{"id":"2fad764a-34b5-429d-92f5-9903cd2be899","arxiv_id":"2507.12933","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A post-training quantization method that combines learned channel scaling and power-of-two scaling keeps diffusion image quality high at 4-bit weight, 6-bit activation precision.","lead":"Diffusion image models normally need heavy compute; this paper shows they can be compressed to 4-bit weights and 6-bit activations with a post-training quantization method that redistributes and removes outliers per channel. The practical payoff is faster, lighter image generation on limited hardware without retraining the model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation confound: in several tables the quantized DMQ model beats the full-precision model on FID, so the claim of 'preserving high image generation quality' is not cleanly established and needs re-testing under a standard sampling protocol.","rationale":"I read the paper in good faith. The proposed components (LES, adaptive timestep weighting, PTS with voting) are technically coherent, the code is promised, and the comparison with four baselines across three generation settings is extensive. The reader's condition is appropriate: the method is plausible and likely improves over prior PTQ at low bit widths, but the magnitude of the improvement is uncertain. My stress-test pass found a different load-bearing concern than the reader's 'static tau/delta' assumption. The reader did note the unexplained pattern of quantized models beating full-precision FID, so there is partial overlap, but they did not treat it as the primary weakness. In my view, the static-scaling assumption is real but secondary: the empirical success across datasets suggests that, at least for the tested models and calibration sizes, a single set of channel-wise factors is adequate. The evaluation confound is more decisive because it directly affects whether the central claim is measured correctly. If the full-precision baseline is run in a deliberately suboptimal regime (20-step DDIM eta=1.0), and if the calibration-time optimization effectively re-trains part of the network, then the reported FID numbers cannot be interpreted as evidence of quantization fidelity. The concrete test I propose would settle this by re-running both full precision and DMQ under a standard, deterministic sampling protocol with error bars, and by rebenchmarking the modified baselines. Until that test is done, the appropriate verdict remains CONDITIONAL: accept only if the re-evaluation confirms that DMQ preserves output quality at W4A6 under the same protocol as full precision. I would not move to REJECT because the method has real merit and the identified issue is correctable rather than fatal; I would not move to ACCEPT because the current evidence is insufficiently clean.","tokens_in":21060,"tokens_out":11497,"duration_ms":124247,"concrete_test":"Re-run full-precision and DMQ under the standard protocols used in prior diffusion-PTQ papers: DDIM with 20 steps and eta=0 for LDM (FFHQ, LSUN, ImageNet) and DDIM with 50 steps and eta=0 for Stable Diffusion, with 3 different random seeds, identical calibration sets, and the same unified sampling code. Report FID and sFID with error bars for both full precision and DMQ at W4A6 and W4A8. If DMQ's W4A6 FID on ImageNet or FFHQ is no longer below the corresponding full-precision FID (e.g., ImageNet W4A6 rises above 11.34), the claim that DMQ preserves full-precision quality must be weakened, and the baselines should be rebenchmarked under the same protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: DMQ achieves stable W4A6 quantization while preserving high image generation quality. The support is Tables 2-4, but those tables contain a pattern that undermines the claim: the quantized model often produces substantially better FID than the full-precision model. In Table 3 (ImageNet), DMQ at W4A6 reaches FID 7.81 versus full-precision 11.34; in Table 2 (FFHQ), DMQ at W4A6 reaches 26.38 versus full-precision 31.34; even the W4A8 ablation in Table 5 ends at FID 30.37 versus full-precision 31.34. If quantization only compresses the model, it is hard to explain such large improvements. The most plausible explanation is that the evaluation protocol gives an unfair advantage to the quantized pipeline. First, the paper uses a DDIM sampler with eta=1.0 and only 20 steps for unconditional generation, which is an unusually noisy and likely suboptimal regime for the full-precision model; the quantized pipeline, by contrast, includes calibration-time optimization (LES factors plus BRECQ weight rounding, trained for 4000-6000 iterations on 5120-10240 calibration samples) that can alter the generative distribution. The paper reports no error bars or multiple-seed variance, so it is impossible to tell whether the apparent 'preservation' or 'improvement' is systematic or an artifact of a single unlucky full-precision run. The claim that previous W4A6 methods 'fail' is also sensitive to the