{"id":"bcc387ab-5ba9-4d8d-9a1b-3c25b43a44a3","arxiv_id":"1908.07985","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MobiSR selects two compressed super-resolution models, maps them onto CPU, GPU and DSP, and uses a total-variation threshold to send easy patches to the accurate model and hard patches to the fast model, cutting latency by roughly 2x at equal average quality.","lead":"Super-resolution on a phone is slow, so MobiSR splits each image into easy and hard patches and sends them to different neural networks on different phone processors. The paper reports 2x to 5x faster upscaling than running one large network, with image quality kept inside a user-set error budget.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline speedups (2.13x/4.79x) come from Table 3 averaged over TV thresholds without enforcing the Eq. (12) PSNR constraint; the equal-quality evaluation in Fig. 10 shows much smaller gains, so the central performance claim is not supported as stated.","rationale":"The reader's verdict of CONDITIONAL remains appropriate, but the decisive weakness is slightly different from the reader's formal weakest_assumption. The reader identifies the TV patch-level proxy as the load-bearing premise; that is a genuine risk, and the paper's evidence for it is image-level (Figs. 3 and 4) rather than patch-level, with no clearly held-out calibration protocol. However, the more immediately checkable and numerically decisive issue is that the abstract's speedup figures are taken from Table 3, which averages over TV thresholds without applying the PSNR constraint of Eq. (12). The paper's own fixed-quality evaluation (Fig. 10) shows much smaller speedups, so the central claim as worded is not supported by the reported measurements. This is a reporting/validation gap rather than a fundamental flaw: if the authors re-report all headline speedups at equal PSNR and clarify the held-out selection of TV_thr, the framework could still be valid. Hence I would keep the CONDITIONAL verdict and require those corrections before acceptance.","tokens_in":20828,"tokens_out":7953,"duration_ms":77014,"concrete_test":"Recompute Table 3 under the Eq. (12) constraint: for each dataset, model pair and each epsilon_max used in Fig. 10, pick the TV_thr that maximizes measured speedup while satisfying measured PSNR(I, mref,0) - PSNR(I,(m1,m2),TV_thr) <= epsilon_max; average these constrained speedups across datasets and report the PSNR drop at the threshold producing the 2.13x/4.79x numbers. If constrained averages match Fig. 10 (~1.3-1.9x) rather than the abstract values, the headline must be restated with equal-quality speedups.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reported 2.13x/4.79x averages are not computed under the quality constraint that the system is claimed to enforce. Table 3 gives speedups of (m1,m2) pairs over mref on CPU/GPU/DSP averaged over TV_thr values, but it does not report the PSNR drop at each threshold. Because the speedup ranges extend up to 5-7x only when TV_thr is low—so that most patches go to the quantized DSP model—averaging over the threshold axis mixes operating points that violate the user's PSNR budget with those that satisfy it. In contrast, Fig. 10, the one evaluation that holds PSNR fixed, reports maximum speedups of 1.47x/1.78x/1.94x/1.29x on Set5/Set14/B100/Urban100. The abstract's 2.13x and 4.79x therefore cannot be verified from an equal-quality comparison as presented. The TV proxy concern (Section 3.3, Eq. (9), Fig. 4) is real but secondary: even with a weak proxy, the optimizer of Eq. (12) could calibrate a threshold on a held-out set; no amount of calibration makes an unconstrained speedup average into a quality-preserving claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents MobiSR, a framework for on-device super-resolution that combines two compressed SR models with a runtime scheduler. The scheduler uses the total variation (TV) of each image patch as a proxy for upscaling difficulty, dispatching easy patches to a larger, more accurate model on the CPU/GPU and hard patches to a compact, quantized model on the DSP, with load balancing. Offline, the framework explores model compression transformations and selects a model pair and a TV threshold to minimize latency subject to a user-specified PSNR drop constraint, formalized as the optimization problem in Eq. (12). The authors report average speedups of 2.13x over optimized parallel difficulty-unaware mappings and 4.79x over single-engine implementations on a Snapdragon 845 board, with additional measurements across Set5, Set14, B100, and Urban100.","tokens_in":21127,"tokens_out":2150,"duration_ms":21661,"significance":"The paper addresses a practically relevant systems problem: efficient on-device super-resolution under a quality constraint. Its strengths are concrete: the evaluation uses real measurements on the Snapdragon 845 via SNPE, the design-space exploration with compression transformations is clearly described, and Fig. 10 honestly reports speedups as a function