{"id":"966deb2e-4d1c-413a-98db-c68f12006a52","arxiv_id":"2412.16700","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TCAQ-DM applies timestep-channel reparameterization, adaptive post-Softmax quantizer selection, and progressive aligned reconstruction to post-train quantize diffusion models down to W4A4 with usable FID.","lead":"Diffusion models produce high-quality images but are slow and memory-heavy. This paper proposes a post-training quantization method that compresses weights and activations to 4 or 6 bits while keeping generated images usable, by adapting quantization ranges across timesteps and channels and recalibrating with data from the compressed model.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline W6A6/W4A4 numbers may keep post-Softmax at 8-bit; stated bit-widths and SOTA comparisons could be misleading.","rationale":"I disagree with the reader's choice of PAR as the weakest link. PAR is applied after TCR+DAQ already reduce W4A4 FID from 371.61 to 9.09, so even if the resampling loop is imperfect, it is not the load-bearing element for the headline W4A4 viability. The more load-bearing issue is the bit-width actually used for post-Softmax layers in the main tables. The paper itself says all experiments use an 8-bit post-Softmax layer unless specifically claimed, and Table 1 does not specify S bit-widths, while Table 2 does. If the default applies, the W6A6 and W4A4 labels are inaccurate, and the comparison to TFMQ-DM/PTQ4DM/Q-Diffusion in W4A4 is not controlled. The proposed check (force S4 and recompute) settles whether the headline is an artifact of the default. If the numbers degrade, the central claim must be revised to 'W4A4 with 8-bit softmax'; if they survive, the paper is strengthened. The paper has real strengths: broad experiments, ablations for all three modules, and an explicit PAR algorithm, but the reporting gap on softmax bit-width is directly about the truth of the headline claim and therefore more load-bearing than the PAR sampling concern.","tokens_in":15033,"tokens_out":8773,"duration_ms":75823,"concrete_test":"Inspect the released configuration or ask the authors for the Table 1 W4A4 and W6A6 runs, and record the bit-width assigned to every post-Softmax layer. Then re-run the W4A4 configuration with post-Softmax forced to 4-bit (S4) and recompute FID. If FID degrades materially (e.g., toward the >100 values reported for prior methods), the headline W4A4 claim depends on an 8-bit softmax exemption; if FID remains near 6.38, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's Implementation Details state: 'All experiments are conducted with an 8-bit post-Softmax layer unless being specifically claimed.' If this default applies to Table 1, then the headline CIFAR-10 results labeled W6A6 (FID 4.40) and W4A4 (FID 6.38) are not pure W6A6/W4A4: the post-Softmax attention activations remain at 8-bit. Table 2 explicitly reports S8/S32 for every method, but Table 1 omits the S column, so the comparison against PTQ4DM*, Q-Diffusion*, and TFMQ-DM* in W4A4 may pair an 8-bit softmax (ours) against 4-bit softmax (baselines). Since the paper's central claim hinges on 'comparable FID to the full precision model in the W6A6 setting' and 'available images in the W4A4 settings', a hidden 8-bit softmax exemption would make the headline comparisons apples-to-oranges. Table 6 shows the authors can quantize post-Softmax to S4/S6, so the main tables should state which bit-width was actually used. This is not a minor formatting issue: it changes what the claimed W4A4 and W6A6 numbers mean.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes TCAQ-DM, a post-training quantization (PTQ) method for diffusion models. The method combines three components: (i) a timestep-channel joint reparameterization (TCR) module that rescales convolution activations along channels with timestep-aware weights; (ii) a dynamically adaptive quantizer (DAQ) that selects between log2 and uniform quantizers for post-Softmax layers based on a likelihood-ratio criterion; and (iii) a progressively aligned reconstruction (PAR) strategy that re-samples the calibration set from the partially quantized model during reconstruction. The method is evaluated on CIFAR-10 with DDIM, LSUN-Bedrooms/LSUN-Churches with LDM-4, and ImageNet with LDM-4, reporting FID, sFID, and IS. The headline results are W6A6 FID 4.40 versus 4.14 for the full-precision model on CIFAR-10, and W4A4 FID 6.38 versus 236.63 for the best prior method.","tokens_in":15288,"tokens_out":7598,"duration_ms":62344,"significance":"If the reported bit-widths are made fully transparent, this is a practically significant contribution: the W4A4 numbers are dramatically better than prior PTQ methods for diffusion models, and the TCR reparameterization is mathematically sound because channel-wise rescaling of inputs and weights preserves the convolution output. The ablation study in Table 5 is consistent with each component adding value, and the evaluation spans multiple datasets and two diffusion-model families. The main reservation is that the paper's default 8-bit post-Softmax setting makes the headline W4A4/W6A6 labels incomplete unless the tables state the post-Softmax width for