely, running CRPA with content fully preserved (HR keys not subsampled in content for LR queries) would isolate the phase effect.","tokens_in":24424,"feed_emoji":"⚡","tokens_out":8579,"duration_ms":78194,"temperature":0.7,"pith_summary":"This paper claims that mixed-resolution denoising with diffusion transformers fails because rotary positional embeddings (RoPE) compare phases sampled at incompatible spatial rates within a single attention head, creating cross-rate aliasing that destabilizes the score landscape. The proposed fix, Cross-Resolution Phase-Aligned Attention (CRPA), enforces the principle \"one attention, one scale\": every query–key pair is evaluated with all positions expressed on the query's grid, so equal physical distances always induce identical phase increments. CRPA is training-free and drop-in, requiring only a change to the RoPE index map, and it stabilizes all heads and layers. The authors show that mixed-resolution inference with CRPA approaches full-resolution quality (video DOVER 75.34 vs 79.12; image FID 32.04 vs 31.50) at roughly four times lower latency, and that a lightweight boundary module further smooths content near resolution transitions.","feed_headline":"RoPE phase alignment rescues mixed-resolution diffusion","feed_subtitle":"Training-free CRPA puts every query-key dot product on one grid, matching full-res quality at ~4x lower latency.","key_machinery":"The central mechanism is the RoPE index map rescaling: before each dot product, key/value positions are multiplied by the ratio of query spacing to key spacing, α_{k→q} = S_q/S_k, so the attention score becomes ⟨q, R(α_{k→q} p_k − p_q) k⟩. This makes the phase increment depend only on physical distance measured on the query's grid. The supporting object is the phase kernel, the measured and derived function κ(Δ) = E[⟨q̂, R(Δ) k̂⟩] = Σ_i C_i(q,k) cos(ω_i Δ + φ_i), which quantifies the head's learned sensitivity to relative offset and explains why cross-rate interpolation fails. A second, lighter mechanism is the Boundary Expand-and-Replace module, which dilates the LR/HR masks and bidirection","core_discovery":"Pretrained DiT attention heads act as sharply tuned, multi-frequency phase kernels over the relative RoPE offset Δ. Empirically, the expected attention score κ(Δ) peaks sharply near Δ=0, oscillates periodically with Δ, and is amplified in RoPE-dominant heads, stably across diffusion timesteps. Formally, the score takes the form Σ_i C_i(q,k) cos(ω_i Δ + φ_i), a mixture of sinusoids at the fixed RoPE frequencies. Interpolating LR and HR tokens into one index space makes the same kernel see different Δ values for the same physical displacement, corrupting the score. CRPA re-indexes every key/value position onto the query's native grid — p_k^(q) = α_{k→q} p_k — so the score depends on a single r","pith_inferences":["The paper's ablation shows the boundary module adds a larger quality gain (DOVER 68.43 → 75.34) than CRPA alone over the PI-LR baseline (63.39 → 68.43). A direct test that ablates content exchange without phase alignment would determine whether phase mismatch or content-fidelity mismatch is the primary driver of the remaining gap.","Because the rule is stated per dot-product, it suggests a more general principle: any positional encoding that defines a phase over token offsets will be unstable in mixed-scale attention unless the effective sampling rate is constant within each dot product. This principle could transfer to other sinusoidal or relative encodings, and to any model mixing tokens at different native densities.","The asymmetry in CRPA — LR queries subsample HR keys both in index and content, while HR queries only lift LR indices — is a design choice that could be tested against a symmetric variant that exchanges content in both directions at every query scale.","The phase-kernel diagnostic (κ(Δ) vs Δ) could serve as a cheap probe for RoPE sensitivity in other architectures, possibly predicting which models will tolerate mixed-resolution inference before running full generation."],"forward_implications":["Mixed-resolution denoising becomes a stable, plug-and-play option for pretrained DiTs: no retraining, no architectural change, only a per-attention index map.","A coarse→mixed (optionally fine) schedule with high resolution restricted to salient regions yields near-full-resolution quality at about 4x lower latency, with the quality–cost curve saturating quickly at small HR token ratios.","The phase-alignment rule extends to arbitrary LR/HR layouts, irregular masks, and mixed spatiotemporal grids, supporting the same stable attention across all heads and layers.","CRPA composes with orthogonal acceleration techniques such as token caching, further cutting latency while keeping quality nearly unchanged (e.g., video DOVER 75.57 at 2.0x, image FID 32.76 at 1.6x)."],"fun_headline_variants":["Phase-aligned RoPE unifies mixed-resolution attention","Align RoPE phases, not positions, for multi-res DiT","Phase-aligned RoPE: one scale for all resolution tokens","Training-free phase alignment boosts multi-res DiT quality","Fix mixed-res RoPE with phase-aligned positions"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the mixed-resolution failure is driven by RoPE phase mismatch rather than by content-fidelity mismatch between LR and HR tokens; the paper's ablations do not isolate the two, and the boundary module (which exchanges content) accounts for the larger share of the gain over the interpolation baseline.","fun_headline_variants_meta":{"raw":{"variants":["Phase-aligned RoPE unifies mixed-resolution attention","Align RoPE phases, not positions, for multi-res DiT","Phase-aligned RoPE: one scale for all resolution tokens","Training-free phase alignment boosts multi-res DiT quality","Fix mixed-res RoPE with phase-aligned positions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001004,"raw_usage":{"total_tokens":4081,"prompt_tokens":742,"completion_tokens":3339,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":3258}},"tokens_in":486,"tokens_out":3339,"duration_ms":26324,"temperature":1.0,"reasoning_tokens":3258,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T20:26:15.014994+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment that turns off phase alignment while keeping the boundary content-exchange module: if the resulting DOVER/FID matches the full method, the phase-mismatch account is falsified. Conversely, running CRPA with content fully preserved (HR keys not subsampled in content for LR queries) would isolate the phase effect.","supporting_citations":[],"review_version":1}