{"id":"07441fc0-f0e6-4c41-850d-780095134539","arxiv_id":"2412.00381","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"DogLayout uses a denoising diffusion GAN with 4 to 12 timesteps to generate layout boxes and discrete labels, sampling up to 175 times faster than LayoutDM, with mixed FID results across tasks.","lead":"DogLayout fuses a GAN with a short diffusion process to generate graphic layouts, including discrete element labels, in a few sampling steps. It reports up to 175 times faster sampling than the LayoutDM diffusion baseline and lower element overlap on one PubLayNet task.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The discrete-label extension rests on an unvalidated continuous relaxation: Gaussian noise on one-hot prototypes has no categorical semantics, and the final softmax/argmax may yield arbitrary labels.","rationale":"The reader's weakest assumption matches my own: the discrete label handling is the most fragile part of the argument. The speed and overlap claims are empirical and can be checked from the tables, but the discrete label mechanism is what makes the method novel, and it is currently a heuristic. The paper claims to solve GANs' discrete data problem, yet it provides no theoretical derivation from discrete diffusion principles and no evaluation of label quality beyond aggregate FID. The proposed test would settle whether the relaxation preserves categorical semantics. I therefore keep the CONDITIONAL verdict: the paper is promising but needs either a principled justification or explicit evidence that the discrete labels are valid.","tokens_in":10690,"tokens_out":4859,"duration_ms":51944,"concrete_test":"Train DogLayout on a synthetic categorical dataset with known label dependencies and compare its generated label sequences to the true distribution using KL divergence and element-type accuracy against LayoutDM. Additionally, examine the learned l0 vectors: compute their entropy and distance to the nearest one-hot vector; if labels are semantically meaningful, the predicted l0 should concentrate near one-hot vertices, and the distribution of noisy labels lt should be approximately Gaussian. A further check: run a variant that samples from the softmax categorical distribution instead of argmax and see if FID/overlap change substantially, indicating that the final decoding rule is critical.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"DogLayout's core novelty is generating discrete labels by adding a diffusion process to GANs. Yet the method treats the one-hot label l0 as a real-valued prototype, corrupts it with Gaussian noise, and reconstructs a continuous vector l0, finally sampling L = argmax softmax(l0) (Eq. 9). This is not derived from a discrete-state diffusion model; there is no proof that the Gaussian forward process on one-hot vectors yields a meaningful noisy categorical state, nor that the reverse GAN learns the correct categorical posterior. The generator's training objective is the adversarial loss on continuous xt-1 (Eq. 8); no term encourages the predicted l0 to be a calibrated probability vector or to match the true label distribution. The discriminator's reconstruction loss is computed on real data and does not directly supervise the generator's label outputs. Consequently, the 'new method to generate discrete label data' reduces to an ad hoc continuous relaxation. This is load-bearing because the model's ability to handle discrete labels is the central claim and the basis for unconditional/completion tasks; the reported weakness on completion (FID 16.8 vs LayoutDM 8.47 on PubLayNet) is consistent with this component failing. Without a formal justification or empirical verification of label semantics, the discrete-data contribution is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DogLayout, a denoising-diffusion GAN for layout generation. The method couples a GAN generator with a small number of denoising steps (T=4/8/12) so that sampling is much faster than discrete diffusion baselines, and it claims that this setup also lets a GAN generate discrete label data directly, by treating one-hot labels as continuous class prototypes that are corrupted and denoised and then converted to discrete labels via argmax over softmax outputs. The paper evaluates DogLayout on Rico and PubLayNet for conditional generation, completion, and unconditional generation, reporting FID, MaxIoU, overlap, alignment, and per-sample sampling time, plus a user study. The main claims are a 175x sampling-speed improvement over LayoutDM, a reduction in overlap on PubLayNet from 16.43 to 9.59, and an expansion of GAN-based layout models to tasks that require discrete label generation.","tokens_in":10934,"tokens_out":6836,"duration_ms":67575,"significance":"If the claims were fully supported, the practical significance would be moderate: a fast GAN-based layout generator that handles discrete labels as well as continuous boxes would be useful for interactive and resource-constrained layout applications. The speed evidence is the paper's strongest point; the 175x figure follows directly from Table 3 for the T=4 conditional tasks. However, the discrete-label contribution is the conceptual core of the paper, and it currently rests on a continuous relaxation of one-hot labels that is neither derived nor directly validated. The quality evidence is also mixed: against LayoutDM, DogLayout is worse on most FID and all MaxIoU entries in Table 2, and the headline overlap improvement comes from a single conditional task. Code availability and the use of standard datasets are positive elements, but they do not by themselves resolve the load-bearing technical concerns.","major_comments":[{"comment":"The central claim that DogLayout 'enables the generation of discrete label data' is not backed by a derivation or direct empirical validation. Equation (9) treats the one-hot label l0 as a continuous class prototype, applies Gaussian noise, and then takes argmax over a softmax of the reconstructed prototype. No argument is given for why Gaussian corruption of a one-hot vector defines a meaningful categorical forward process or why the reverse GAN learns the correct posterior over labels. The training objective in Eq. (8) contains no term that supervises the reconstructed label vector to be calibrated or matched to the true label distribution, and the discriminator sees