{"id":"00ed23ea-582c-4ff9-9e32-e311d87a6adc","arxiv_id":"2411.17106","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"PassionSR quantizes one-step diffusion super-resolution models to 6-8 bits via learnable quantizer boundaries, equivalent transformations, and a two-stage calibration, while keeping quality close to full precision.","lead":"This paper presents a post-training quantization method, PassionSR, that compresses a one-step diffusion super-resolution model to 8-bit or 6-bit weights and activations with small quality loss. It does this by simplifying the model to UNet and VAE, then tuning learnable quantizer boundaries and activation rescaling on a small calibration set.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"6-bit 'comparable visual quality' claim is contradicted by Table 2: LPIPS/DISTS degrade 25-75%, NIQE roughly doubles, and MANIQA/CLIP-IQA drop 31-55% relative to the FP backbone.","rationale":"The reader's weakest assumption is that quantization results on PassionSR-FP transfer to OSEDiff. I regard that as a generalizability concern, but the paper explicitly evaluates against PassionSR-FP and frames the contribution as a new simplified backbone; Table 2 shows the simplification is roughly neutral on most metrics, and even improves PSNR by 1.13 dB on RealSR. The more direct problem is that the central claim of 'comparable visual results at 8-bit and 6-bit' is contradicted by the paper's own main quantitative table, especially at 6-bit. The deltas on LPIPS, DISTS, NIQE, MANIQA, and CLIP-IQA are large enough that the phrase 'comparable' is unsupported without a threshold or perceptual validation. There is also an internal inconsistency in the DQC ablation: Table 4 shows LBQ+LET reaching 25.40 PSNR while LBQ+LET+DQC reaches 24.41, yet the text says DQC gives slight performance enhancement; this is secondary but reinforces the need for careful re-reporting. Credit is due for a well-motivated pipeline, extensive baseline comparisons, and a fair attempt to quantify compression, but the headline claim needs either stronger evidence or qualified wording. I would keep the reader's CONDITIONAL verdict pending the proposed bootstrap test and a corrected DQC description.","tokens_in":14274,"tokens_out":12286,"duration_ms":107333,"concrete_test":"Compute per-image deltas between 6-bit/8-bit PassionSR and PassionSR-FP for LPIPS, DISTS, NIQE, MANIQA, and CLIP-IQA on RealSR, DRealSR, and DIV2K val, then bootstrap the mean deltas with 10,000 resamples. Pre-register 'comparable' as the 95% confidence interval for every mean delta containing zero and a mean absolute relative change below 10% on all metrics. Based on the reported Table 2 means, 6-bit will clearly fail on LPIPS, DISTS, NIQE, MANIQA, and CLIP-IQA, and 8-bit will fail on NIQE. If the test confirms this, the abstract and conclusion must be revised to metric-specific, threshold-defined claims.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 2 reports 6-bit (W6A6) PassionSR results versus the FP backbone PassionSR-FP. On RealSR, LPIPS rises from 0.3339 to 0.4199 (+25.8%), DISTS from 0.1765 to 0.2592 (+46.9%), NIQE from 4.336 to 8.618 (+98.7%), while MANIQA falls from 0.4686 to 0.2131 (-54.5%) and CLIP-IQA from 0.7520 to 0.4612 (-38.7%). On DIV2K val, LPIPS rises from 0.2373 to 0.3440 (+45.0%), DISTS from 0.1185 to 0.2075 (+75.1%), NIQE from 3.573 to 7.039 (+97.0%), MANIQA falls from 0.4402 to 0.2267 (-48.5%), and CLIP-IQA from 0.6958 to 0.4802 (-31.0%). Even 8-bit shows NIQE degradation: 6.070 vs 4.336 on RealSR (+40.0%) and 4.424 vs 3.573 on DIV2K val (+23.8%). These deltas are large on both reference-based perceptual metrics (LPIPS, DISTS) and no-reference metrics. Since the abstract and conclusion state that both 8-bit and 6-bit obtain visual results comparable to full precision, and no threshold or user study is provided, the paper's central claim is not supported by its own main table. This does not invalidate the quantization method relative to baselines, but it does invalidate the headline claim as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PassionSR, a post-training quantization (PTQ) pipeline for one-step diffusion image super-resolution models. The method first simplifies OSEDiff into a UNet+VAE architecture (PassionSR-FP), then introduces a Learnable Boundary Quantizer (LBQ), a Learnable Equivalent Transformation (LET), and a Distributed Quantization