{"id":"92c4eee5-bd1f-4aca-856b-b03ccdcce007","arxiv_id":"2508.07246","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MiraMo turns a static image into a video using a linear-attention transformer that learns inter-frame motion residuals, with DCT-based noise refinement and a user-controllable dynamics knob.","lead":"MiraMo animates a single photo into a short video by combining a linear-attention transformer, motion-residual learning (predicting how frames differ rather than the frames themselves), and a DCT-based noise-refinement trick during inference. The paper reports competitive quality with faster and cheaper generation, but its own tables only partly support the headline superiority claims.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Linear-attention quality parity is asserted, never tested: no controlled swap with vanilla attention, so the O(N) efficiency claim lacks its quality counterpart.","rationale":"The reader's verdict of CONDITIONAL is appropriate. My concern overlaps with the reader's weakest assumption---that cosine temporal linear attention matches vanilla attention---but I identify a broader and more structural gap: the entire paper lacks a controlled linear-vs-vanilla attention quality comparison, not just for the newly introduced temporal attention but for the T2V backbone. The paper asserts in the introduction that linear attention is 'better, faster, and cheaper' (Fig. 3) and in Sec. III-B that it 'can achieve performance comparable to vanilla attention,' yet no ablation or experiment backs this with an apples-to-apples comparison. The internal ablations in Table V are valuable and support MRL and DCTInit, and the paper deserves credit for the disclaimed visual comparisons and the acknowledgement of VAE-induced flicker. However, the absence of the core attention-type ablation is a concrete, addressable defect, not a fatal one: if a controlled swap shows parity, the claim is substantiated; if not, the central advantage dissolves. The mixed quantitative results (e.g., FVD worse than Cinemo on both I2V test sets, motion smoothness lower than Cinemo in Table III) make this gap more salient because the claimed 'best appearance consistency and motion smoothness' is already not uniformly supported by the paper's own numbers. I therefore keep the reader's CONDITIONAL/UNCHANGED verdict; no adjustment is needed, but the requested ablation should be mandatory before the linear-attention efficiency claim is accepted.","tokens_in":23510,"tokens_out":5458,"duration_ms":54402,"concrete_test":"Run a controlled head-to-head experiment: fix the full MiraMo pipeline (MRL, DCTInit, dynamics bucket, training data, 16 frames at 512x512, same optimizer, steps, classifier-free guidance, and inference solver). Create exactly one variant that replaces only the cosine temporal linear attention (Eq. 8-9) with vanilla softmax temporal attention (with RoPE), leaving everything else identical. Report VBench T2V average, I2V FVD/IS/FID/CLIPSIM on MSR-VTT and UCF-101, and per-sample latency/GPU memory. Also report a second variant where all linear attentions are replaced by vanilla attention to validate Fig. 3. If quality is within noise (e.g., <1% VBench average, <5% relative FVD), the linear-attention claim holds; if vanilla is meaningfully better, the central efficiency-quality claim fails. Additionally, the authors must specify the exact vanilla baseline behind Fig. 3 so the +4.76% claim i","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that replacing vanilla attention with linear attention in a video Transformer preserves generation quality while reducing cost (Sec. III-B, Fig. 3). This is the load-bearing premise of the entire method: if linear attention does not match vanilla attention in temporal modeling, the frame-only cosine attention of Eq. 8-9 and the O(N) efficiency argument collapse, leaving only the orthogonal contributions (MRL, DCTInit) which could be built on any backbone. However, no controlled experiment is reported that swaps linear attention for vanilla attention while holding data, training, and inference fixed. Fig. 3 compares throughput and memory and shows a '+4.76% VBench' number, but the vanilla-attention baseline is not described: is it the same SANA-style architecture with vanilla self-attention? Is it a DiT? Was it trained under the same data and compute? Table I compares against external T2V methods with different architectures, data scales, and training budgets, so it cannot isolate attention type. The claim in Sec. III-B that frame-dimension-only temporal attention outperforms full 3D temporal attention is presented without any quantitative comparison. The Eq. 7 instability argument for ReLU linear attention with RoPE is plausible but never empirically demonstrated; and Eq. 9 itself has a denominator that can approach zero in degenerate configurations, yet no practical resilience analysis is given. These gaps mean the key efficiency-quality tradeoff assertion rests on an untested assumption, making the paper's central contribution unverified in its most essential part.