only continuous xt-1, so it cannot directly enforce label semantics. Because discrete-label generation is the first contribution and underlies the unconditional and completion tasks, this is load-bearing. I recommend either providing a formal connection to discrete-state diffusion or an empirical validation: per-class accuracy or confusion matrices on generated labels, a distributional comparison between generated and real label histograms, and an ablation comparing Eq. (9) with a Gumbel-softmax or straight-through estimator.","section":"DogLayout for Discrete Data (Eq. 9)"},{"comment":"The numerical evidence does not consistently support the abstract's claim of 'surpassing' and 'competitive performance'. Against LayoutDM, DogLayout is worse on the majority of FID entries (e.g., C+S→P PubLayNet FID 9.47 vs 4.29; Completion PubLayNet FID 16.8 vs 8.47) and on all MaxIoU entries reported in Table 2. The overlap improvement of 16.43 to 9.59 is taken from a single row (C→S+P PubLayNet), while the unconditional overlap is worse (16.3 vs 13.43). No error bars, repeated runs, or significance tests are reported, so it is unclear whether the local improvements are stable. Please report variance across multiple seeds, give a balanced summary of wins and losses against the strongest baseline, or revise the claims to match the evidence.","section":"Quantitative Comparisons, Tables 1 and 2"},{"comment":"The ablation for discrete label generation reports only the discriminator's real/fake probability curves (Figure 5) and does not measure whether the generated labels are semantically correct. The FID and alignment metrics can be insensitive to label errors, and the completion task's poor FID on PubLayNet (16.8 vs. LayoutDM's 8.47) is consistent with the unvalidated continuous relaxation. A dedicated experiment that isolates label quality, such as label accuracy on a held-out set or a comparison of generated and real categorical distributions, is necessary to support the paper's first contribution.","section":"Ablation Study, Discrete Layout Generation"}],"minor_comments":[{"comment":"The sentence 'Our model outperforms all non-Diffusion models, with a few exceptions in tasks involving MaskGIT and BLT' is self-contradictory; if there are exceptions, the model does not outperform all non-diffusion models.","section":"Quantitative Comparisons"},{"comment":"Please state explicitly which timestep setting corresponds to which task in the 175x speed claim, and clarify whether LayoutDM and DogLayout were timed on identical hardware and with identical batching and implementation practices.","section":"Sampling Time Comparison, Table 3"},{"comment":"The user study report lacks the number of pairwise comparisons per task, the exact preference counts, and any significance test; please provide these details so the reader can assess the strength of the preference result.","section":"User Study, Figure 3"},{"comment":"The number of denoising timesteps T is selected using the same FID metric used in the main comparison; please state whether this introduces selection bias and report validation-set numbers separately from test-set numbers.","section":"Ablation Study, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is an application of denoising diffusion GANs to layout generation, and its novelty rests on the discrete-label relaxation. The speed contribution is credible, but the discrete-label mechanism and the aggregate quality claims need substantially stronger support before the paper can be recommended for acceptance. The code link is a positive signal, but I would want to see the requested label-validity experiments and repeated-run statistics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Speed is the real story here. DogLayout takes the Denoising Diffusion GAN recipe from DDGAN/SIDDM, applies it to layout generation with T=4-12 instead of 50, and reports a 175x sampling speedup (Table 3). That number is consistent with the method and easy to believe. The paper also deserves credit for an honest limitation section and for clearly attributing the core diffusion-GAN machinery to prior work.\n\nThe genuinely new piece is the discrete-label handling: corrupt a one-hot label prototype with Gaussian noise, denoise it as a continuous vector, and sample via softmax/argmax. It is a simple, differentiable way to let a GAN output categories, and it plausibly explains why the model can do unconditional and completion tasks that previous layout GANs couldn't. That is worth a careful look.\n\nThe soft spots are real but not fatal. The quality evidence is inconsistent: DogLayout is worse than LayoutDM on several FID scores and on unconditional overlap (16.3 vs 13.43), and the headline overlap improvement (16.43 to 9.59) comes from a single task. There are no error bars or significance tests. On the discrete side, the paper gives no derivation from discrete-state diffusion and no direct metric for label correctness; the stress-test concern that the continuous relaxation has no categorical semantics is a fair critique of the justification, though calling it 'arbitrary labels' is stronger than the evidence supports. The completion result on PubLayNet (FID 16.8 vs 8.47) is consistent with the discrete component being underpowered, so this is the part a referee should probe.\n\nThe ablation over T (Table 4) shows the model is sensitive to timestep choice, and T is tuned on the same FID metric used in the main comparison. That is standard model selection, not circular, but it would be nice to see the sweep on a separate validation set.\n\nThis paper is a solid, if not spectacular, contribution to layout generation. The speed win is reproducible in principle, the code is promised, and the discrete-label trick is interesting enough to merit referee time. I would send it out. My own verdict would be a conditional accept: the authors need to add label-accuracy numbers, error bars, and tone down the 'outperforms' language, but the core idea is sound enough to build on.","headline":"A genuinely fast diffusion-GAN for layout generation; the discrete-label trick is interesting but under-justified, and the quality claims are oversold.","tokens_in":11447,"tokens_out":2903,"would_cite":false,"duration_ms":30379,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-12T05:26:01.697976+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}