Calibration (DQC) strategy. The authors evaluate on RealSR, DRealSR, and DIV2K validation at W8A8 and W6A6 settings, comparing against MaxMin, LSQ, Q-Diffusion, and EfficientDM. The central claims are that PassionSR achieves visual results comparable to full precision at both 8-bit and 6-bit, and that it significantly outperforms prior low-bit quantization methods for image super-resolution.","tokens_in":14660,"tokens_out":3342,"duration_ms":30493,"significance":"If the claims held, this would be a useful step toward deploying one-step diffusion super-resolution on resource-constrained hardware. The evaluation is broad in coverage: three datasets, eight metrics, and multiple bit widths, with all baselines re-run on the same backbone. The proposed components are clearly motivated, and the ablations show that LBQ and LET bring large gains over simple min-max quantization. However, the headline claim of FP-comparable visual quality at 6-bit is not supported by the paper's own main table, and the use of a simplified model as the full-precision reference introduces a transferability caveat. The method still appears competitive relative to the baselines, so the issues are fixable by reframing and additional evidence rather than by discarding the approach.","major_comments":[{"comment":"The claim in the Abstract and Section 5 that '8-bit and 6-bit obtains comparable visual results with full-precision model' is not supported by Table 2. On RealSR at W6A6, compared with PassionSR-FP, LPIPS rises from 0.3339 to 0.4199 (+25.8%), DISTS from 0.1765 to 0.2592 (+46.9%), NIQE from 4.336 to 8.618 (+98.7%), MANIQA drops from 0.4686 to 0.2131 (-54.5%), and CLIP-IQA drops from 0.7520 to 0.4612 (-38.7%). Similar large degradations appear on DIV2K val at W6A6. Even at W8A8, NIQE worsens from 4.336 to 6.070 on RealSR and from 3.573 to 4.424 on DIV2K val. Since no threshold or user study is provided, the 'comparable visual results' claim is unsupported by the main experimental table. The method may still be useful as a strong PTQ baseline, but the headline claim must be substantially qualified or replaced with a measured perceptual-comparison study.","section":"§4.2, Table 2"},{"comment":"The simplification from OSEDiff to PassionSR-FP is not neutral on all datasets, and this affects the scope of the quantization claims. Table 2 shows that on RealSR, PassionSR-FP has PSNR 26.70 versus OSEDiff's 25.57, a gain of more than 1 dB, and also differs on LPIPS (0.3339 vs 0.3447). Since the full-precision reference for all quantization comparisons is PassionSR-FP, the reported 'comparable to full precision' results are only established for the simplified architecture, not for OSEDiff itself. Replacing the DAPE and CLIPEncoder branches with a constant empty-prompt embedding may change behavior on inputs where text conditioning matters; the paper should either evaluate quantization on the original OSEDiff or explicitly limit the claims to the simplified model.","section":"§3.2, Table 2"},{"comment":"The ablation text states that DQC produces 'slight performance enhancement,' but Table 4 shows the opposite on the main distortion metrics: adding DQC to LBQ+LET reduces PSNR from 25.40 to 24.41 and SSIM from 0.7529 to 0.7374 on RealSR at W6A6. Some perceptual and no-reference metrics improve (LPIPS, DISTS, NIQE, MUSIQ), so the effect is mixed, not a clear enhancement. The authors should either describe DQC as improving convergence and memory cost while having mixed metric effects, or provide a principled explanation of why the selected metrics favor the DQC configuration.","section":"§4.3, Table 4"}],"minor_comments":[{"comment":"The table has formatting errors: 'PassionSR-FTW32A32' should likely be 'PassionSR-FP W32A32', and the entry '3,732 ↓17.50%)' is missing an opening parenthesis.","section":"Table 3"},{"comment":"The caption cites 'MaxMin [16]' but the method is otherwise referenced as [12]; please make the citation numbers consistent.","section":"Figure 6"},{"comment":"The sentence 'LET playes an important role' contains a typo ('playes' should be 'plays').","section":"§4.4"},{"comment":"Equation (7) writes the equivalent transformation as 's ⊙ K = K̃' but does not define the corresponding transformation for the query matrix Q; the notation in the displayed equation should be made symmetric and self-contained.","section":"§3.3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid empirical core in the comparison against PTQ baselines, but the central claim about FP-comparable visual quality is contradicted by its own Table 2 at 6-bit and partially at 8-bit. This is a fixable framing issue rather than a fatal flaw, but it must be addressed before publication. The use of the simplified PassionSR-FP as the reference also deserves explicit scope limitation, especially given the >1 dB PSNR difference on RealSR between OSEDiff and PassionSR-FP."