every entry.","major_comments":[{"comment":"The Implementation Details state that all experiments use an 8-bit post-Softmax layer unless specifically claimed, but Tables 1, 3, and 4 do not include a post-Softmax bit-width column or a caption note. As a result, the labels \"W6A6\" and \"W4A4\" for the proposed method in these tables are incomplete: the post-Softmax activations are S8, not 6-bit or 4-bit. This is not purely cosmetic, because the comparisons against PTQ4DM*, Q-Diffusion*, and TFMQ-DM* in W4A4 may pair an S8 model with baselines whose post-Softmax width differs. Please add an S column (or explicit caption statements) to Tables 1, 3, and 4, and for every baseline report the exact post-Softmax bit-width used; if some baselines use S32 or S4, the comparison should be re-labeled so that the reader can see whether the comparison is same-precision.","section":"Implementation Details / Tables 1, 3, 4"},{"comment":"The PAR module resamples the calibration set from the partially quantized model (Algorithm 1, lines 5-6). The ablation in Table 5 attributes a 2.71 FID improvement at W4A4 (from 9.09 to 6.38) to PAR, but the design does not isolate the effect of aligned sampling from the effect of additional reconstruction iterations, since each PAR round performs 10,000 extra iterations. Without a control that runs the same number of extra iterations while drawing calibration samples from the full-precision model, or that keeps the calibration set fixed, the specific mechanism behind the PAR gain is not established. Please add such a control experiment.","section":"Progressively Aligned Reconstruction / Table 5"},{"comment":"The selection variable Rg is introduced as a \"ratio\" of likelihood estimation results but is compared against zero. A likelihood ratio is non-negative, so either Rg is a log-likelihood ratio or the definition is incomplete. The text should give the exact formula for Rg, specify the candidate distributions (power law, log-normal, exponential) and how the MLE fit is performed per layer and timestep, and clarify the decision rule in Eq. (8). Without this information, the DAQ module is not reproducible.","section":"Dynamically Adaptive Quantizer / Eq. (8)"}],"minor_comments":[{"comment":"The subscript in Eq. (3) uses j while Eq. (5) uses d for the same channel index; please unify the notation.","section":"Methodology, Eq. (3) vs Eq. (5)"},{"comment":"The text refers to \"the basic reconstruction with BRECQ\" while Algorithm 1 says \"perform the basic Adaround operation\"; please make the reconstruction procedure terminology consistent.","section":"Methodology vs Algorithm 1"},{"comment":"The caption says \"except for the post-Softmax quantization bit-width\" but the table lists several S values for both our method and the baselines; please spell out which setting applies to which method.","section":"Table 2 caption"},{"comment":"No error bars or multiple seeds are reported; several high-bit comparisons have small FID differences (for example, 3.21 vs 3.14 in Table 2), so reporting variance or the number of evaluation seeds would help the reader judge significance.","section":"Experimental Results"},{"comment":"In the abstract, \"fail to tackle with the large variations\" should be \"fail to handle the large variations\".","section":"Abstract"},{"comment":"The reference \"Fig. Ca (b) and (c)\" appears to be a typo; it should be \"Fig. C(b) and C(c)\".","section":"Supplementary Material, Figure C"}],"recommendation":"major_revision","confidential_remarks":"The post-Softmax bit-width ambiguity is likely to be the main point of contention at review; the authors should be required to make every table self-contained regarding the S bit-width. The PAR control experiment is inexpensive and should be mandatory. I do not see evidence of fabrication or internal inconsistency, but the current presentation overstates the purity of the W4A4/W6A6 settings."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a genuinely useful PTQ-for-diffusion paper with strong W4A4 numbers, but the headline results may be quoted with an 8-bit post-Softmax layer while some baselines are quoted at lower softmax precision. If that's true, the claimed advantage over prior art is partly apples-to-oranges. The fix is easy—report S in every table—and the underlying method still looks sensible.\n\nWhat's new: they combine timestep-channel reparameterization (TCR), per-timestep quantizer selection for post-Softmax layers (DAQ), and progressive recalibration from the quantized model (PAR). The combination is genuinely new, and the W4A4 results are impressive: 6.38 FID on CIFAR-10 versus collapsed baselines in the hundreds is a real advance, even if the softmax is at 8-bit. The algebra for TCR is correct—it's a valid linear reparameterization that preserves the output—and the ablations are consistent with the story. The comparison across three datasets and two model families is broader than most papers in this area.\n\nThe soft spots: first, the post-Softmax bit-width ambiguity is load-bearing. The implementation details say \"all experiments are conducted with an 8-bit post-Softmax layer unless being specifically claimed.