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"MiraMo is an image-to-video (I2V) framework built by extending the SANA linear-attention text-to-image model into a text-to-video model with a new cosine-kernel temporal linear attention compatible with RoPE, then adapting it to I2V via motion-residual learning (predicting z_i - z_1), DCT-based low-frequency noise initialization during inference, and a dynamics-degree control based on MS-SSIM buckets. The T2V backbone is trained on roughly 2M Vimeo25M videos and post-trained on about 150k synthetic videos; I2V is evaluated on UCF-101/MSR-VTT and VBench against U-Net and DiT baselines, with throughput/memory comparisons against vanilla attention. The paper claims high appearance consistency, smooth controllable motion, and O(N) attention complexity with no quality loss.","tokens_in":23825,"tokens_out":7031,"duration_ms":65628,"significance":"If substantiated, this would be a useful demonstration that linear attention can support video diffusion, with orthogonal contributions (motion-residual learning and DCTInit) that could transfer to other backbones. The paper's strengths include a focused ablation in Table V supporting MRL over direct frame prediction and DCTInit over FFT initialization, a quantitative dynamics-control analysis in Fig. 10, and a user study component. However, the load-bearing claim of linear-attention parity with vanilla attention is not directly tested, and some reported results conflict with the text (Table VI). The authors also state a self-admitted limitation (image-level VAE may cause flickering) that tempers the consistency claim.","major_comments":[{"comment":"The claim that linear attention 'preserves generation quality' is not established by a controlled experiment. Fig. 3 shows throughput/memory and a '+4.76% VBench' number but does not specify the vanilla-attention baseline's architecture, training data, or compute, so it cannot isolate the attention mechanism. Table I compares external T2V models with different data and budgets. Moreover, the statement that frame-dimension temporal attention 'outperforms' the full 3D variant is given without quantitative comparison. Without a same-cost linear-vs-vanilla ablation, the central O(N)-without-quality-loss conclusion is unsupported. Table II further shows Ours FVD 201.90 vs Cinemo 168.16 on UCF-101, so external comparisons do not supply the missing parity evidence.","section":"Sec. III-B, Fig. 3; Table II"},{"comment":"The synthetic-data claim is contradicted by the table. The text states that the model fine-tuned on synthetic data achieves 'over a 3% improvement' over real data, but Table VI reports Sync data values (45.50, 88.40, 88.40, 35.50) that are all lower than Real data (52.91, 92.00, 90.55, 37.18). If the table rows are swapped or the metrics mislabeled, this must be corrected; as written, Sec. III-F and Sec. IV-C's central claim about post-training data selection has no support.","section":"Sec. IV-C, Table VI"},{"comment":"The abstract/introduction claim of 'best appearance consistency and motion smoothness' is stronger than the evidence. In Table II, MiraMo's FVD is worse than Cinemo on UCF-101 (201.90 vs 168.16) and marginally worse on MSR-VTT (93.91 vs 93.51); in Table III, Cinemo has higher Motion Smoothness (98.21 vs 96.93) and Camera Motion (30.68 vs 21.57). The user study (Table IV) only includes SVD and SEINE, not Cinemo or PIA. The defensible claim is 'comparable FVD/CLIPSIM with better aesthetic quality and dynamic degree at lower compute.' The text should be revised to that claim, or additional experiments provided.","section":"Sec. I, Sec. IV-B, Tables II-III"},{"comment":"Eq. (9) as written does not have a non-negative denominator: D_i = n + (q_i/||q_i||) Σ_j (k_j/||k_j||)^T can approach zero because the sum of unit vectors has no lower bound. The paper's justification for rejecting Eq. (7) is that RoPE may make inner products negative, but the same concern applies to the denominator of Eq. (9). No empirical stability analysis is given. Since this attention is load-bearing for the efficiency claim, the authors should either provide a proof or regularizer ensuring D_i is bounded away from zero, or report numerical safeguards and a stability ablation.","section":"Sec. III-B, Eq. (9)"},{"comment":"The motion-residual training formulation is underspecified. The notation z1 is used both for the first-frame latent and for the Gaussian endpoint in the flow-matching interpolation z_t = t z1 + (1-t) z0 (Sec. III-A). In Algorithm 1, M = {z_i - z_1} is computed, then 'noised Mt via the diffusion process' and input X_t = cat([z_1, M_t + z_1]) are defined, but the flow-matching latent variables z_0 and z_1 for the residual distribution are never defined, nor is the velocity target v_t in Eq. (14) expressed in terms of M. This makes the training objective ambiguous and prevents reproducibility. Please define the noising schedule and target for residuals explicitly and reconcile the notation.","section":"Sec. III-C, Algorithm 1"}],"minor_comments":[{"comment":"The text says 'σ is the sigmoid function', but the scaled dot-product attention in Eq. (3) uses softmax, not sigmoid. This typo is confusing in a central equation.","section":"Eq. (3)"},{"comment":"Typo: 'Finlay' should be 'Finally'. Also the phrase 'increase both training and inference efficiency' should likely be 'increase computational cost'.","section":"Sec. III-B"},{"comment":"The implementation details state 'MiraMo is identical to SANA [9]' but the SANA reference is [38], not [9]. Please correct the citation.","section":"Sec. IV-A"},{"comment":"The inference-speed claim would be stronger with wall-clock generation times per video, not only throughput/memory plots. The current ablation reports no end-to-end latency for the I2V pipeline.","section":"Sec. IV-A"},{"comment":"Table V reports 'Average' and 'Average Consistency' without defining their composition beyond a parenthetical in the