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTwo things to know. First, this is the first post-training quantization pipeline aimed specifically at one-step diffusion super-resolution, and it quantizes the VAE as well as the UNet, which is the right move since the VAE dominates compute. Second, the headline claim that 6-bit results are 'comparable' to full precision is not supported by the paper's own numbers.\n\nWhat's actually new: the combination of a learnable boundary quantizer (a reparameterization of LSQ), a SmoothQuant-style equivalent transformation, and a two-stage calibration that stabilizes training and cuts calibration time roughly in half. Quantizing the VAE and simplifying OSEDiff by replacing the CLIP branch with a constant embedding are practical steps that make sense for deployment. The evaluation is solid: three datasets, eight metrics, baselines re-run on the same backbone. That is more careful than most quantization papers.\n\nThe soft spots are real but manageable. At 6-bit, Table 2 shows LPIPS up 25-75%, DISTS up 46-75%, NIQE roughly doubled, and MANIQA/CLIP-IQA down 31-55% relative to the FP backbone. That is not 'comparable.' At 8-bit the reference perceptual metrics are close, but NIQE still degrades 24-40%. The abstract and conclusion overclaim. Also, the DQC ablation (Table 4) shows PSNR dropping about 1 dB when DQC is added, while the text says 'slight performance enhancement.' Other metrics improve, so not fatal, but the description is misleading. And the 'full precision' reference is the authors' simplified PassionSR-FP, not the original OSEDiff; on RealSR the simplification alone changes PSNR by over 1 dB, so the transfer claim needs care.\n\nThe method itself looks sound, and the relative advantage over existing quantization methods is clear. No error bars or released code yet, which limits verification.\n\nThis deserves a serious referee. The reviewer should ask for a toned-down claim or a user study for 6-bit, a corrected DQC ablation description, and a clear statement of which full-precision model is the reference. Address those and it's a useful contribution for anyone deploying one-step diffusion SR.","headline":"First PTQ pipeline for one-step diffusion SR with VAE quantization, but the 6-bit 'comparable to full precision' claim is contradicted by the paper's own Table 2.","tokens_in":15188,"tokens_out":3814,"would_cite":true,"duration_ms":31860,"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":"A post-training quantization pipeline quantizes one-step diffusion super-resolution models to 8-bit and 6-bit with visual quality comparable to full precision, while cutting parameters by up to 86% and operations by up to 82%.","keywords":["image super-resolution","one-step diffusion models","post-training quantization","low-bit quantization","learnable boundary quantizer","learnable equivalent transformation","quantization calibration","efficient inference"],"falsifier":"Apply the same LBQ, LET, and DQC pipeline directly to the original OSEDiff architecture without removing DAPE and the CLIPEncoder, using the same 500-pair calibration set; if the 8-bit model's quality no longer tracks the full-precision model, then the reported success comes from the architectural simplification rather than from the quantization components.","tokens_in":14095,"feed_emoji":"🖼️","tokens_out":7582,"duration_ms":62771,"temperature":0.7,"pith_summary":"One-step diffusion super-resolution models restore images in a single denoising step, but their storage and compute costs are still too high for phones and embedded devices. This paper tries to establish that such models can be post-training quantized to 8-bit or 6-bit weights and activations without a visible quality loss, by first simplifying the architecture to its UNet and VAE cores and then calibrating trainable quantizer parameters. The paper's central claim is that its three components—learnable clipping boundaries, per-channel