fact that the authors modified PTQD and TFMQ-DM to quantize skip-connection/downsample/upsample layers that their original implementations left in full precision; this may be a legitimate equalization, but it means the baselines are not the published versions. The method itself is coherent and the components are well motivated, but the headline result is not yet supported by a clean quantization-preservation measurement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes DMQ, a post-training quantization method for diffusion models that combines Learned Equivalent Scaling (LES) with an adaptive timestep weighting loss and channel-wise Power-of-Two Scaling (PTS) with a voting-based factor selection. LES learns channel-wise scaling factors that rebalance quantization error between weights and activations; the adaptive weighting up-weights timesteps whose small errors still have large influence on final output; PTS targets layers with extreme inter-channel variance such as residual-block skip connections. The method is evaluated on LDM-based unconditional and class-conditional generation and on Stable Diffusion text-conditioned generation at W4A8 and W4A6, where it generally outperforms prior PTQ baselines.","tokens_in":21455,"tokens_out":8835,"duration_ms":98356,"significance":"DMQ is a clearly specified PTQ framework with a credible mechanism: the analysis of timestep-dependent error influence in Fig. 4 is interesting, the bit-shift formulation in Eq. (13) is hardware-conscious, and the voting algorithm is a plausible safeguard against small-calibration overfitting. The paper also releases code and documents implementation details thoroughly. If the reported gains survive a more rigorous evaluation protocol, the work would be a valuable contribution to low-bit diffusion quantization. However, the current evaluation contains several confounds that make the central 'preserving image generation quality' claim hard to assess, so the significance is conditional on fixing those issues.","major_comments":[{"comment":"It is unexplained why the quantized model often produces substantially better FID than the full-precision model, e.g., FFHQ W8A8 FID 26.78 vs 31.34 and ImageNet W4A6 FID 7.81 vs 11.34. This undermines the 'preserving' framing and suggests that the evaluation protocol (DDIM with eta=1.0 and 20 steps for unconditional generation) is far from optimal for the full-precision model, while the calibration-time optimization in DMQ acts as an additional adaptation. Please report full-precision results under the same protocol and seed, report multiple seeds with error bars, and include results under a standard sampling protocol (e.g., DDIM eta=0 with 50 or 100 steps); if the quantized model still outperforms full precision, the reason should be discussed rather than treated as preservation.","section":"Section 5.1, Tables 2 and 3"},{"comment":"The LES factors tau and PTS exponents delta are learned once from calibration data and then used for every denoising step. The paper justifies this through Fig. 4, which shows layer-wise average quantization error trends, not channel-wise outlier stability. Since per-channel scaling factors are only meaningful if the identity of outlier channels does not shift across timesteps, please provide channel-wise stability evidence (e.g., how often the max-magnitude channel or the selected delta* changes with t) or an ablation comparing static factors with per-timestep recalibrated factors.","section":"Sections 4.1 and 4.2"},{"comment":"The authors state that PTQD and TFMQ-DM were modified to quantize skip-connection, downsample, and upsample layers for fair comparison. Because these are exactly the layers where DMQ applies PTS, such a modification can change baseline behavior substantially. Please report per-method modification details, state explicitly which numbers in Tables 2-4 include the modified code, and clarify whether the time-embedding caching described in Section 5 applies to all methods or only to DMQ; otherwise the baseline comparison is not fully specified.","section":"Appendix B"},{"comment":"All FID/sFID numbers are reported without error bars or multiple-seed statistics. For comparisons where the differences between methods are often smaller than one FID point, a single run cannot support the claim of 'stable' performance. Please provide at least three sampling seeds (and, where feasible, several calibration/optimization runs) with means and standard deviations.","section":"Section 5.1, Tables 2-4"},{"comment":"The claim that W4A6 quantization 'preserves high image generation quality' is not consistently supported: in Table 2 (LSUN-Church) the W4A6 FID is 14.28 vs 11.28 for full precision, and in Table 4 (MS-COCO, Stable Diffusion) W4A6 degrades FID from 20.38 to 29.58 and sFID from 62.55 to 67.05. Please qualify the claim or report the average degradation across all tasks.","section":"Abstract and Section 5.1"}],"minor_comments":[{"comment":"The word 'qauntization' should be 'quantization'.","section":"Figure 