of allowed PSNR degradation. If the central speedup claim is properly connected to the quality constraint, the work would be a useful contribution to mobile systems and embedded deep learning. However, the headline numbers in the abstract and Table 3 are not computed under the equal-quality protocol used in Fig. 10, and the TV-based difficulty proxy is validated only at the image level while being applied at the patch level. These issues affect the central claim and require revision.","major_comments":[{"comment":"The headline speedups of 2.13x and 4.79x are not supported by the equal-quality evaluation. Table 3 reports speedup ranges and averages over TV threshold values without reporting the PSNR drop at each operating point, so the averages mix configurations that violate the Eq. (12) constraint with those that satisfy it. In contrast, Fig. 10, which compares systems at the same PSNR drop, shows maximum speedups of 47%, 78%, 94%, and 29% on Set5, Set14, B100, and Urban100, respectively. The abstract should either report speedups under the PSNR constraint or clearly label the Table 3 numbers as unconstrained operating points.","section":"Abstract; Table 3; Section 4.4, Fig. 10"},{"comment":"The TV-based difficulty criterion is validated at the image level on DIV2K, but the scheduler operates on individual patches. The paper does not provide patch-level evidence that high-TV patches are indeed nearly equally hard for the large and compact models, nor does it describe a held-out calibration protocol for selecting TV_thr. Because the scheduler's quality guarantee depends on this transfer, the authors should add a patch-level analysis (e.g., PSNR difference vs. TV for patches) and an explicit calibration/validation split for threshold selection.","section":"Section 3.3, Eq. (9), Fig. 4"},{"comment":"The performance model in Eq. (11) assumes that patches are assigned to m1 or m2 solely by the TV threshold, but Algorithm 1 allows hard patches to be processed by m1 when the DSP is oversubscribed. This means the analytic latency estimate can understate the running time by assuming more DSP utilization than the scheduler actually provides. The authors should either incorporate the load-balancing policy into the performance model or justify that the discrepancy is negligible for the evaluated configurations.","section":"Section 3.4, Eq. (11); Algorithm 1"}],"minor_comments":[{"comment":"The caption reads 'SDM854' but the platform is Snapdragon 845 (SDM845); this typo should be corrected.","section":"Fig. 9 caption"},{"comment":"The x-axis of Fig. 10 is described only as 'error degradation'; the exact PSNR drop values used for each interval are not stated in the text, making it hard to assess which operating points are of practical interest.","section":"Section 4.4, Fig. 10"},{"comment":"The table would be easier to interpret if it also reported the PSNR drop of each configuration at the listed speedups, or if the speedups were restricted to configurations satisfying the Eq. (12) constraint.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read: the system is real and well-engineered, and the headline numbers are not the equal-quality numbers. The core idea—a two-model heterogeneous schedule (accurate model on CPU/GPU, compact model on DSP) with a TV-based patch difficulty predictor—is a legitimate contribution. The design-space search over compressed models is coherent, the Eq. (12) optimization is properly formulated, and the evaluation is genuine: SNPE measurements on a Snapdragon 845, four standard SR datasets, latency and PSNR both reported. The image-level TV correlation in Figs. 3–5 is reasonable evidence for the proxy, and the paper is honest about the scheduler's load-balancing behavior.\n\nThe soft spot is the performance claim. The 2.13x/4.79x averages in Table 3 and the abstract are computed without enforcing the PSNR constraint that Eq. (12) is supposed to enforce. They are averaged over TV thresholds, including operating points that violate the user's error budget. When the PSNR budget is actually held fixed (Fig. 10), the speedups drop to at most 1.47x on Set5, 1.78x on Set14, 1.94x on B100, and 1.29x on Urban100. That is still a useful gain—especially for a fully on-device system—but it is not 4.79x. A revised abstract should lead with the equal-quality numbers.\n\nTwo smaller issues. First, the TV threshold selection is not clearly held out from the test benchmarks; the paper mentions a calibration set but never shows the calibration split or the sensitivity of the chosen threshold. Second, the baselines are all variants of their own models in single-engine or unconstrained parallel mode. They do not compare against CARN-M or FSRCNN, which are the relevant efficient SR baselines. The citation pattern otherwise looks solid, and the related-work positioning against cascades and early-exit classifiers is fair.