activation rescaling, and a two-stage calibration schedule—recover most of the full-precision behavior while cutting parameters by up to 86% and operations by up to 82%. If true, single-step diffusion super-resolution becomes deployable on practical hardware instead of requiring large GPU clusters.","feed_headline":"One-step diffusion SR shrinks to 8-bit with 82% fewer parameters","feed_subtitle":"A new calibration scheme keeps 6- and 8-bit outputs looking full-precision while slashing parameters up to 86%.","key_machinery":"The design is carried by a fake-quantization operator whose bounds are trainable: LBQ clips activations and weights to $[B_l, B_u]$, computes $\\alpha=(B_u-B_l)/(2^N-1)$ and $\\beta=B_l$, and produces $X_q=\\alpha\\lfloor (X_c-\\beta)/\\alpha\\rceil+\\beta$. LET then rescales each channel of the input as $\\tilde{X}=(X-\\delta)\\oslash s$ while transforming weights and bias to $\\tilde{W}=s\\odot W$ and $\\tilde{B}=B+\\delta W$, leaving the full-precision output mathematically unchanged but making activations smoother; these factors merge into neighboring layers after calibration, so the quantized runtime has no extra cost. DQC separates calibration into two stages so that LET's scale factors and LBQ's boundaries are not optimized simultaneously, which stabilizes training and lowers GPU memory. The UNet loss is computed in latent space through the one-step transformation $I(Z_l,\\varepsilon)=\\sqrt{1/\\hat{\\alpha}}Z_l-\\sqrt{(1-\\hat{\\alpha})/\\hat{\\alpha}}\\,\\varepsilon(Z_l)$, and the VAE encoder and decoder are each anchored to their full-precision outputs by MSE losses.","core_discovery":"Working from OSEDiff, the paper replaces the DAPE and CLIPEncoder conditioning branch with a constant empty-prompt embedding to obtain a simplified full-precision model, PassionSR-FP. The central claim is that this simplified model, when quantized with the proposed pipeline, delivers W8A8 and W6A6 outputs that are visually comparable to the 32-bit models while reducing parameters by 81.77% (8-bit) and 86.32% (6-bit) and operations by 76.56% and 82.42% in the UNet–VAE configuration. Quantizing the VAE is presented as essential because the VAE accounts for over 80% of the computational load once denoising is reduced to one step. The paper reports that prior diffusion-quantization baselines, which were designed for multi-step denoising, degrade sharply on one-step models, whereas PassionSR's learned quantizers adapt the distribution to the available bit widths.","pith_inferences":["Beyond the paper, the same two-stage calibration idea could transfer to other single-step generative models, such as one-step text-to-image or video diffusion, wherever quantizer parameters and activation scales are trained jointly.","Because the conditioning branch is replaced by a constant embedding, the quantized model's behavior for prompt-dependent inputs remains untested; a direct extension would quantize the DAPE and CLIPEncoder branches instead of dropping them, at the cost of some compression.","A testable extension would be to vary the calibration dataset's content (e.g., faces, text, low-light scenes) and measure whether the learned boundaries and scales shift, indicating that task-specific calibration matters."],"forward_implications":["At 8-bit precision, the quantized model is claimed to match or beat the full-precision reference on several metrics while using 81.77% fewer parameters and 76.56% fewer operations, so mobile inference becomes feasible.","At 6-bit precision, the pipeline remains usable with up to 86.32% parameter compression and 82.42% operation compression, extending low-bit post-training quantization to one-step diffusion models.","Quantizing the VAE rather than only the UNet is required for one-step models because the VAE dominates compute; including it raises the compression ratio from about 81% to 86% at 6-bit.","Existing multi-step diffusion quantization methods do not transfer directly to one-step models; the paper's comparisons show large drops for those baselines, supporting the need for one-step-specific calibration."],"supporting_citations":[{"why":"OSEDiff is the one-step diffusion backbone that PassionSR simplifies and quantizes; it supplies the full-precision reference values.","marker":"[42]"},{"why":"SmoothQuant+ is the source of the equivalent-transformation idea and the initialization used for LET.","marker":"[25]"},{"why":"OmniQuant