{"id":"60597b51-4cf6-46bd-ab90-cdb4054f7aab","arxiv_id":"2412.14628","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Qua2SeDiMo learns layer-level quantization sensitivity with a GNN and constructs sub-4-bit mixed-precision weights for several diffusion models with FID close to or better than full precision.","lead":"Qua2SeDiMo uses graph neural networks to learn which layers of a diffusion model can tolerate aggressive quantization and which cannot, then builds mixed-precision configurations averaging roughly 3.4 to 3.9 bits per weight. If the results hold, this offers a calibration-free path to sub-4-bit weight compression for image generation models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central attribution—using GNN node-embedding L1 norms as per-layer sensitivity scores (Eq. 6 and text after Eq. 6)—is unvalidated and not implied by the ranking loss, which supervises only the norm of the averaged graph embedding.","rationale":"The reader's weakest assumption coincides with the most load-bearing concern. The ranking loss in Eq. 6 supervises only the norm of the averaged graph embedding; it does not constrain per-node embeddings to be faithful attributions. The paper's inference that averaging 'forces' the GNN to learn per-node contributions is not justified: a GNN can achieve high SRCC/NDCG on graph-level FID by encoding global configuration statistics. The supplementary validation (Fig. 18) reports only graph-level SRCC/NDCG, not any node-level or counterfactual fidelity, so the attribution mechanism is entirely unvalidated. The selection procedure independently optimizes per-node scores, which requires additivity and ignores interactions; this is an additional unsupported assumption. The reported FID/CLIP tables are specific and falsifiable, and I see no internal numerical contradiction that would warrant rejection on the evidence in the paper, aside from minor typos (e.g., the mislabeled Q-Diffusion OAQ row in Table 3). If the proposed node-level perturbation test shows no correlation between GNN scores and true FID effects, the paper's central contribution (quantifiable sensitivity and the derived insights) collapses, and the verdict should be REJECT. Until then, CONDITIONAL remains the right call, requiring code release, error bars, and the attribution validation. Therefore I recommend UNCHANGED relative to the reader's verdict.","tokens_in":27079,"tokens_out":7717,"duration_ms":119110,"concrete_test":"On PixArt-α, take the trained Qua2SeDiMo predictor that produced the 3.4-bit configuration (Hybrid Block-level, Fig. 4). Hold out 10% of the 372 sampled configurations for validation. Select 30 weight layers spanning t-Embed, attention Q/K/V/out, feedforward, and Out-Proj. For each, flip that layer alone to the quantization setting that minimizes (or maximizes) its node-embedding norm (or subgraph root norm) while holding all other layers at the chosen 3.4-bit configuration. Generate 1k images per variant using the paper's COCO-2017 protocol and measure FID. Compute the Spearman correlation between the GNN's per-node score difference and the measured FID difference. If rho is not significantly positive (e.g., < 0.3), the node norms are not faithful sensitivity scores and the central attribution claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the leap from Eq. 6 to the claim (Methodology, after Eq. 6) that 'we are able to treat the scalar norm of the node embedding ... as a numerical score.' The ranking loss Lrank is applied to ||h^m_G||_1, the L1 norm of the graph embedding obtained by averaging node embeddings (Eq. 5). A GNN can satisfy the ranking constraint on this aggregate norm without individual node norms carrying causal information: it can encode y in global statistics (e.g., average bit precision, method mix) or in a few nodes with large, aligned embeddings, leaving the remaining node norms meaningless. Nothing in the loss forces per-node norms to be additive, counterfactual, or even monotone in the true FID effect of changing that node's quantization. The selection procedure then independently maximizes each node's (or subgraph root's) norm, implicitly assuming a sum-separable, faithful attribution, which is never checked against direct single-layer perturbations. The scarcity of training data (only 340-447 configurations for search spaces of size 6^#W with #W up to 803, Table 2) makes it more likely the GNN latches onto global summaries rather than per-node sensitivity, so the extracted insights (DiT prefers K-Means, U-Net ResBlocks fragile, sinusoidal layer-wise scores) and the constructed configurations are not reliably tied to true quantization sensitivity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Qua²SeDiMo, a mixed-precision post-training quantization (PTQ) framework for diffusion-model denoisers. The denoiser is represented as a directed acyclic graph over weight layers; a GNN is trained on a few hundred randomly sampled quantization configurations using a ranking loss on the L1 norm of the averaged graph embedding (Eq. 6). The authors then interpret the L1 norm of each node embedding as a per-layer sensitivity score and greedily select per-layer quantization methods and precisions (K-Means channel-wise, K-Means tensor-wise, UAQ; 3 or 4 bits) to optimize a FID-plus-average-bit objective. The resulting configurations are evaluated on PixArt-α, PixArt-Σ, Hunyuan-DiT, SDXL, SDv1.5 and DiT-XL/2, with sub-4-bit average precisions (e.g., 3.4-bit on PixArt-α) reported and compared against Q-Diffusion, TFMQ-DM, and ViDiT-Q in FID, CLIP, and a small human-preference study.","tokens_in":27376,"tokens_out":6807,"duration_ms":47878,"significance":"If the per-node sensitivity scores are valid, the paper contributes a scalable, calibration-free way to obtain mixed-precision sub-4-bit weight-only PTQ configurations for both U-Net and DiT denoisers. The final reported configurations are measured, not fitted, and the comparisons in Tables 3 to 6 are extensive, covering three activation-precision levels, multiple architectures, and a human-preference study. The claimed architectural insights (time-embedding importance, U-Net ResBlock fragility, DiT preference for K-Means) are also potentially useful. However, the central attribution from a graph-level ranking loss to per-node scores is not established, and the experimental protocol has a serious FID-protocol inconsistency between search and final evaluation, with no error bars or multiple seeds. These issues are fixable with additional validation experiments, but they currently weaken both the interpretability claim and the statistical reliability of the headline numbers.","major_comments":[{"comment":"The central claim that the L1 norm of a node embedding is a valid per-layer sensitivity score is not implied by the loss in Eq. (6). The ranking term supervises the scalar norm of the averaged graph embedding ||h^m_G||_1 (Eq. 5), which can be maximized by encoding global summaries such as average bit precision or method mix, or by concentrating signal in a small number of nodes; nothing in the loss forces per-node norms to be additive, counterfactual, or monotone in the true FID effect of changing that node's quantization. The subsequent greedy selection that maximizes each node's score independently therefore rests on an unvalidated attribution. I request a direct validation: hold out random configurations, measure the change in FID when a single layer is toggled from 4-bit to 3-bit (or its method is changed) while other layers are held fixed, and report the rank correlation between these measured marginal effects and the GNN node scores; also compare the greedy per-node choice against a joint search on a small subnetwork. Without this, the qualitative insights in the 'Extracted Insights' section and the claim that the method 'can learn to assign the optimal configuration to each layer' are not supported.","section":"Methodology, Eq. (6)"},{"comment":"The FID protocol changes between search and final evaluation. The search-time corpus in Table 2 and the Pareto frontiers in Figure 4 use 1k-image FID computed on the COCO 2017 validation split, while Tables 3, 4, 6 and 8 report 10k-image FID on COCO 2014. The gap is not cosmetic: the full-precision PixArt-α FID is 99.67 at 1k images (Table 2) versus 34.05 at 10k images (Table 3). Because the predictor is trained on the 1k scores, λ is tuned for that protocol, and the best configurations are selected from these noisy labels, the reported 10k FID gains may be affected by FID-estimation noise and dataset mismatch. Please report search-time FID using the same 10k COCO-2014 protocol, or at minimum show that the selected configurations are stable when the search objective is re-evaluated on the final protocol, and include multiple seeds or bootstrap intervals.","section":"Experimental Results and Discussion; Tables 2 vs 3–6"},{"comment":"The paper evaluates many predictor variants (three ranking losses, Op-level vs Block-level, two objectives) and then selects a small number of configurations for detailed investigation, marked with purple circles in Figures 4 and 9. The selection rule for these circles is not stated. If the choice was made after inspecting test-set FID, the headline comparisons in Tables 3–6 are a form of best-of-many selection and need correction, e.g., report the full distribution of FID across all variants, or pre-register a selection criterion based only on validation performance. This matters because the reported sub-4-bit results are the best among many searched configurations rather than the output of a single prespecified model selection rule.","section":"Figures 4 and 9, purple-circle selection"},{"comment":"The predictor validation metrics in Fig. 18 are reported as ensemble weights, but there is no analysis of the error of the predicted FID itself or of the sensitivity scores. The search space has size 6^#W with #W up to 803, while only 340–447 samples per model are used. A predictor can rank the sampled configurations well while the per-node score attribution is wrong, because the