\n\nBottom line: this deserves a serious referee and probably acceptance after revision. The framework is sound, the evaluation is extensive, and the central mechanism holds up. What needs fixing is the reporting: separate the unconstrained speedup from the equal-quality speedup, provide the calibration protocol, and add at least one external efficient SR baseline. I would bring it to a reading group and cite it if I worked on mobile SR systems.","headline":"MobiSR is a genuinely useful mobile SR scheduling system, but the abstract's 2.13x/4.79x speedups are unconstrained averages; the equal-quality speedups in Fig. 10 are at most ~1.94x.","tokens_in":682,"tokens_out":845,"would_cite":true,"duration_ms":28242,"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":"MobiSR claims that a difficulty-aware scheduler using total variation can cut on-device super-resolution latency by 2.13x to 4.79x while staying within a user-specified PSNR drop.","keywords":["super-resolution","on-device inference","heterogeneous mobile processors","difficulty-aware scheduling","total variation","model compression","PSNR-latency trade-off","design space exploration"],"falsifier":"Run both candidate models on every patch of a held-out dataset that differs from the calibration set, compute per-patch PSNR against the full-precision large model, and count high-TV patches where the compact model's PSNR drop exceeds the user tolerance; if this fraction is large, the TV-threshold scheduler cannot meet its quality constraint, and the reported speedup at fixed quality would not transfer.","tokens_in":20653,"feed_emoji":"⚡","tokens_out":7417,"duration_ms":67196,"temperature":0.7,"pith_summary":"MobiSR sets out to show that on-device super-resolution can be made much faster by treating image patches differently: easy patches go to a large, accurate model, while hard patches go to a compact, low-precision model. The paper's central assertion is that a patch's total variation is a usable run-time proxy for its upscaling difficulty, so a simple threshold can dispatch patches without exceeding a user-set average PSNR budget. The framework automatically searches compressed variants of a supplied super-resolution network, prunes to Pareto-optimal models per compute engine, and tunes the threshold to minimize latency subject to the error tolerance. On a mobile SoC with CPU, GPU, and DSP, the resulting two-model designs achieve an average speedup of 2.13x over parallel difficulty-unaware mappings and 4.79x over single compute engine implementations. If right, this gives mobile zoom and image-upscaling applications a practical path to local processing with bounded quality loss.","feed_headline":"A patch scheduler cuts phone super-resolution latency by up to 4.79x","feed_subtitle":"Two models split image patches by difficulty, keeping average quality within the user's error budget.","key_machinery":"The load-bearing mechanism is the total-variation threshold scheduler. Total variation (TV) is the sum of absolute intensity differences between neighboring pixels in a patch, and the scheduler computes it per patch at run time; if $TV(p) \\le TV_{\\mathrm{thr}}$ the patch is sent to the larger model on the PSNR-preserving engines (CPU/GPU), otherwise to the compact model on the low-precision engine (DSP), with load balancing across engines and an overflow rule that lets hard patches fall back to the large model when the DSP is busy. The threshold $TV_{\\mathrm{thr}}$ is not hand-set: a calibration set is used to estimate the domain's TV range, and the optimization over model pairs and thresholds is driven by an analytical performance model that estimates per-engine latency from measured per-patch execution times. The design space itself is generated by model transformations—residual bottlenecks, group, depthwise-separable, and separable convolutions, inverted residuals, channel shuffle, and channel split—which are pruned per engine to the Pareto front.","core_discovery":"The discovery is that image patches differ systematically in how much a large super-resolution model outperforms a compact one, and that this difference tracks total variation. For low-TV (easy) patches the large model wins by a meaningful margin, so those patches should be sent to it; for high-TV (hard) patches the two models are nearly equally weak, so sending them to a fast compact model costs little quality and saves time. MobiSR turns this into a deployable system: it generates a family of compressed models from a user-supplied network, measures their latency on each compute engine, keeps only Pareto-optimal (model, engine) pairs, and then selects two models and a TV threshold that minimize estimated latency while keeping average PSNR within a user-specified drop. The quantitative claim is an average speedup of 2.13x over highly optimized parallel difficulty-unaware mappings and 4.79x over highly optimized single-engine implementations, measured on one mobile SoC.","pith_inferences":["Beyond the paper: the scheduler's quality guarantee is average PSNR, not per-patch or worst-case quality; applications sensitive to artifacts in individual patches would need a stricter dispatch rule.","Beyond