\" Table 1 has no S column, so its W6A6 and W4A4 rows are presumably W6A6S8 and W4A4S8. Baselines re-run from open code may have used lower softmax precision. That would make the W4A4 comparison in Table 1 unfair. Table 2 explicitly reports S8/S32, so the authors know the difference; they just didn't apply it consistently to Table 1. This is a reporting flaw, but it changes the meaning of the headline numbers.\n\nSecond, there are no error bars, no code or checkpoints released, and several hyperparameters (TCR clamp range, DAQ likelihood threshold, PAR rounds) are tuned on the benchmarks themselves. The PAR resampling mechanism—recalibrating with samples from the partially quantized model—could in principle amplify distortion, though the ablation shows it helps. A control with full-precision recalibration would strengthen that section. These are fixable but currently make exact replication hard.\n\nDespite the reporting issues, the method is coherent and the W4A4 results, even with S8 softmax, are worth taking seriously. I'd send this to peer review. The authors need to disclose softmax bit-width in every table, release code, and run the PAR control. If the numbers hold under those conditions, it's a solid acceptance.","headline":"Strong W4A4 PTQ results for diffusion models, but the headline tables may be comparing 8-bit softmax against lower-bit softmax baselines without disclosing it; still worth a serious referee.","tokens_in":15845,"tokens_out":3262,"would_cite":true,"duration_ms":27880,"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 method, TCAQ-DM, claims 4-bit diffusion models can generate usable images, cutting CIFAR-10 FID from over 236 to 6.38.","keywords":["post-training quantization","diffusion models","low-bit quantization","activation reparameterization","adaptive quantizer","calibration alignment","FID","image generation"],"falsifier":"Run the PAR procedure for 4-bit weights and activations on CIFAR-10 with two variants: resampling from the quantized model (as proposed) and resampling from the full-precision model at every round. If the full-precision-resampled control matches or beats the proposed PAR FID of 6.38, the alignment mechanism is not responsible for the gain and the central claim about input mismatch is undercut.","tokens_in":14812,"feed_emoji":"🖼️","tokens_out":8513,"duration_ms":57573,"temperature":0.7,"pith_summary":"Diffusion models produce high-quality images but are expensive to run, and post-training quantization can cut that cost with little retraining. The paper argues that prior quantization methods fail on diffusion models because activation ranges swing wildly across channels and timesteps, and because the data used during calibration does not match the data seen during iterative inference. To fix this, it proposes three linked modules: a timestep-channel reparameterization that evens out activation ranges, a dynamically adaptive quantizer that picks log or uniform quantization per timestep for softmax outputs, and a progressively aligned reconstruction that recalibrates on samples drawn from the partially quantized model. On CIFAR-10, the method reports FID 4.40 at 6-bit weights and activations against 4.14 for the full-precision model, and FID 6.38 at 4-bit weights and activations where prior methods collapse to FIDs above 200.","feed_headline":"4-bit quantized diffusion models generate usable images","feed_subtitle":"Three calibration fixes cut CIFAR-10 FID from 236 to 6.4 at 4-bit weights and activations.","key_machinery":"The load-bearing object is a per-channel, per-timestep scaling vector derived from the pooled maximum activation of each channel across timesteps, applied as a reparameterization that shifts value range from activations into weights, plus a per-layer, per-timestep likelihood-ratio test that selects between a $\\log_2$ quantizer and a uniform quantizer for post-Softmax activations. The third piece is an iterative recalibration loop, described in Algorithm 1, in which the calibration set is resampled from the partially quantized model and used to refine quantization parameters over successive rounds. Together they let a single set of quantization parameters handle distributions that change shape as denoising proceeds.","core_discovery":"The central claim is that a post-training quantization pipeline can quantize diffusion models to very low bit-widths without catastrophic quality loss, provided it adapts to three properties of diffusion computation: the joint fluctuation of activation ranges over timesteps and channels, the timestep-dependent shape of post-Softmax distributions, and the input mismatch between one-shot calibration and iterative inference. The authors introduce the timestep-channel joint reparameterization (TCR) module, which rescales activations channel-wise using a timestep-averaged weight so that outlier channels are tamed; the dynamically adaptive quantizer (DAQ), which uses a maximum-likelihood power-law fit to decide per timestep whether a $\\log_2$ quantizer or a uniform quantizer better matches the post-Softmax distribution; and the progressively aligned reconstruction (PAR) strategy, which resamples the calibration set from the quantized model during reconstruction so the