4 caption"},{"comment":"The sentence mentioning 'the scaled activation \\hat{X} and weights \\hat{X}' should read '\\hat{W}' for the weights.","section":"Figure 3 caption"},{"comment":"The header involving 'T /c' is confusing; it should clearly state that T is the number of sampling steps for calibration data and n is the number of samples per step.","section":"Table 8"},{"comment":"The phrase 'for all channel index' should be 'for each channel index', and the tie-breaking rule for the mode of delta* is unspecified.","section":"Algorithm 1"},{"comment":"The hyperparameters alpha, kappa, xi, and D are not accompanied by sensitivity analyses; at least a brief discussion of their ranges would help the reader assess robustness.","section":"Table 8"}],"recommendation":"major_revision","confidential_remarks":"The core method is plausible and the writing is generally clear, but the evaluation protocol needs to be tightened before publication. The biggest risk is that the apparent improvement over full precision is an artifact of the unusual DDIM eta=1.0 20-step protocol; I would ask the authors for seed-level statistics and a standard-protocol baseline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a plausible, well-motivated PTQ method for diffusion models, with a clear combination of existing ideas (equivalent scaling, power-of-two scaling) plus a new adaptive timestep weighting and a voting mechanism for factor selection. The ablations are coherent and each component earns its place. But the evaluation has a load-bearing problem: the quantized model beats the full-precision baseline on FID by a large margin in several settings (FFHQ W8A8 26.78 vs 31.34; ImageNet W4A6 7.81 vs 11.34). That is not a sign of \"preserving\" quality; it is a sign that the measurement is not a clean quantization-preservation test. No error bars or seeds are reported, so we cannot tell whether the effect is systematic or an artifact of a single run. The paper also discloses that PTQD and TFMQ-DM were modified to quantize layers their original implementations left in full precision; that is transparent and defensible, but it means the baselines are not the released versions.\n\nWhat is genuinely new: the learned equivalent scaling formulation with timestep-adaptive loss, and the voting-based channel-wise power-of-two scaling. The paper's analysis of where outliers live (skip connections, inter-channel variance) is useful and the ablations support the design choices. The code is released, so the method is reproducible.\n\nThe main soft spot is the evaluation protocol. The authors use a 20-step DDIM sampler with eta=1.0, which is an unusual and likely suboptimal regime for the full-precision model; the quantized pipeline, by contrast, includes calibration-time optimization of scaling factors and weight rounding (4000-6000 iterations) that can shift the generative distribution. Quantization should be a small perturbation, not an improvement of 3-4 FID points over full precision. The most plausible reading is that the comparison protocol is tilted toward the quantized pipeline. The authors should re-test under a standard sampling protocol (e.g., 50-step DDIM eta=0, or 100-step DDIM) and report multiple seeds and confidence intervals. They should also report the full-precision baseline with the same calibration-time optimization applied (if that is even well-defined) or otherwise explain the improvement. A secondary concern: the channel-wise scaling factors are static across timesteps, and the paper only shows layer-wise error trends, not channel-wise stability. That said, the method is coherent, the components are well motivated, and I would send it to a serious referee. The evaluation needs major work, but the core idea and the ablations are worth refereeing.","headline":"Plausible PTQ method for diffusion models, but the evaluation is confounded by quantized models beating full precision, so the quality-preservation claim needs a clean re-test.","tokens_in":22027,"tokens_out":3477,"would_cite":false,"duration_ms":36171,"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":"A post-training quantization method called DMQ keeps diffusion image generators accurate at 4-bit weights and 6-bit activations by learning channel-wise scaling and taming outlier channels with power-of-two bit shifts.","keywords":["diffusion models","post-training quantization","outlier channels","equivalent scaling","power-of-two scaling","timestep weighting","low-bit inference","W4A6"],"falsifier":"Re-optimize $\\tau$ and $\\delta$ separately for each denoising timestep using the same calibration protocol, and compare W4A6 FID with the static single-factor version; if per-timestep re-optimization clearly improves FID on FFHQ or ImageNet, the timestep-invariant outlier assumption is false, and if the two are indistinguishable, it is confirmed. A supporting observation would be a channel-level correlation matrix of activation outliers across timesteps for skip-connection layers, showing