text. Please specify exactly which VBench dimensions are averaged and the number of prompts/videos used.","section":"Sec. IV-C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an extension of the authors' CVPR 2025 paper, and much of the novelty over [37] is the linear-attention backbone plus DCTInit. The missing controlled linear-vs-vanilla attention experiment and the internal contradiction in Table VI are the main blockers. Both are fixable in a revision if the authors can run the controlled ablation and correct/justify the synthetic-data results. The journal fit is otherwise appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on MiraMo. The genuinely new piece is the temporal linear attention with the cosine kernel (Eq. 8-9) that works with RoPE, plus DCTInit as an alternative to FFT noise refinement. The MRL part is carried over from Cinemo but applied to a new backbone, and the ablation in Table V does support that it helps. I also credit the synthetic-data post-training study: it's an interesting question, and the authors openly list limitations and disclaim the visual comparisons.\n\nThe soft spots are real. First, the paper's central claim—linear attention preserves generation quality while cutting cost—is never tested with a controlled swap. Fig. 3 shows throughput/memory and a +4.76% VBench number, but the vanilla baseline is not described, and no training-time comparison is reported. The claim that frame-dimension temporal attention outperforms full 3D temporal attention is also stated without quantitative support. These are load-bearing, because the entire method is framed around linear attention. If linear attention doesn't actually hold quality, the paper reduces to MRL + DCTInit, which are independent of the backbone.\n\nSecond, Section IV-C contradicts Table VI. The text says the synthetic dataset gives over a 3% improvement, but Table VI shows the real data wins on every listed dimension (e.g., Multiple Object 52.91 vs 45.50). Either the table labels are swapped or the claim is wrong. That needs to be fixed before this can be trusted.\n\nThird, the own-numbers issue: on UCF-101, FVD is 201.9 vs Cinemo's 168.2, and on MSR-VTT it's 93.9 vs 93.5. That doesn't match 'comparable' if you read closely, and it weakens the 'appearance consistency' headline, though IS and FID are better.\n\nOverall, this is a solid engineering paper with a clear gap in verification and one internal inconsistency. I wouldn't desk-reject it; a good referee could get it into shape. The reader's conditional verdict is fair, though I'd emphasize the Table VI issue over the missing ablation—the ablation is a question, but the table is a factual contradiction.\n\nRecommend: send to peer review; ask for the linear-vs-vanilla controlled experiment and a corrected Table VI.","headline":"Solid engineering paper with a useful linear-attention recipe, but the central quality-parity claim lacks a controlled ablation and the synthetic-data section has a clear internal contradiction.","tokens_in":24404,"tokens_out":3434,"would_cite":true,"duration_ms":32697,"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":"MiraMo claims image animation can be made faster and more consistent by replacing quadratic attention with a RoPE-compatible cosine linear attention, learning motion residuals rather than frames, and initializing inference noise with low-fr","keywords":["image animation","image-to-video generation","linear attention","diffusion transformer","flow matching","motion residual learning","DCT noise refinement","dynamics control"],"falsifier":"Replace MiraMo's temporal linear attention with vanilla quadratic attention under identical training and inference; if the full-attention variant scores strictly better on temporal consistency and motion smoothness at comparable compute on the VBench I2V suite, the paper's core efficiency-quality tradeoff claim is refuted.","tokens_in":23327,"feed_emoji":"🎬","tokens_out":7495,"duration_ms":68575,"temperature":0.7,"pith_summary":"MiraMo is an image-animation model built on a linear-attention diffusion transformer. The paper claims that three design choices let it beat U-Net-based animators on appearance consistency and motion smoothness while running faster: a cosine-kernel temporal attention that keeps O(N) cost, training on latent motion residuals rather than frames themselves, and a DCT-based noise initialization at inference. The authors also report that post-training on synthetic videos works better than on a small curated real dataset, producing higher VBench scores. If these claims hold, transformer-based image animation can be both cheaper and steadier without sacrificing prompt alignment or dynamics.","feed_headline":"Linear attention makes image animation consistent and cheap","feed_subtitle":"MiraMo's linear attention and motion-residual learning cut GPU cost while improving consistency.","key_machinery":"The core mechanism is a RoPE-compatible cosine linear attention (Eq. 8–9), where the similarity between query $q_i$ and key $k_j$ is $1 + (q_i/\\|q_i\\|)(k_j/\\|k_j\\|)^T$, allowing shared sums over keys/values to be computed once per sequence and reducing attention cost from $O(N^2)$ to $O(N)$. This attention is inserted as a temporal-attention layer in each SANA transformer block to capture frame-to-frame relationships. Two supporting mechanisms carry the consistency claim: motion residual learning, which subtracts the first frame's latent from all subsequent latents so the model predicts only inter-frame differences, and