the paper: the measured speedups come from one mobile SoC; on newer heterogeneous chips the relative latencies of CPU, GPU, and DSP change, so the optimal model pair and threshold would shift, though the search and scheduling procedure should transfer.","Beyond the paper: a natural testable extension is to make the threshold content-adaptive per image or to learn it from a small labeled calibration set instead of a single global value, which could improve the quality-latency frontier."],"forward_implications":["For a fixed error tolerance, an on-device SR system can run faster by splitting the image by difficulty rather than running one model on all patches.","The framework's reported speedups are 2.13x over parallel difficulty-unaware mappings and 4.79x over single-engine implementations, with larger speedups as the allowed PSNR drop increases.","Because the optimization is parametrized by the target SoC's compute engines, the same procedure applies to newer chips with neural accelerators, not only to CPU/GPU/DSP designs.","The Pareto-front pruning and the constraint that the fast model be more compact than the accurate model keep the design search small enough to be exhaustive."],"supporting_citations":[{"why":"Supplies the reference model mref whose compressed variants define the model space.","marker":"[62]"},{"why":"Provides the DIV2K training and validation sets used for training, calibration, and the TV-to-PSNR correlation plots.","marker":"[50]"},{"why":"Defines total variation, the metric used by the difficulty evaluation unit.","marker":"[44]"},{"why":"Inspires the ms2 compact model via channel split, depthwise convolution, and channel shuffle building blocks.","marker":"[39]"},{"why":"Provides the SRCNN baseline that MobiSR's reference model outperforms in the evaluation.","marker":"[6]"},{"why":"Provides the VDSR baseline used for comparison of on-device latency and quality.","marker":"[30]"},{"why":"Provides the FEQE-P mobile SR baseline from the PIRM challenge that mref is compared against.","marker":"[53]"},{"why":"Introduces depthwise separable convolutions used as a model transformation.","marker":"[19]"}],"fun_headline_variants":["TV-based patch routing speeds up on-device super-resolution by 4.79x","MobiSR: Heterogeneous processors speed up super-resolution by 4.79x","Split super-resolution patches by difficulty, gain 4.79x speedup","Patch difficulty router yields 2.13x to 4.79x faster mobile SR","Difficulty-aware SR scheduling: up to 4.79x speedup on phones"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Total variation of a patch is a reliable run-time proxy for how hard that patch is to upscale, and on hard patches the compact model's quality loss is small enough that the user's average PSNR budget is still met.","fun_headline_variants_meta":{"raw":{"variants":["TV-based patch routing speeds up on-device super-resolution by 4.79x","MobiSR: Heterogeneous processors speed up super-resolution by 4.79x","Split super-resolution patches by difficulty, gain 4.79x speedup","Patch difficulty router yields 2.13x to 4.79x faster mobile SR","Difficulty-aware SR scheduling: up to 4.79x speedup on phones"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001021,"raw_usage":{"total_tokens":4313,"prompt_tokens":958,"completion_tokens":3355,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":3246}},"tokens_in":574,"tokens_out":3355,"duration_ms":22038,"temperature":1.0,"reasoning_tokens":3246,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:52:44.062325+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run both candidate models on every patch of a held-out dataset that differs from the calibration set, compute per-patch PSNR against the full-precision large model, and count high-TV patches where the compact model's PSNR drop exceeds the user tolerance; if this fraction is large, the TV-threshold scheduler cannot meet its quality constraint, and the reported speedup at fixed quality would not transfer.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the reference model mref whose compressed variants define the model space."},{"cited_title":"Timofte et al","cited_arxiv_id":null,"evidence_quote":"Provides the DIV2K training and validation sets used for training, calibration, and the TV-to-PSNR correlation plots."},{"cited_title":"Rudin, Stanley Osher, and Emad Fatemi","cited_arxiv_id":null,"evidence_quote":"Defines total variation, the metric used by the difficulty evaluation unit."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Inspires the ms2 compact model via channel split, depthwise convolution, and channel shuffle building blocks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the SRCNN baseline that MobiSR's reference model outperforms in the evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the VDSR baseline used for comparison of on-device latency and quality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the FEQE-P mobile SR baseline from the PIRM challenge that mref is compared against."}],"review_version":1}