{"id":"6eca923d-064f-44ec-8432-a49f92400cc6","arxiv_id":"2411.12306","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"DPQ applies product quantization with compressed codebooks and DDPM-loss calibration to compress DiT diffusion models to 1-bit weights with FID 14.03, far better than VQ's 91.71.","lead":"This paper compresses diffusion image generators down to 1-bit weights, about 24x smaller, by using product quantization with compressed codebooks and a calibration step that retrains the codebook on the diffusion objective. The method keeps image quality much closer to the full-precision model than vector quantization at the same bit-width.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"§3.2's codebook-pool size derivation requires projection storage ≤ pool, which implies m ≥ 16k; most DiT-XL/2 layers violate this, so the reported 24.57x size ratio may be overstated.","rationale":"The central claim is that DPQ reaches 1-bit compression with a 24.57x size reduction while keeping FID at 14.03, far better than VQ at the same nominal bit-width. I examined the storage accounting because the size ratio is the quantitative anchor of the abstract. The paper's derivation of N' rests on the assertion that projection storage is at most equal to the codebook pool. Working through the stated formats, this assertion is equivalent to m ≥ 16k, which fails for many DiT-XL/2 layers. The missing appendix makes this impossible to verify, and the paper does not clarify whether the projection is included in the reported size ratios. This is a concrete, checkable issue that directly affects whether the headline compression ratio is real. The reader's weakest assumption also flagged the missing appendix and projection storage, but did not identify the specific arithmetic inconsistency; hence partial agreement. The tau_c sensitivity is a secondary concern: no ablation of the 0.05 threshold is given, but the method reports a working configuration and the central comparison would likely survive a modest threshold change. I therefore keep the reader's CONDITIONAL verdict unchanged, as the issue is addressable with a corrected accounting or an explicit statement of what is included in 'codebook size.'","tokens_in":12322,"tokens_out":19170,"duration_ms":186169,"concrete_test":"Recompute the 1-bit DPQ size ratio for DiT-XL/2 as original FP32 bytes divided by (assignment bytes + C' pool bytes + \\hat C mapping bytes), using the formats stated in §3.2 (INT8 assignments, FP16 pool, INT16 mapping). Verify the inequality m ≥ 16k for each quantized layer's shape. If the corrected ratio is materially below 24.57x, or if the mapping table must be added to the reported size, the compression claim is overstated. Also recompute the 1-bit VQ size ratio under the same accounting to confirm the comparison is fair.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step for the headline compression ratio is the codebook-pool size formula in §3.2. The paper sets N' = mn/(16d^2) after asserting that the projection from the original PQ codebook to the pool requires storage at most equal to the pool itself. With k=256 INT8 assignments and d=8 (the 1-bit setting), the projection is a mapping table of (n/d)·k INT16 entries, costing 2nk/d bytes, while the pool costs mn/(8d) bytes. Requiring projection ≤ pool gives m ≥ 16k = 4096. Most DiT-XL/2 linear layers (qkv, fc1, adaLN) have m=1152 and fail this inequality, meaning the projection is larger than the pool by a factor of about 3.5 for those layers. If the projection is included in the Tables 1–3 size ratios, the reported 24.57x is not derivable from the stated formulas; if it is excluded, the 'over 24x reduction' is inflated relative to a deployable model. The missing appendix is precisely where this derivation must be checked, and the paper does not specify whether the projection is counted in 'codebook size.'","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Diffusion Product Quantization (DPQ), a post-training compression method for diffusion transformers. It replaces a single vector quantizer with product quantization over weight rows, compresses the resulting PQ codebooks via a usage-weighted 'codebook pool' with a similarity threshold tau_c, and calibrates the assignments and codebooks using activation-weighted reassignment and the DDPM loss. Experiments on DiT-XL/2 at ImageNet 256x256 report FID 14.03 at 1-bit with a 24.57x model-size reduction, versus FID 91.71 for a VQ baseline at 28.65x, and favorable comparisons against GPTQ, Q-DiT, and VQ4DiT; an additional DDIM/CIFAR-10 experiment shows applicability to a U-Net diffusion model.","tokens_in":12589,"tokens_out":8442,"duration_ms":85627,"significance":"If the size-accounting and comparisons hold, DPQ is a useful step for extreme low-bit diffusion-model compression: it demonstrates that PQ's larger effective codebook capacity translates into much better generative quality than VQ at 1-2 bits, while the codebook pool keeps storage overhead under control. The paper is transparent about limitations (no inference speedup, costly calibration), and the VQ-versus-DPQ comparison under the same calibration pipeline is a clean internal control. The main caveats are the deferred projection-storage derivation and the need for size-matched external comparisons.","major_comments":[{"comment":"The headline size ratios are not verifiable as written. The