{"id":"aa12681d-381d-4292-b3ba-377614376090","arxiv_id":"2502.04363","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A training-free pipeline makes diffusion text-to-video generation run on an iPhone 15 Pro with quality close to GPU output, at the cost of slower generation.","lead":"This paper shows that pre-trained text-to-video diffusion models can run on a smartphone without retraining, using three efficiency tricks: skipping denoising steps, merging video frames in attention layers, and loading model pieces on demand. A smart generalist might read it to see whether on-device video generation is now practical on a consumer phone like the iPhone 15 Pro.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No experiment measures quality and latency for the same final configuration: Tab. 5's 638 s 'All' uses TDTM for all 30 steps, but Tab. 4 shows that setting cuts Dynamic Degree to 0.06 (vs 0.23 baseline), and Tab. 2's 0.27 Dynamic Degree is inconsistent with it.","rationale":"The reader's CONDITIONAL verdict is appropriate. The feasibility result—a pre-trained rectified-flow model can be executed on an iPhone 15 Pro with the proposed techniques—is supported by real measurements and open-source code. The main gap is not LPL correctness in the abstract; LPL's own ablation in Tab. 3 shows VBench scores essentially unchanged at 16/30 steps and in the dynamic setting. The more immediately load-bearing issue is that the paper never reports quality metrics for the exact 'All' configuration whose 638 s latency is highlighted. Tab. 4 indicates that full TDTM is precisely the component that damages Dynamic Degree, and Tab. 2's Dynamic Degree of 0.27 is inconsistent with full TDTM. This means the central claim 'high-quality videos at 638 s' may pair a quality measurement from one configuration with a latency measurement from another. A single experiment measuring the complete pipeline on both axes would settle this. Because the core feasibility and the general direction of the contributions are credible, the verdict remains CONDITIONAL, not REJECT.","tokens_in":827,"tokens_out":731,"duration_ms":158018,"concrete_test":"Run the exact 'All' pipeline from Sec. 7.4 (LPL stop at step 15, TDTM applied to all 30 steps, CI-DL enabled) on an iPhone 15 Pro over the same VBench prompts used for Tab. 2/Tab. 6, and report all seven VBench metrics plus end-to-end latency. If Dynamic Degree lands near 0.06-0.12 while latency is ~638 s, the 'comparable quality at 638 s' formulation must be revised to name the actual configuration; if metrics match Tab. 2 at 638 s, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires simultaneous evidence of high quality and low latency from the complete pipeline. The paper does not provide this. Sec. 7.4 defines the 'All' latency configuration as LPL activated at the 15th denoising step and TDTM applied throughout all steps; Tab. 5 reports 638.09 s end-to-end at 256x256. But the VBench table for that configuration is never given. Tab. 4 shows TDTM 30/30 on VBench reduces Dynamic Degree from 0.23 to 0.06 and TDTM 15/30 to 0.12, while Tab. 2 reports On-device Sora Dynamic Degree 0.27, higher than Open-Sora's 0.21. These numbers cannot all describe the same configuration. The quality table and the latency table therefore appear to be generated under different LPL/TDTM settings, so the headline '638 s with comparable quality' is not directly established. The reader's LPL-straightness concern is legitimate but secondary: Tab. 3 shows LPL alone preserves VBench scores well, so the unresolved gap is the joint configuration, not LPL in isolation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents On-device Sora, a training-free framework for running diffusion-based text-to-video generation on mobile devices. The system combines three techniques: Linear Proportional Leap (LPL), which halts denoising early and replaces remaining steps with a scaled Euler leap; Temporal Dimension Token Merging (TDTM), which merges consecutive temporal tokens in attention layers; and Concurrent Inference with Dynamic Loading (CI-DL), which partitions models for memory-constrained execution. The authors implement the system on an iPhone 15 Pro using Open-Sora as the backbone and evaluate quality with VBench, reporting that On-device Sora achieves video quality comparable to Open-Sora on an NVIDIA A6000 while cutting end-to-end latency to about 638 seconds at 256x256 resolution. The paper also includes an ablation study of Pyramidal Flow with LPL. The central claim is that these techniques together enable efficient, high-quality on-device text-to-video generation without retraining.","tokens_in":22145,"tokens_out":3259,"duration_ms":28994,"significance":"If the central claim is substantiated, this is a useful and timely contribution: it avoids the large GPU training budgets required by prior on-device video