ranking loss only evaluates graph-level order. Please report the calibration and error of the graph-level predictor on held-out configurations and, more importantly, the per-node validation described in the first major comment.","section":"Supplementary, Fig. 18"}],"minor_comments":[{"comment":"In the W4A6 block of Table 3, the row labeled 'Q-Diffusion OAQ W4A8' (FID 70.96) should presumably be labeled 'Q-Diffusion OAQ W4A6'; as printed, the same precision label appears in two different activation-precision blocks.","section":"Table 3"},{"comment":"The sentence 'K-Means C quantizes each applies output channel cout separately' contains a typo and should read 'K-Means C quantizes each output channel cout separately'.","section":"Methodology, 'Operation-Level Sensitivity via Graphs'"},{"comment":"Table 1 reports FID values for PixArt-α/Σ that are close to but not identical to the corresponding values in Table 3, yet the caption does not specify the prompt set or generation protocol; please reconcile these numbers with the COCO-2017/COCO-2014 protocol distinction.","section":"Table 1 and Tables 2–3"},{"comment":"The stacked bar plots in Figures 8 and 15 show quantization-method distributions for a single constructed configuration; statements such as 'DiT blocks have a slight preference for K-Means' would be more convincing if accompanied by variability across configurations, ensemble folds, or random seeds.","section":"Extracted Insights and Figures 8/15"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for the venue and the empirical effort is substantial. My main concern is not the authors' integrity but the unvalidated leap from graph-level ranking loss to per-node sensitivity scores; this is the load-bearing component of the method and should be addressed with direct single-layer perturbation experiments before publication. The authors should also clarify the selection protocol for the reported best configurations and reconcile the 1k/10k FID protocols."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about this paper is that the headline numbers are probably real. The authors construct sub-4-bit mixed-precision weight-only PTQ configurations for PixArt-α/Σ, Hunyuan-DiT, SDXL, and DiT-XL/2, evaluate them on held-out FID/CLIP, and they beat prior W4 methods like Q-Diffusion, TFMQ-DM, and ViDiT-Q. Those comparisons use 10k generated images and standard COCO prompts, and the tables are extensive. If those numbers hold up on independent reimplementation, this is a genuinely useful result for deploying diffusion models.\n\nWhat's new is the surrogate-based search: a GNN regressor is trained on a few hundred sampled quantization configs (6 choices per weight layer: K-Means C/A or UAQ at 3/4 bits), and then per-node or per-subgraph embedding norms are used as sensitivity scores to greedily pick the best setting for each layer. That's a real departure from Hessian/activation heuristics. The paper also does a good job of reporting search-space statistics, predictor validation performance, and qualitative insights (DiT prefers K-Means, U-Net ResBlocks fragile, t-Embed important).\n\nThe soft spots are real but not fatal. The central attribution step—treating the L1 norm of a GNN node embedding as a faithful per-layer sensitivity score—is never directly validated. The ranking loss in Eq. 6 supervises only the norm of the averaged graph embedding, and nothing forces individual node norms to be monotone or additive in the true effect of changing that node's quantization. With only 340–447 training configs for spaces of size 6^#W, the GNN could easily satisfy the ranking constraint using global statistics. A simple ablation—compare the selected configs against random search on the same budget, or measure single-layer perturbation effects—would settle this. The paper does compare against its own random Pareto frontier, though, so the selection is not completely unfounded.\n\nSecond, the FID reporting is inconsistent: Table 2 uses 1k-image FID (FP PixArt-α at 99.67) while Tables 3/4 use 10k-image FID (FP at 34.05). The gap likely comes from the different evaluation sets, but the paper never reconciles it, and it makes the supplementary tables hard to interpret. Third, no code or checkpoints are available, so the specific sub-4-bit configs cannot be verified directly. Fourth, MixDQ is cited but not compared, which is a missed baseline.\n\nOverall, the paper is a serious piece of work. The measured configs and the empirical comparisons are the load-bearing part, and they are not forced by the surrogate. The attribution insight needs validation, but the core result is plausible and important enough to deserve a serious referee.