derivation of the projection storage is explicitly deferred to a missing appendix ('see appendix'), and the text does not state whether the projection table is included in the 'Size ratio', 'C ratio', or 'codebook size' columns. Please provide the full per-layer calculation, including the number of projection entries ((n/d)*k), their integer width, and a comparison with the pool size for the actual DiT-XL/2 layer shapes, and state unambiguously whether all compressed components are counted in the reported size ratios. Without this, the 24.57x claim is not derivable from the equations in §3.2.","section":"§3.2, Tables 1-3"},{"comment":"The comparison against VQ4DiT and the other external baselines is confounded by effective model size. At 2-bit/250 steps, DPQ has a size ratio of 12.08x, whereas VQ4DiT reports 15.75x; at 3-bit, DPQ has 9.92x versus VQ4DiT 10.59x. In both cases DPQ is less compressed, so part of the FID gain may be attributable to a larger model rather than to the quantization method. The paper should either match baselines at equal compressed size, plot FID/sFID versus size ratio for all methods, or explicitly discuss this trade-off when claiming to 'consistently outperform' other PTQ methods.","section":"Table 3"},{"comment":"The codebook-compression heuristic is load-bearing, but no sensitivity analysis is reported. The threshold tau_c is fixed at 0.05 with no ablation, and the paper does not report how many centroids are merged per layer or how the pool size N' relates to the number of surviving centroids. Since the size/quality trade-off of the whole method depends on this heuristic, an ablation over tau_c (or over N') is needed to support the claim that redundancy can be removed without significant quality loss.","section":"§3.2, Eq. (4)-(5)"}],"minor_comments":[{"comment":"The sentence 'While DPQ has a slightly larger size ratio due to the additional codebook size' contradicts Table 2, which shows DPQ with smaller size ratios (i.e., larger compressed models) at every bit-width; please correct the wording.","section":"§4.2"},{"comment":"The ablation table is under-specified: the three rows do not make clear which combination of PQ, DDPM loss, blockwise distillation, and codeword adjustment is being compared, and the text's claim that blockwise distillation is suboptimal is not backed by a displayed row.","section":"Table 5"},{"comment":"The expression c_{j,p} x_j is ambiguous because c_{j,p} is a d-dimensional row vector and x_j is a matrix or vector; please specify the shapes and the dimension over which the norm is taken.","section":"Eq. (7)"},{"comment":"The text says the listed bit-width does not represent actual bits per value, yet Tables 2 and 3 still label rows '1 bit', '2 bit', etc.; please add an effective-bits-per-value column or otherwise define the bit-width convention precisely.","section":"§4.1"},{"comment":"Several annotations in Figures 2 and 4 are too small or undefined; in particular, the right panel of Figure 2 should define the x-axis (diffusion step?) and clarify which blocks are selected for the MSE comparison.","section":"Figures 2 and 4"},{"comment":"The paper credits code from [21] and [25] but provides no repository URLs or version identifiers; please add reproducibility details for the quantization and calibration pipelines.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable candidate after revision. The internal VQ-versus-DPQ comparison with the same calibration pipeline is compelling, and the limitation statements are honest. However, the missing appendix and the size-matched comparison issue are load-bearing for the main claims, so I could not endorse acceptance in the current form. I would be willing to re-review once the projection-storage derivation and the sensitivity analysis are provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Shorthand: this is the PQ-for-DiT paper, and it is worth a serious look. The core result is that replacing VQ with PQ plus a compressed shared codebook pool and activation-aware recalibration gives a large, reproducible-looking gain at low bit-widths. On DiT-XL/2 at 1-bit, they report FID 14.03 versus 91.71 for their VQ baseline, with 24.57x size reduction. The head-to-head comparison in Table 2 uses the same calibration pipeline for both, so it is clean. The relative gain at 2-bit (6.84 vs 12.34) is also substantial.\n\nWhat is new is not PQ itself — Stock et al. and Martinez et al. already did PQ for neural-network weights — but the specific package for diffusion transformers: the codebook pool compression that keeps PQ codebooks from eating the compression ratio, and the calibration that re-assigns codes in the forward pass using activation error and updates codebooks with the DDPM loss. That is a legitimate engineering contribution, and the paper is honest about its limitations (no inference speedup, calibration cost). It borrows code from [21] and [25] with attribution, so no circularity concern.