generation methods such as SnapGen-V, and it is accompanied by an open-source implementation and real-device latency measurements. The system is concretely implemented on commodity smartphone hardware, and the three techniques are clearly separable and tested in isolation. However, the paper's headline result—high quality and low latency from the complete pipeline—is not directly evidenced: no VBench evaluation is reported for the exact configuration used in the latency table, and the quality metrics are saturated around 0.99 without error bars or significance tests. The LPL straightness assumption is also only validated indirectly. These gaps currently prevent a full endorsement of the claimed result.","major_comments":[{"comment":"The central claim of \"comparable quality at 638 s\" is not established because no quality measurement is reported for the exact 'All' configuration that produces the 638-second latency. Section 7.4 states that 'All' uses LPL activated at the 15th denoising step and TDTM applied throughout all steps. However, Table 4 shows that applying TDTM for 30/30 steps on VBench reduces Dynamic Degree to 0.06 (and 15/30 to 0.12), while Table 2 reports On-device Sora's Dynamic Degree as 0.27; these cannot describe the same configuration. The quality table and the latency table therefore appear to have been generated under different LPL/TDTM settings. Please provide VBench scores for the exact 'All' configuration used in Table 5, ideally with per-prompt error bars, and clarify which settings were used for each row of Table 2.","section":"§7.4, Table 5 vs. §7.3, Table 4"},{"comment":"The quality comparison relies on heavily saturated metrics: Temporal Flickering, Motion Smoothness, Subject Consistency, and Background Consistency are all around 0.97–0.99 for both methods, and the reported differences (e.g., a 0.03 average drop in frame-wise quality) are within the noise of a single VBench run. No error bars, per-category standard deviations, or significance tests are provided, so the claim that quality is \"nearly equivalent\" is statistically unsupported. Please report confidence intervals (e.g., bootstrap over the 800 prompts) or per-prompt distributions, and consider supplementing with a less saturated metric such as FVD or human evaluation.","section":"§7.1, Table 2"},{"comment":"The LPL method rests on the assumption that the reverse-process trajectory is approximately straight after a certain number of steps, but this is validated only by cosine similarities between adjacent drifts on the same models; the paper itself concedes in App. C that LPL is inapplicable to non-rectified-flow models such as CogVideoX. The runtime stopping rule also involves hand-tuned parameters (a 1e-4 tolerance and a 50% minimum step fraction). Please provide a sensitivity analysis of these thresholds and, more importantly, a direct downstream evaluation (e.g., VBench or FVD) comparing LPL's output at early stop against the full-trajectory output, to confirm that the straightness assumption preserves quality rather than merely correlating with drift similarity.","section":"§4.2, Fig. 4, App. C"}],"minor_comments":[{"comment":"The paper contains two naming typos: \"Conference Inference and Dynamic Loading\" in Section 1 should be \"Concurrent Inference and Dynamic Loading,\" and \"Linear Professional Leap\" in Section 9 should be \"Linear Proportional Leap.\"","section":"§1, §9"},{"comment":"The text repeatedly uses \"retrained state\" (e.g., \"loaded in a retrained state\" and \"with a retrained state\"); this should be \"retained state.\"","section":"App. A.1"},{"comment":"The baseline latency of 1768.32 seconds for STDiT comes from Table 1, which reports the 50-step configuration, whereas all other rows in Table 5 presumably use a 30-step schedule; please state the number of denoising steps for each row to make the comparison fair and reproducible.","section":"Table 5"},{"comment":"Tables 3 and 4 report speedups relative to a 30/30 baseline, but Table 2's \"On-device Sora\" row does not state which LPL/TDTM configuration was used; please specify this so the reader can map the headline quality numbers to the actual pipeline.","section":"§7.2, §7.3"},{"comment":"The introduction claims broad applicability to \"pre-trained video generative models, e.g., Open-Sora, Pyramidal Flow,\" but only LPL is experimented on Pyramidal Flow; TDTM and CI-DL are only evaluated with Open-Sora. Please clarify the scope of the generality claim.","section":"§1, §8"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to be a good fit for the venue if the authors can close the configuration gap: a single experiment that runs VBench on the exact 'All' configuration (LPL at step 15 + TDTM throughout) and shows acceptable quality would directly support the headline claim. My main concern is that the current evidence for 'comparable