is cited alongside SmoothQuant as inspiration for the learnable equivalent transformation of activations and weights.","marker":"[29]"},{"why":"MaxMin provides the fake-quantization formulation in Eq. (3) and is the baseline that learns no parameters.","marker":"[12]"},{"why":"LSQ is the learnable-step-size baseline and a direct comparison point in the quantized experiments.","marker":"[8]"},{"why":"Q-Diffusion is a leading diffusion-quantization method that the paper adapts to its backbone and uses as a baseline.","marker":"[18]"},{"why":"EfficientDM is a quantization-aware fine-tuning baseline that supplies a comparison for low-bit accuracy.","marker":"[9]"},{"why":"The straight-through estimator is the mechanism that lets gradients flow through LBQ's rounding operation.","marker":"[24]"},{"why":"DIV2K is the source of the 500 cropped LR-HR pairs used to build the calibration dataset.","marker":"[1]"}],"fun_headline_variants":["Diffusion SR at 6-bit: 86% fewer parameters, same looks","Adaptive quantizer slashes SR model size by 86%","One-step diffusion SR: 6-bit matches full precision","Post-training quantization tames one-step diffusion SR","PassionSR: 6-bit super-resolution without quality loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simplified PassionSR-FP, with the DAPE and CLIPEncoder branches removed and replaced by a constant empty-prompt embedding, behaves closely enough like the original OSEDiff for quantization results to transfer; the paper's Table 2 itself shows a 25.57 to 26.70 PSNR shift on RealSR, so this premise is doing real work.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion SR at 6-bit: 86% fewer parameters, same looks","Adaptive quantizer slashes SR model size by 86%","One-step diffusion SR: 6-bit matches full precision","Post-training quantization tames one-step diffusion SR","PassionSR: 6-bit super-resolution without quality loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000163,"raw_usage":{"total_tokens":1255,"prompt_tokens":968,"completion_tokens":287,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":200}},"tokens_in":584,"tokens_out":287,"duration_ms":3227,"temperature":1.0,"reasoning_tokens":200,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:30:06.498234+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the same LBQ, LET, and DQC pipeline directly to the original OSEDiff architecture without removing DAPE and the CLIPEncoder, using the same 500-pair calibration set; if the 8-bit model's quality no longer tracks the full-precision model, then the reported success comes from the architectural simplification rather than from the quantization components.","supporting_citations":[{"cited_title":"One-step effective diffusion network for real- world image super-resolution","cited_arxiv_id":null,"evidence_quote":"OSEDiff is the one-step diffusion backbone that PassionSR simplifies and quantizes; it supplies the full-precision reference values."},{"cited_title":"Omniquant: Omnidi- rectionally calibrated quantization for large language models","cited_arxiv_id":null,"evidence_quote":"OmniQuant is cited alongside SmoothQuant as inspiration for the learnable equivalent transformation of activations and weights."},{"cited_title":"Quantization and train- ing of neural networks for efficient integer-arithmetic- only inference","cited_arxiv_id":null,"evidence_quote":"MaxMin provides the fake-quantization formulation in Eq. (3) and is the baseline that learns no parameters."},{"cited_title":"Learned step size quantization","cited_arxiv_id":null,"evidence_quote":"LSQ is the learnable-step-size baseline and a direct comparison point in the quantized experiments."},{"cited_title":"Q-diffusion: Quantizing diffusion models","cited_arxiv_id":null,"evidence_quote":"Q-Diffusion is a leading diffusion-quantization method that the paper adapts to its backbone and uses as a baseline."},{"cited_title":"Nonuniform-to-uniform quantization: Towards accurate quantization via gen- eralized straight-through estimation","cited_arxiv_id":null,"evidence_quote":"The straight-through estimator is the mechanism that lets gradients flow through LBQ's rounding operation."},{"cited_title":"Ntire 2017 chal- lenge on single image super-resolution: Dataset and study","cited_arxiv_id":null,"evidence_quote":"DIV2K is the source of the 500 cropped LR-HR pairs used to build the calibration dataset."}],"review_version":1}