\n\nSoft spots. First, the external comparisons are not size-matched. At 2-bit, DPQ's effective size ratio is 12.08x while VQ4DiT's is 15.75x, so DPQ is spending about 30% more storage for its better FID. Some of the gain may be capacity, not method. The paper needs a rate-distortion-style comparison or at least a discussion of this confound. Second, the codebook merge threshold tau_c is load-bearing and there is no sensitivity analysis; a referee should ask whether the results are stable around tau_c=0.05. Third, the appendix containing the projection memory derivation and 512x512 results is missing from this version. That matters because the compression ratio depends on whether the projection table is counted. The stress-test note's specific arithmetic, though, is wrong: with INT16 projection of (n/d)k centroids and an FP16 pool of size mn/(16d^2), projection <= pool iff m >= 2k, not m >= 16k. DiT-XL/2 layers satisfy m >= 512, so the note's claim that most layers violate the inequality does not hold up. The missing derivation is still a real completeness issue.\n\nMy read: the main claim is likely correct. The method is practical, the internal comparison is fair, and the issues are fixable with more experiments and a complete appendix.\n\nRecommendation: send it to peer review. A competent referee can spend time productively on the size-matched comparison and the codebook compression ablation. If I were writing on low-bit diffusion quantization, I would cite it.","headline":"A useful, mostly sound adaptation of product quantization to DiT at extreme low bit-widths; the internal VQ comparison is convincing, but the external baselines are not size-matched and one derivation is stashed in a missing appendix.","tokens_in":13134,"tokens_out":7247,"would_cite":true,"duration_ms":77402,"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":"Product quantization lets 1-bit diffusion models generate sharp images while cutting model size over 24 times.","keywords":["product quantization","diffusion models","model compression","vector quantization","low-bit quantization","DiT","codebook compression","ImageNet generation"],"falsifier":"Reproduce the 1-bit DiT-XL/2 experiment on ImageNet with two variants: the published DPQ pipeline, and the same pipeline with $\\tau_c = 0$ (no centroid merging, with the bit budget held constant by reducing the subspace dimension $d$). If the unmerged variant's FID is substantially better than DPQ's reported 14.03, the codebook-pool merging step, not product quantization itself, is where generative quality is lost. Also verify the projection-storage calculation promised in the appendix, since the claimed size ratio depends on it.","tokens_in":1865,"feed_emoji":"🗜️","tokens_out":2715,"duration_ms":84346,"temperature":0.7,"pith_summary":"This paper claims that product quantization, not ordinary vector quantization, is the right tool for compressing diffusion models to very low bit-widths. Applied to the DiT image-generation transformer on ImageNet, the proposed DPQ method reaches 1-bit weights and a 24.57x model-size reduction while keeping FID at 14.03, whereas plain vector quantization collapses to 91.71 at the same bit-width. The reason, the paper argues, is that product quantization reconstructs high-dimensional weight vectors far more precisely and gives an exponentially larger effective codebook, which matters because diffusion models run many forward passes and accumulate quantization error. The paper also shows that the codebooks themselves must be compressed, and introduces a usage-frequency-based merging step plus an end-to-end calibration that adjusts assignments on activations and fine-tunes codebooks on the diffusion loss.","feed_headline":"1-bit diffusion compression: FID 92 to 14","feed_subtitle":"Product quantization plus codebook trimming keeps 1-bit DiT images coherent on ImageNet at 24x smaller size.","key_machinery":"The load-bearing object is the product-quantization decomposition: the weight matrix is split into sub-vectors of dimension $d$, and for each subspace position there is a separate codebook $C \\in \\mathbb{R}^{N \\times k \\times d}$, so the effective number of representable vectors is $k^N$ while the codebook memory grows only linearly. Around that core sit two mechanisms the paper adds: a codebook pool (Eqs. 4-5) that ranks centroids by assignment frequency, merges any centroid within L2 distance $\\tau_c = 0.05$ of a more important one, and stores the result in FP16 so the compressed codebook stays within the designed size budget; and a two-part calibration (Eqs. 7-8) that reassigns codewords in the forward pass using the activation-weighted error $\\|Wx - W'x\\|^2$ and updates codewords in the backward pass through the DDPM loss $L_{\\mathrm{DDPM}}$.","core_discovery":"The central claim is that existing vector-quantization compression fails for diffusion models at low bit-widths not because quantization is wrong, but because a single shared codebook cannot represent high-dimensional sub-vectors precisely, and because errors compound over the iterative denoising steps. DPQ replaces the shared codebook with product quantization: each weight row is split into subspaces, each subspace has its own codebook, and the effective number of codewords grows exponentially while codebook storage grows only linearly. To stop the codebooks from canceling the compression gain, DPQ stores centroids in FP16 and merges similar, rarely used centroids into a codebook pool according to a distance threshold, keeping the codebook no more than a quarter of the assignment size. Finally, calibration updates the assignments in the forward pass by minimizing the activation-weighted reconstruction error $\\|Wx - W'x\\|^2$ and fine-tunes the codebooks in the backward pass with the DDPM noise-prediction