to high-end GPUs' is indirect, since the quality tables and the latency table describe different settings. The saturated metrics and absence of error bars also weaken the quality claim, but those are fixable with additional analysis. The 'first training-free on-device text-to-video' claim should be checked against any contemporaneous work, but I did not find an instance in the references that would invalidate it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read of arXiv:2502.04363. The genuinely new thing here is the demonstration: a pre-trained rectified-flow video model (Open-Sora) running training-free on an iPhone 15 Pro, with open-source code and concrete latency numbers. That alone is worth knowing about. The three components—LPL, TDTM, CI-DL—are adaptations of known ideas, but the combination is new and the system-level result is plausible.\n\nHowever, the central claim—'comparable quality at 638 seconds'—is not actually measured. Sec. 7.4 defines the 'All' latency configuration as LPL at the 15th step and TDTM on all 30 steps, and reports 638 s end-to-end at 256x256. But no VBench table is given for that configuration. Tab. 4 shows TDTM 30/30 drops Dynamic Degree from 0.23 to 0.06, Tab. 3 shows LPL 16/30 gives 0.18, and Tab. 2 reports On-device Sora's Dynamic Degree as 0.27. Those numbers cannot all describe the same setup, so the quality/latency trade-off of the final pipeline is unestablished. That is the load-bearing soft spot.\n\nSecondary issues: the VBench scores are saturated near 0.99, there are no error bars or significance tests, and the LPL thresholds (10^-4 tolerance, 50% minimum) are hand-tuned with no sensitivity analysis. The straightness assumption for LPL is only validated on the same models, but the paper is honest about that limitation in App. C.\n\nCredit where it is due: the ablation structure is clear, the implementation details (CoreML, int8 T5, float32 STDiT) are concrete, and App. C is candid about latency still being minutes rather than GPU-like. The code being public is a real plus.\n\nI would send this to peer review. The core feasibility result is real, and the paper is the first to show it, but the authors need to either report VBench scores for the exact 'All' configuration or stop claiming parity. This is for mobile-AI systems people, not core generative modeling researchers. A serious referee can handle it, provided they push on the configuration ambiguity.","headline":"A real first demonstration of training-free on-device diffusion text-to-video, but the headline quality/latency claim is not backed by a single joint measurement.","tokens_in":22737,"tokens_out":2559,"would_cite":true,"duration_ms":20333,"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":"The paper claims that three training-free techniques — Linear Proportional Leap, Temporal Dimension Token Merging, and Concurrent Inference with Dynamic Loading — let a pre-trained text-to-video diffusion model run on an iPhone 15 Pro and…","keywords":["text-to-video generation","on-device inference","diffusion models","rectified flow","token merging","training-free acceleration","mobile devices","model partitioning"],"falsifier":"Apply Linear Proportional Leap to a flow-matching video model whose consecutive-drift cosine similarity stays below 0.95 through the second half of its schedule (CogVideoX, which the paper itself names as non-straight, is a direct candidate), and compare VBench frame-wise quality against full-step generation: a drop in imaging quality or aesthetic quality larger than the roughly 0.03 the paper reports on Open-Sora would show that LPL depends on the straightness premise, not on the leap formula alone. A sharper probe is to print the cosine-similarity curve for a 50-step rectified-flow generation and check whether it is monotone non-decreasing past step 25; if it dips, the dynamic stopping rule's 10⁻⁴ improvement tolerance can halt too early or too late, and the reported 1.94× speedup is not generally reproducible.","tokens_in":21691,"feed_emoji":"📱","tokens_out":9888,"duration_ms":79449,"temperature":0.7,"pith_summary":"The paper claims that a pre-trained diffusion text-to-video model can run directly on a smartphone, with no re-training, distillation, or compression, and still generate videos nearly as good as those from a high-end GPU. The authors attribute this to three techniques applied to the existing Open-Sora model: Linear Proportional Leap, which skips most denoising steps by moving in one scaled jump along the model's nearly straight rectified-flow trajectory; Temporal Dimension Token Merging, which averages adjacent video-frame tokens so attention processes half as many tokens; and Concurrent Inference with Dynamic Loading, which splits the large text and video transformers into blocks and overlaps CPU loading with GPU inference. On an iPhone 15 Pro, the