\n\nRecommendation: send it to peer review, but require the authors to validate the attribution (single-layer perturbation or a random-search control), reconcile the FID evaluations, and release code before acceptance.","headline":"Measured sub-4-bit configs are the real contribution; the GNN attribution story is the unvalidated soft spot.","tokens_in":27944,"tokens_out":4133,"would_cite":true,"duration_ms":33433,"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":"This paper claims that quantization sensitivity of individual layers in diffusion image denoisers can be read directly off a graph neural network, and that these readings can be used to build sub-4-bit mixed-precision weight…","keywords":["post-training quantization","mixed-precision quantization","diffusion models","graph neural networks","sensitivity attribution","sub-4-bit quantization","text-to-image generation","FID optimization"],"falsifier":"Measure, for one denoiser, the paper's predicted per-layer sensitivity scores for its optimal sub-4-bit configuration, then independently quantize each single layer to 3-bit UAQ while leaving all other weights at full precision and record the FID change. If the Spearman correlation between the predicted scores and these single-layer FID deltas is not clearly positive, the attribution that the entire method rests on is falsified.","tokens_in":26846,"feed_emoji":"🎨","tokens_out":9043,"duration_ms":62722,"temperature":0.7,"pith_summary":"This paper claims that the quantization sensitivity of every weight layer in a diffusion image denoiser can be read directly off a graph neural network: a GNN trained to predict end-to-end FID from random mixed-precision configurations yields node embeddings whose L1 norms serve as per-layer sensitivity scores. The authors use these scores to assign each layer one of six choices (3/4-bit precision times K-Means channel-wise, K-Means whole-tensor, or uniform affine quantization), producing sub-4-bit weight configurations — 3.4 bits on PixArt-alpha, 3.9 on PixArt-Sigma, 3.65 on Hunyuan-DiT, 3.7 on SDXL — that match or beat the full-precision model's FID and CLIP while using no calibration data. If true, this would make sub-4-bit weight-only post-training quantization practical for both U-Net and DiT denoisers, and would replace hand-crafted heuristics with automatically extracted, architecture-specific sensitivity rankings. The same surrogate also yields qualitative insights: DiT models prefer cluster-based quantization, U-Net ResNet blocks are the fragile components, and time-embedding layers are consistently the most sensitive.","feed_headline":"Diffusion models quantize below 4 bits without quality loss","feed_subtitle":"Layer-wise graph scores build 3.4-bit PixArt-alpha and 3.7-bit SDXL, beating 4-bit baselines.","key_machinery":"The load-bearing mechanism is the GNN-attribution identity: after a graph neural network is trained to regress a denoiser's end-to-end quality (negative FID, optionally penalized by average bit precision), the L1 norm of a node's embedding at hop $m$, $\\|h^m_v\\|_1$, is interpreted as the sensitivity score of the $m$-hop subgraph rooted at weight layer $v$. The objective adds a ranking loss to the regression loss at every GNN layer, applied to the norms of the graph-level embeddings, so the norms are forced to order the sampled quantization configurations by quality. Because message passing spreads a node's features across its neighborhood, the hop-$m$ norm of a root node aggregates the quantization choices of an entire block structure (self-attention, feedforward, ResNet block, time-embedding module). Selecting, for each node or subgraph root, the quantization setting that maximizes this norm converts the exponential search space into per-node decisions and yields the final mixed-precision configuration.","core_discovery":"The central discovery is that quantitative per-layer sensitivity can be derived from a GNN surrogate rather than from Hessians or activation outliers. Encoding each quantizable weight layer as a graph node whose features describe quantization method, precision, operation type, and position, the paper trains an ensemble of GNNs to regress a combined objective $y = -\\mathrm{FID} - \\lambda\\,\\overline{\\mathrm{Bits}}$ from only 340–447 sampled configurations per denoiser. A ranking loss (differentiable Spearman rho, LambdaRank/NDCG, or their sum) is applied at every message-passing layer to the L1 norm of the graph-level embedding, which compels the embedding norms to rank configurations by quality; the paper then treats $\\|h^m_v\\|_1$ as the sensitivity score of the $m$-hop subgraph rooted at layer $v$. Enumerating the six quantization options per node and keeping the highest-scoring setting yields mixed-precision configurations at 3.4–3.9 average weight bits that achieve lower FID than the full-precision model on PixArt-$\\alpha$, SDXL, Hunyuan-DiT, and DiT-XL/2, and that outperform Q-Diffusion, TFMQ-DM, and ViDiT-Q at W4A8/W4A6. The paper further claims architecture-level insights from the score distributions: time-embedding and output-projection layers are consistently sensitive, DiT blocks favor K-Means clustering while U-Nets favor uniform affine quantization, and U-Net ResNet blocks are more sensitive than Transformer blocks.","pith_inferences":["A testable transfer claim the paper leaves implicit: because the GNN encodes only quantization