loss. On DiT-XL/2 at 256x256, the method reports FID 6.84 at 2 bits and 14.03 at 1 bit, with size ratios of 12.08x and 24.57x respectively, consistently beating the vector-quantization and post-training-quantization baselines compared in the paper.","pith_inferences":["If the reported 1-bit result reproduces, product quantization should be tried on larger text-to-image DiTs, where per-subspace codebooks could absorb higher weight dimensionality without the exponential codebook blow-up that defeats ordinary VQ.","The error-accumulation analysis suggests a testable extension: calibrating against a subset of denoising timesteps might cut the roughly 5-hour calibration cost while preserving the FID gains, since the paper shows saturation after 5 epochs.","Because DPQ does not accelerate inference, a natural next step is combining DPQ with activation quantization or pruning; the codebook structure may also enable lookup-table-based fast inference.","The codebook-pool merging rule is an importance-based pruning inside the codebook; a principled alternative would select centroids to merge by their contribution to reconstruction error on real activations rather than usage frequency alone."],"forward_implications":["At 1 bit, the DPQ-compressed DiT-XL/2 retains usable generation (FID 14.03, IS 110.23, precision 0.6863), where ordinary VQ collapses to FID 91.71.","At 2 bits, DPQ gives FID 6.84 and a 12.08x size reduction, beating GPTQ, Q-DiT, and VQ4DiT at 250, 100, and 50 sampling steps.","The method transfers to U-Net diffusion models: on DDIM/CIFAR-10 it reaches FID 5.19 at 4 bits and 7.01 at 2 bits.","Compressing the codebook itself is mandatory for PQ in this setting; keeping the codebook within one quarter of the assignment size is what preserves the overall size ratio.","Calibrating with the DDPM loss outperforms block-wise distillation, and forward-pass codeword reassignment adds a further gain in the reported ablation."],"supporting_citations":[{"why":"Introduces product quantization, the central compression mechanism the paper adapts to diffusion models.","marker":"[17]"},{"why":"Supplies the DiT-XL/2 model that all compression experiments target.","marker":"[25]"},{"why":"Defines the DDPM noise-prediction loss used in the calibration objective.","marker":"[14]"},{"why":"VQ4DiT is the vector-quantization baseline whose compression results DPQ is compared against and whose block-wise calibration is contrasted.","marker":"[5]"},{"why":"GPTQ provides a post-training quantization baseline for transformer weights in the comparison tables.","marker":"[8]"},{"why":"Q-DiT provides a second post-training quantization baseline for diffusion transformers.","marker":"[3]"},{"why":"Supplies the product-quantization implementation the paper builds on and the observation that codeword reassignment is often omitted in calibration.","marker":"[21]"}],"fun_headline_variants":["1-bit diffusion: 24x smaller, FID 14 on ImageNet","Product quantization: 1-bit diffusion at 24x size cut","Diffusion compressed to 1 bit via product quantization","Codebook trimming yields 1-bit diffusion with FID 14"],"cache_read_input_tokens":15232,"weakest_assumption_plain":"The claimed 24x compression at 1 bit assumes that merging similar, rarely used codebook centroids with a fixed threshold $\\tau_c = 0.05$ preserves generative quality, a heuristic the paper does not analyze for sensitivity.","fun_headline_variants_meta":{"raw":{"variants":["1-bit diffusion: 24x smaller, FID 14 on ImageNet","Product quantization: 1-bit diffusion at 24x size cut","Diffusion compressed to 1 bit via product quantization","Codebook trimming yields 1-bit diffusion with FID 14"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000896,"raw_usage":{"total_tokens":3879,"prompt_tokens":980,"completion_tokens":2899,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":2824}},"tokens_in":596,"tokens_out":2899,"duration_ms":21797,"temperature":1.0,"reasoning_tokens":2824,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:41:20.520233+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the 1-bit DiT-XL/2 experiment on ImageNet with two variants: the published DPQ pipeline, and the same pipeline with $\\tau_c = 0$ (no centroid merging, with the bit budget held constant by reducing the subspace dimension $d$). If the unmerged variant's FID is substantially better than DPQ's reported 14.03, the codebook-pool merging step, not product quantization itself, is where generative quality is lost. Also verify the projection-storage calculation promised in the appendix, since the claimed size ratio depends on it.","supporting_citations":[{"cited_title":"Prod- uct quantization for nearest neighbor search","cited_arxiv_id":null,"evidence_quote":"Introduces product quantization, the central compression mechanism the paper adapts to diffusion models."},{"cited_title":"GPTQ: Accurate post-training quantization for gener- ative pre-trained transformers","cited_arxiv_id":null,"evidence_quote":"GPTQ provides a post-training quantization baseline for transformer weights in the comparison tables."},{"cited_title":"Permute, quantize, and fine-tune: Efficient compression of neural networks","cited_arxiv_id":null,"evidence_quote":"Supplies the product-quantization implementation the paper builds on and the observation that codeword reassignment is often omitted in calibration."}],"review_version":1}