combined methods reduce the denoising process from about 1,768 seconds to 293–454 seconds (depending on resolution) and end-to-end latency to roughly 638 seconds, while VBench quality scores stay within about 0.01 of Open-Sora running on an NVIDIA A6000. The paper frames this as the first training-free route to on-device text-to-video generation, removing the GPU-intensive model re-training that earlier on-device video generators require.","feed_headline":"No training: video diffusion runs on a phone at near-GPU quality","feed_subtitle":"Leaping over denoising steps and merging frame tokens cuts Open-Sora's runtime on an iPhone to ~638 seconds.","key_machinery":"The load-bearing object is the rectified-flow drift $v(P_k, t_k)$ and the leap identity $z_k = z_n + v(P_{n+1}, t_{n+1}) t_{n+1}$, which says that once the trajectory is straight, the accumulated update of all remaining denoising steps equals one scaled Euler step taken with the next drift. LPL uses the cosine similarity between $v(P_n, t_n)$ and $v(P_{n-1}, t_{n-1})$ as a runtime straightness probe, stopping the denoising loop when similarity stops improving. TDTM's machinery is a merge/unmerge pair on the temporal axis: input tokens of shape $[B, ST, C]$ become $[B, ST/2, C]$ by averaging consecutive frames, and the attention output is replicated back; this exploits the fact that adjacent frames in a video are near-duplicates. CI-DL's machinery is a block pipeline: the transformer is cut into segments, the CPU loads segment $i+1$ while the GPU executes segment $i$, and a dynamic number of segments stay resident in memory across denoising iterations, so the same blocks are not reloaded each step.","core_discovery":"On-device Sora's central claim is that the three bottlenecks of mobile video diffusion — too many denoising steps, quadratic attention cost, and model weights that exceed device memory — can each be bypassed without changing any weights. In the rectified-flow formulation used by Open-Sora, the denoising trajectory becomes nearly straight after the early steps; LPL detects this straightness from the cosine similarity of consecutive drifts and replaces all remaining steps with the single update $z_k = z_n + v(P_{n+1}, t_{n+1}) t_{n+1}$, cutting steps from 30 to as few as 16 with VBench scores essentially unchanged. TDTM merges pairs of adjacent tokens along the temporal dimension, computes attention on half the tokens, and replicates the output to restore the original shape, reducing self-attention by a factor of four and cross-attention by half. CI-DL partitions T5 and STDiT into blocks that are loaded by the CPU while the GPU executes the previous block, retaining a memory-fit subset across denoising iterations. The experimental section reports that with all three applied, Open-Sora on an iPhone 15 Pro produces 256×256 videos whose temporal-quality scores are 0.96–0.99 versus 0.97–0.99 for the same model on an A6000, with an average frame-wise drop of 0.03 and a 0.06 improvement in dynamic degree.","pith_inferences":["The leap identity in LPL is a zero-training shortcut that effectively distills the model's learned drift into a near-one-step update in the straight region; the same reasoning should apply to rectified-flow image diffusion, where even larger step reductions are plausible because the trajectory straightness assumption is not video-specific.","CI-DL's block pipeline is a general technique for any transformer whose weights exceed device memory, not just T5 and STDiT; the paper's latency formula predicts that gains grow with the number of blocks, so other over-capacity models should see comparable speedups.","A natural extension the authors do not explore is merging both spatial and temporal tokens: combining TDTM with spatial token merging would cut self-attention by up to 16×, but the dynamic-degree drop observed with full temporal merging suggests such aggressive merging would need selective scheduling to stay usable.","The paper runs only on the GPU and leaves the device's Neural Engine idle; if the device vendor's software eventually supports the required diffusion operators, offloading T5 or the VAE to a dedicated neural processor could cut the remaining ~500 seconds of end-to-end latency substantially."],"forward_implications":["Any pre-trained rectified-flow text-to-video model becomes deployable on a phone without re-training, eliminating the GPU-heavy distillation and compression pipelines that earlier on-device video generators require.","LPL's step reduction and TDTM's attention reduction are multiplicative, and both stack with CI-DL's memory pipelining, so the combined 6x denoising speedup is not the sum but the product of the individual gains.","Because the methods touch only the sampling schedule and token layout, they transfer to more compact DiT-based video generators as those models appear, making on-device