features and architecture position, a predictor trained on one PixArt model may rank configurations for its sibling model without retraining, which would slash the evaluation budget.","One could validate the attribution mechanism itself by ablating each layer to a fixed low-bit setting and correlating measured FID deltas with predicted scores; the paper reports ensemble validation accuracy but not this direct check.","The same surrogate-search recipe should transfer to video diffusion models or other generative architectures with hundreds of layers, since it needs only a few hundred FID evaluations and the compute is dominated by sampling, not GNN training.","The paper briefly notes a sinusoidal pattern of layer-wise sensitivity across transformer depth; an extension would be to use these scores to guide block pruning or layer dropping, analogous to depth-pruning policies in large language models."],"forward_implications":["Sub-4-bit weight-only post-training quantization becomes practical: PixArt-alpha runs at 3.4 average weight bits, PixArt-Sigma at 3.9, Hunyuan-DiT at 3.65, SDXL at 3.7, and DiT-XL/2 at 3.5 bits, with FID and CLIP scores comparable to or better than the W16A16 reference.","The extracted sensitivity rankings tell developers where to spend bits: time-embedding layers and output projections are consistently fragile, ResNet blocks in U-Nets are more sensitive than Transformer blocks, and patchify/skip-connection layers matter in DiTs.","Architecture-aware quantization-method choice is quantified: DiT-family models prefer K-Means clustering, U-Nets prefer uniform affine quantization (UAQ), and mixing methods by layer yields better configurations than any single method applied uniformly.","Pairing these weight configurations with online 6-bit activation quantization (no calibration set) beats calibration-based baselines (Q-Diffusion, TFMQ-DM, ViDiT-Q) on both FID and CLIP for PixArt-alpha/Sigma and Hunyuan-DiT."],"supporting_citations":[{"why":"Provides the TFMQ-DM baseline that Qua2SeDiMo must beat, along with earlier evidence that time-embedding layers are quantization-sensitive.","marker":"Huang et al. 2024"},{"why":"Provides the Q-Diffusion baseline, the code base extended by this work, and the long-residual-connection handling that informs the subgraph design.","marker":"Li et al. 2023"},{"why":"Supplies the ViDiT-Q baseline for DiT quantization, which also uses ZeroQuant online activation quantization.","marker":"Zhao et al. 2024"},{"why":"Supplies the online, patch-based activation quantization used by both ViDiT-Q and Qua2SeDiMo.","marker":"Yao et al. 2022"},{"why":"Supplies the differentiable Spearman rank-correlation loss used as one ranking objective.","marker":"Blondel et al. 2020"},{"why":"Supplies the LambdaRank loss that maximizes NDCG and prioritizes correctly ranking high-quality configurations.","marker":"Burges 2010"},{"why":"Provides the GATv2 attention message-passing layer used inside the GNN regressor.","marker":"Brody, Alon, and Yahav 2022"},{"why":"Provides the K-fold predictor-ensemble training and score re-scaling scheme used to stabilize the sensitivity estimates.","marker":"Mills et al. 2024"}],"fun_headline_variants":["GNN scores guide 3.4-bit diffusion quantization","Diffusion quantization hits sub-4 bits, beats 4-bit baselines","Sub-4-bit diffusion quantization beats full-precision FID","Explainable GNN picks layer precisions for 3.4-bit diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the L1 norm of a GNN node embedding, learned from a few hundred randomly sampled quantization configurations, faithfully measures how much that layer or block actually contributes to end-to-end image quality; if the embedding norms do not track true sensitivity, the constructed configurations and extracted insights would not be reliable.","fun_headline_variants_meta":{"raw":{"variants":["GNN scores guide 3.4-bit diffusion quantization","Diffusion quantization hits sub-4 bits, beats 4-bit baselines","Sub-4-bit diffusion quantization beats full-precision FID","Explainable GNN picks layer precisions for 3.4-bit diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000895,"raw_usage":{"total_tokens":3943,"prompt_tokens":1118,"completion_tokens":2825,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":734,"completion_tokens_details":{"reasoning_tokens":2749}},"tokens_in":734,"tokens_out":2825,"duration_ms":21453,"temperature":1.0,"reasoning_tokens":2749,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:03:32.056884+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure, for one denoiser, the paper's predicted per-layer sensitivity scores for its optimal sub-4-bit configuration, then independently quantize each single layer to 3-bit UAQ while leaving all other weights at full precision and record the FID change. If the Spearman correlation between the predicted scores and these single-layer FID deltas is not clearly positive, the attribution that the entire method rests on is falsified.","supporting_citations":[],"review_version":1}