whether the same channels stay extreme throughout the denoising trajectory.","tokens_in":20864,"feed_emoji":"🖼️","tokens_out":11639,"duration_ms":117246,"temperature":0.7,"pith_summary":"Diffusion image generators normally must be quantized gently: at 4-bit weights and 6-bit activations, earlier post-training methods collapse, because a few outlier channels stretch the quantization range so far that the rest of the tensor is quantized coarsely. This paper proposes a two-part recipe that keeps such aggressive quantization stable. A learned equivalent scaling factor optimizes a channel-wise rescaling that shifts quantization difficulty between weights and activations, using a timestep-weighted loss that treats early denoising steps as more consequential than their raw error suggests. A power-of-two channel scaling then directly shrinks outliers in skip-connection layers, with a voting rule over calibration samples selecting the scaling exponents reliably. Across unconditional, class-conditional, and text-to-image models, the paper reports that the recipe beats prior methods and preserves image quality at W4A6 and W4A8.","feed_headline":"Diffusion models keep quality at 4-bit weights, 6-bit activations","feed_subtitle":"A learned channel scaling plus power-of-two bit-shifts tames the outliers that broke earlier low-bit methods.","key_machinery":"The load-bearing object is the channel-wise scaling pair: $\\tau \\in \\mathbb{R}^{C_{\\text{in}}}$ for Learned Equivalent Scaling and $\\delta \\in \\mathbb{R}^{C_{\\text{in}}}$ for Power-of-Two Scaling. LES uses the identity $Y = (X/\\tau)(\\tau^\\top \\odot W)$ to move quantization difficulty between activation and weight channels, with $\\tau$ tuned by minimizing layer-output MSE under the adaptive timestep weighting $\\lambda_t = (1 - \\Lambda_t / \\sum_{t'}\\Lambda_{t'})^\\alpha$, where $\\Lambda_t$ is a momentum-averaged accumulated loss. PTS applies $\\tilde{X} = \\mathrm{clamp}(\\lfloor X / (2^{\\delta} \\odot s^{(X)}) \\rceil, l, u)$ to activation channels and folds $2^{\\delta_k}$ into the loaded weights as a left bit-shift $\\tilde{W}_{kj} \\ll \\delta_k$, so extreme outliers are suppressed without per-step computation; the exponents come from a mode-voting algorithm with agreement threshold $\\kappa$ over calibration samples. Both $\\tau$ and $\\delta$ are learned once and fused into static scales, which is what keeps inference overhead low.","core_discovery":"The central claim is that the failure of low-bit diffusion quantization is dominated by channel-wise outliers, and that those outliers can be handled by two complementary mechanisms without retraining. Learned Equivalent Scaling (LES) optimizes a channel-wise scale vector $\\tau \\in \\mathbb{R}^{C_{\\text{in}}}$ so that the reparametrization $Y = (X/\\tau)(\\tau^\\top \\odot W)$ minimizes layer-output quantization error over calibration samples; because early denoising steps have small but accumulating errors, the optimization uses an adaptive timestep weighting that up-weights those steps. The second mechanism, channel-wise Power-of-Two Scaling (PTS), applies $2^{\\delta}$ factors to activation channels of layers with extreme inter-channel variance, such as skip connections, removing outliers rather than transferring them to the weights; the exponents are selected by a per-sample voting procedure with an agreement threshold so that a small calibration set does not overfit. PTS is implemented as a bit-shift on the loaded weights, so the paper argues the added cost is minimal. With both mechanisms, the paper reports stable W4A6 quantization, with image quality close to full precision on FFHQ, LSUN-Bedroom, LSUN-Church, ImageNet LDM, and Stable Diffusion text-to-image generation.","pith_inferences":["Beyond the paper's U-Net experiments, the static $\\tau$ and $\\delta$ premise is most likely to break in transformer-based diffusion backbones such as DiT, where activations vary more strongly with timestep and conditioning; a timestep-conditional variant of the voting rule would be a direct test.","The voting mechanism suggests a general recipe for choosing outlier-scaling hyperparameters from small calibration sets: take the per-sample best candidate, keep only the modal choice, and default to no scaling when agreement is low; that recipe could transfer to other iterative generative models such as flow matching or consistency models.","The adaptive timestep weighting is heuristic in form (a focal-loss-style power law); a natural extension is replacing it with a measured sensitivity that tracks how much a unit of injected noise at each timestep propagates to the final image, which would put the step prioritization on firmer ground.","Because power-of-two scaling is applied only to a few skip-connection layers, the reported speedup is mostly attributable to low-bit GEMM itself; one could test whether the same outlier analysis identifies other high-variance layers, such as attention projections in transformer blocks, where bit-shift scaling would