generation progressively cheaper without new optimization work.","Deploying the full pre-trained model on-device means prompts and videos never need to leave the phone, which makes private, offline text-to-video possible on commodity hardware.","TDTM's motion cost matters: merging all 30 denoising steps drops dynamic degree from 0.23 to 0.06 on VBench, so applications wanting lively videos should merge only the first half of steps and sacrifice part of the speedup."],"supporting_citations":[{"why":"Open-Sora — the pre-trained backbone whose STDiT architecture and rectified-flow sampling schedule LPL and TDTM modify.","marker":"[83]"},{"why":"Rectified Flow — provides the straight-trajectory drift formulation and the Euler-step notation that LPL's leap identity relies on.","marker":"[38]"},{"why":"VBench — the benchmark whose temporal and frame-wise scores ground the claim that on-device output matches an A6000's.","marker":"[28]"},{"why":"Token Merging — the merge/unmerge attention acceleration that TDTM adapts to the temporal dimension.","marker":"[7]"},{"why":"T5 — the 18 GB text encoder that motivates CI-DL; its size defines the memory constraint CI-DL must overcome.","marker":"[48]"},{"why":"SnapGen-V — the training-required on-device video generator used as the baseline that On-device Sora's training-free claim is contrasted with.","marker":"[70]"},{"why":"Pyramidal Flow — a second rectified-flow text-to-video model used to show that LPL transfers beyond Open-Sora.","marker":"[29]"}],"fun_headline_variants":["No training, no recompression: diffusion video on a phone","Skip steps and merge tokens to get diffusion video on iPhone","Phone video gen: cut denoising steps and attention cost","Diffusion video on a phone at near-GPU quality, training-free","From 30 steps to 16: phone generates video like a GPU"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the later stages of the rectified-flow reverse process are approximately straight, so one scaled Euler step with the final drift can replace all remaining denoising steps without visible quality loss; if the trajectory curves, LPL's leap overshoots and quality degrades.","fun_headline_variants_meta":{"raw":{"variants":["No training, no recompression: diffusion video on a phone","Skip steps and merge tokens to get diffusion video on iPhone","Phone video gen: cut denoising steps and attention cost","Diffusion video on a phone at near-GPU quality, training-free","From 30 steps to 16: phone generates video like a GPU"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000868,"raw_usage":{"total_tokens":3851,"prompt_tokens":1125,"completion_tokens":2726,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":741,"completion_tokens_details":{"reasoning_tokens":2636}},"tokens_in":741,"tokens_out":2726,"duration_ms":16290,"temperature":1.0,"reasoning_tokens":2636,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T10:40:49.043917+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply Linear Proportional Leap to a flow-matching video model whose consecutive-drift cosine similarity stays below 0.95 through the second half of its schedule (CogVideoX, which the paper itself names as non-straight, is a direct candidate), and compare VBench frame-wise quality against full-step generation: a drop in imaging quality or aesthetic quality larger than the roughly 0.03 the paper reports on Open-Sora would show that LPL depends on the straightness premise, not on the leap formula alone. A sharper probe is to print the cosine-similarity curve for a 50-step rectified-flow generation and check whether it is monotone non-decreasing past step 25; if it dips, the dynamic stopping rule's 10⁻⁴ improvement tolerance can halt too early or too late, and the reported 1.94× speedup is not generally reproducible.","supporting_citations":[{"cited_title":"Open-sora: Democratizing efficient video production for all, 2024","cited_arxiv_id":null,"evidence_quote":"Open-Sora — the pre-trained backbone whose STDiT architecture and rectified-flow sampling schedule LPL and TDTM modify."},{"cited_title":"Vbench: Comprehensive bench- mark suite for video generative models","cited_arxiv_id":null,"evidence_quote":"VBench — the benchmark whose temporal and frame-wise scores ground the claim that on-device output matches an A6000's."},{"cited_title":"Exploring the limits of transfer learning with a unified text-to-text transformer","cited_arxiv_id":null,"evidence_quote":"T5 — the 18 GB text encoder that motivates CI-DL; its size defines the memory constraint CI-DL must overcome."},{"cited_title":"SnapGen-V: Generating a Five-Second Video within Five Seconds on a Mobile Device","cited_arxiv_id":"2412.10494","evidence_quote":"SnapGen-V — the training-required on-device video generator used as the baseline that On-device Sora's training-free claim is contrasted with."}],"review_version":1}