inputs match inference-time data flow. Supported by ablations, each module is reported to contribute to the final FID, with the full method producing usable images at 4-bit weights and activations where baseline PTQ methods produce FIDs of hundreds.","pith_inferences":["The PAR loop assumes the quantized model's own outputs are a good proxy for inference inputs; one could test this by comparing PAR against a control that resamples from the full-precision model at every round, and if the control performs equally well, the alignment effect is not actually driving the gain.","The TCR reparameterization is not limited to diffusion: any temporally iterated network with channel-wise activation outliers, such as video or recurrent models, could adopt the same timestep-averaged rescaling.","The DAQ likelihood-ratio choice between log and uniform quantizers could be extended to more than two quantizer families, or applied to other heavy-tailed activation layers beyond Softmax, such as certain normalization outputs.","The reported W4A4 gains hinge on keeping the input and output layers and the post-Softmax layer in higher precision; a fully 4-bit pipeline would need further work."],"forward_implications":["At 6-bit weights and activations, quantized DDIM on CIFAR-10 reaches FID 4.40, nearly matching the full-precision 4.14, so low-bit diffusion can run with little perceivable quality loss.","At 4-bit weights and activations, the method produces usable images (FID 6.38 on CIFAR-10) where prior PTQ methods collapse to FID over 200, opening W4A4 as a practical operating point.","The same three modules transfer across unconditional and conditional generation: results are reported on CIFAR-10, LSUN-Bedrooms, LSUN-Churches, and ImageNet with DDIM and LDM-4.","DAQ's per-timestep quantizer selection yields stable gains across softmax bit-widths from 8 down to 4 bits, whereas fixed log or uniform quantizers degrade sharply.","Because the method is post-training with a small calibration set, it avoids retraining and can be deployed on a single RTX 4090, per the paper's implementation details."],"supporting_citations":[{"why":"Supplies the BRECQ block-reconstruction base that the paper initializes and then refines with PAR.","marker":"Li et al. 2021"},{"why":"Provides the TFMQ-DM baseline and the experimental setting for LSUN and ImageNet that the paper follows.","marker":"Huang et al. 2024b"},{"why":"Provides the Q-Diffusion baseline and the CIFAR-10 evaluation setting the paper adopts.","marker":"Li et al. 2023b"},{"why":"Contributes the scale-reparameterization idea for vision transformers that TCR adapts to diffusion timesteps.","marker":"Li et al. 2023c"},{"why":"Gives the power-law maximum-likelihood fitting used by DAQ to decide log versus uniform quantization.","marker":"Clauset, Shalizi, and Newman 2009"},{"why":"Provides the log2 quantizer that DAQ applies to power-law-like post-Softmax activations.","marker":"Lin et al. 2022"}],"fun_headline_variants":["Adaptive quantization tames diffusion outliers for 4-bit inference","Three fixes make 4-bit diffusion quantization practical","Diffusion models hit 4-bit weights with FID 6.4 via adaptive PTQ","Timestep-channel reparameterization unlocks 4-bit diffusion","Power-law quantizer selection improves diffusion PTQ"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that sampling a fresh calibration set from the partially quantized model yields inputs representative of what the final quantized model will see during iterative inference, so if the early quantized outputs are badly distorted the recalibration could reinforce the distortion rather than correct it.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive quantization tames diffusion outliers for 4-bit inference","Three fixes make 4-bit diffusion quantization practical","Diffusion models hit 4-bit weights with FID 6.4 via adaptive PTQ","Timestep-channel reparameterization unlocks 4-bit diffusion","Power-law quantizer selection improves diffusion PTQ"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000514,"raw_usage":{"total_tokens":2564,"prompt_tokens":1079,"completion_tokens":1485,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":695,"completion_tokens_details":{"reasoning_tokens":1398}},"tokens_in":695,"tokens_out":1485,"duration_ms":9593,"temperature":1.0,"reasoning_tokens":1398,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:20:54.736787+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the PAR procedure for 4-bit weights and activations on CIFAR-10 with two variants: resampling from the quantized model (as proposed) and resampling from the full-precision model at every round. If the full-precision-resampled control matches or beats the proposed PAR FID of 6.38, the alignment mechanism is not responsible for the gain and the central claim about input mismatch is undercut.","supporting_citations":[{"cited_title":"R.; and Newman, M","cited_arxiv_id":null,"evidence_quote":"Gives the power-law maximum-likelihood fitting used by DAQ to decide log versus uniform quantization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the log2 quantizer that DAQ applies to power-law-like post-Softmax activations."}],"review_version":1}