pay for itself."],"forward_implications":["Diffusion image generators can be deployed at W4A6 with FID and sFID close to full precision, which removes a major obstacle to running them on memory-limited hardware.","Static, precomputed scaling factors are sufficient; because $\\tau$ and $2^{\\delta}$ are fused into scales and weight loads, low-bit inference does not pay a per-timestep rescaling cost.","Outlier channels in skip connections are the primary bottleneck at low bit widths; targeting only those layers with power-of-two scaling is enough to recover most of the lost quality.","The recipe transfers across model families: latent diffusion models for unconditional, class-conditional, and text-to-image generation all improve, suggesting the outlier structure is a general property of diffusion U-Nets."],"supporting_citations":[{"why":"Supplies the baseline equivalent-scaling scheme that the paper shows fails on diffusion models, motivating the learned alternative.","marker":"[66]"},{"why":"Provides the calibration-data protocol and the main prior PTQ baseline; its skip-connection bimodal observation is extended to ResBlock shortcuts.","marker":"[32]"},{"why":"The strongest prior timestep-specific quantization baseline, compared at every bit-width setting across all datasets.","marker":"[22]"},{"why":"Prior noise-correction PTQ baseline used in comparisons; the paper distinguishes its outlier handling from noise correction.","marker":"[16]"},{"why":"Prior distribution-alignment PTQ baseline compared in the main tables and identified as orthogonal to the proposed method.","marker":"[39]"},{"why":"BRECQ weight quantization is applied after learning the equivalent scaling factors, so it is part of the pipeline.","marker":"[33]"},{"why":"AdaRound's layer-wise reconstruction objective is the template for the LES optimization.","marker":"[42]"},{"why":"Focal loss is the stated inspiration for the adaptive timestep weighting that prioritizes early denoising steps.","marker":"[38]"}],"fun_headline_variants":["DMQ tames outliers for 4-bit diffusion quantization","4-bit diffusion weights with stable image quality","Outlier-aware scaling boosts low-bit diffusion models","Post-training quantization for diffusion at W4A6","Learned scaling beats outlier challenge in diffusion PTQ"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the channel-wise outlier pattern in activations stays roughly the same across all denoising timesteps, so a single set of scaling factors learned from calibration data can serve every step.","fun_headline_variants_meta":{"raw":{"variants":["DMQ tames outliers for 4-bit diffusion quantization","4-bit diffusion weights with stable image quality","Outlier-aware scaling boosts low-bit diffusion models","Post-training quantization for diffusion at W4A6","Learned scaling beats outlier challenge in diffusion PTQ"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000202,"raw_usage":{"total_tokens":1447,"prompt_tokens":1073,"completion_tokens":374,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":689,"completion_tokens_details":{"reasoning_tokens":300}},"tokens_in":689,"tokens_out":374,"duration_ms":4918,"temperature":1.0,"reasoning_tokens":300,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:34:58.039425+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-optimize $\\tau$ and $\\delta$ separately for each denoising timestep using the same calibration protocol, and compare W4A6 FID with the static single-factor version; if per-timestep re-optimization clearly improves FID on FFHQ or ImageNet, the timestep-invariant outlier assumption is false, and if the two are indistinguishable, it is confirmed. A supporting observation would be a channel-level correlation matrix of activation outliers across timesteps for skip-connection layers, showing whether the same channels stay extreme throughout the denoising trajectory.","supporting_citations":[{"cited_title":"Smoothquant: Accurate and effi- cient post-training quantization for large language models","cited_arxiv_id":null,"evidence_quote":"Supplies the baseline equivalent-scaling scheme that the paper shows fails on diffusion models, motivating the learned alternative."},{"cited_title":"Q-diffusion: Quantizing diffusion models","cited_arxiv_id":null,"evidence_quote":"Provides the calibration-data protocol and the main prior PTQ baseline; its skip-connection bimodal observation is extended to ResBlock shortcuts."},{"cited_title":"Tfmq-dm: Temporal feature maintenance quantization for diffusion models","cited_arxiv_id":null,"evidence_quote":"The strongest prior timestep-specific quantization baseline, compared at every bit-width setting across all datasets."},{"cited_title":"Ptqd: Accurate post-training quantization for diffusion models.Advances in Neural Information Pro- cessing Systems, 36, 2024","cited_arxiv_id":null,"evidence_quote":"Prior noise-correction PTQ baseline used in comparisons; the paper distinguishes its outlier handling from noise correction."}],"review_version":1}