DCTInit, which replaces the low-frequency part of the inference noise wi","core_discovery":"The central claim is that image animation does not need to trade quality for speed. By adapting the SANA linear-attention text-to-image architecture into a text-to-video model with a new temporal linear attention (Eqs. 8–9) that stays compatible with RoPE, then fine-tuning it to predict latent motion residuals (differences from the first frame) rather than full frames, and finally refining the inference-time noise with low-frequency DCT components of the input image, MiraMo produces videos that hold the input image's appearance across frames, move smoothly, and follow the prompt, all at O(N) attention cost with a 0.74B-parameter model. In the reported comparisons it achieves the best or near","pith_inferences":["Editorial inference: The same cosine-kernel attention could be dropped into the spatial attention of existing video diffusion transformers, since RoPE compatibility already exists there, potentially cutting memory for high-resolution and long-video generation.","Editorial inference: The paper's residual representation suggests a general recipe for training-free motion transfer: invert motion residuals from any source video and apply them to a modified first frame, a protocol that could extend to other animators beyond MiraMo.","Editorial inference: DCTInit's rationale predicts that other energy-compacting transforms, such as wavelets, would show similar or better suppression of abrupt motion; a simple ablation could test whether DCT is optimal or merely sufficient.","Editorial inference: If synthetic post-training data routinely beats small curated real data, it suggests that SFT gains depend more on distribution coverage and prompt diversity than on pixel-level realism, a hypothesis the paper does not directly test."],"forward_implications":["Image animation can be built on linear-attention transformers, closing the architectural gap with text-to-video models at a fraction of the compute.","Motion residual learning becomes a viable alternative to direct frame prediction for appearance preservation in I2V generation.","DCT-based noise initialization can replace FFT-based FreeInit-style injection to avoid color drift and abrupt motion artifacts.","Synthetic video data can substitute for carefully curated real data in post-training, lowering the cost of high-quality fine-tuning.","The residual representation supports motion transfer and video editing via flow inversion, as the paper's applications demonstrate."],"supporting_citations":[{"why":"Supplies the base linear-attention text-to-image architecture (SANA) that MiraMo extends with a temporal module.","marker":"[38]"},{"why":"Provides RoPE positional encoding, whose compatibility motivates the cosine-kernel linear attention in Eqs. 8–9.","marker":"[33]"},{"why":"The prior conference version (Cinemo), which MiraMo extends and uses as the closest comparison baseline for I2V quality.","marker":"[37]"},{"why":"Supplies the FFT-based FreeInit noise-initialization method that DCTInit replaces and improves upon.","marker":"[35]"},{"why":"A consistency-focused I2V baseline that also uses FFT-based low-frequency injection, serving as the key comparison for appearance consistency.","marker":"[18]"},{"why":"Provides the inspiration for dynamics-degree control via frame-similarity buckets.","marker":"[15]"},{"why":"One of the two T2V models used to synthesize the post-training video dataset.","marker":"[27]"},{"why":"The other T2V model used to synthesize the post-training video dataset.","marker":"[28]"},{"why":"Supplies the Vimeo25M pretraining subset and acts as a text-to-video baseline in VBench comparisons.","marker":"[9]"}],"fun_headline_variants":["O(N) attention powers MiraMo's smooth, consistent animations","MiraMo: linear attention for cheap, consistent image animation","Linear attention trims cost, keeps animation consistent","Motion residuals + linear attention = smoother animation at lower cost","MiraMo: faster image animation without sacrificing consistency"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The whole efficiency-without-quality-loss claim rests on the premise that the cosine-kernel temporal linear attention captures video temporal relationships as well as full quadratic attention, a comparison the paper asserts but does not quantitatively demonstrate.","fun_headline_variants_meta":{"raw":{"variants":["O(N) attention powers MiraMo's smooth, consistent animations","MiraMo: linear attention for cheap, consistent image animation","Linear attention trims cost, keeps animation consistent","Motion residuals + linear attention = smoother animation at lower cost","MiraMo: faster image animation without sacrificing consistency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000317,"raw_usage":{"total_tokens":1645,"prompt_tokens":778,"completion_tokens":867,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":801}},"tokens_in":522,"tokens_out":867,"duration_ms":7076,"temperature":1.0,"reasoning_tokens":801,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:16:43.417666+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace MiraMo's temporal linear attention with vanilla quadratic attention under identical training and inference; if the full-attention variant scores strictly better on temporal consistency and motion smoothness at comparable compute on the VBench I2V suite, the paper's core efficiency-quality tradeoff claim is refuted.","supporting_citations":[],"review_version":1}