{"id":"01d3c705-4c6a-48d9-badb-7c7f7ee68528","arxiv_id":"2412.16859","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"An adversarially trained latent diffusion model with long encoder-decoder skip connections reports state-of-the-art mIoU of 74.4 and 67.2 on two unsupervised domain adaptation benchmarks.","lead":"ICCLD is a latent-diffusion segmentation model that adds skip connections between encoder and decoder plus adversarial training during denoising. On synthetic-to-real benchmarks it reports mIoU of 74.4 (GTA5 to Cityscapes) and 67.2 (Synthia to Cityscapes), slightly above prior state of the art.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The diffusion loss in Eq. 6 is degenerate: conditioning on the clean latent makes the target noise a deterministic function of the inputs, so the Step-2 ablation cannot be attributed to diffusion-based alignment as claimed.","rationale":"The paper's SOTA claim depends on the assertion that the second training phase, adversarial learning on the denoising UNet, pushes performance from 68.4 to 74.4 on GTA5→Cityscapes. For that assertion to hold, Eq. 6 must be a meaningful diffusion objective and Eq. 7 must implement a coherent adversarial alignment mechanism. Eq. 6 fails this test because conditioning on the same clean latent used to construct the noisy input makes the target noise an analytically recoverable function of the two network inputs; the loss can be minimized without learning any generative representation. This is a correctness risk, not a disagreement with consensus. The ablation in Table 1 therefore cannot be interpreted as evidence for diffusion-based domain alignment. The additional issues with Eq. 7 (no defined discriminator loss or alternating updates) and the undefined 'Mask Encoder c Conditioning' block in Fig. 1(b) further weaken the description, but the degeneracy of Eq. 6 is the most load-bearing single concern because it undermines the entire Step-2 mechanism. The paper does provide public datasets, standard baselines, and an ablation table, and I am not questioning the authors' intent; the concern is that the written method is internally inconsistent and the reported gains are not attributable to the claimed mechanism as specified. If the authors can supply code showing that the actual conditioning is not the clean latent and that the adversarial term is implemented as a proper minimax objective, the concern would be resolved. Without that, the existing REJECT verdict is appropriate, so I recommend no change.","tokens_in":14668,"tokens_out":10112,"duration_ms":91323,"concrete_test":"Implement or obtain from the authors the Step-2 training loop. For fixed z0*, verify analytically or numerically that ϵ_opt(x_t,z0*,t) = (x_t − √ᾱ_t z0*)/√(1−ᾱ_t) achieves zero loss in Eq. 6 for all t. Then run two ablations on GTA5→Cityscapes: (a) replace the z0* conditioning with a constant/null condition while keeping Eq. 7; (b) keep z0* conditioning but replace Eq. 7 by the KL-to-uniform term only. If (a) yields mIoU within ~1 point of 74.4, the diffusion loss is not the source of the gain; if (b) matches 74.4, the adversarial cross-entropy term is unnecessary. Also report the mean relative L2 error between the final denoised latent and z0* to check whether inference degenerates to an identity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central mechanism rests on Step 2: training the denoising UNet ϵθ with Eq. 6 plus Eq. 7, with the ablation (Table 1) attributing a +8.0/+6.5 mIoU gain to this phase. This attribution is not supported because the stated diffusion objective is degenerate. In Eq. 6, the noisy input is √ᾱ_t z* + √(1−ᾱ_t)ϵ and the conditioning variable z0* is the clean latent extracted from the same image. For any t>0, the target noise is then determined by the two inputs: ϵ = (x_t − √ᾱ_t z0*)/√(1−ᾱ_t). A network can drive L_ldm to zero by learning this inverse, without learning any generative or denoising behavior. At inference, iterated denoising with such a network returns z0*, so the diffusion path is close to identity and the decoder is essentially fed the encoder's own latent. The paper never rules out this shortcut. Eq. 7 is also not a valid adversarial objective: the cross-entropy term and the KL-to-uniform term pull fdis in opposite directions when optimized jointly, and no discriminator loss or alternating min-max update is defined. Finally, Fig. 1(b) shows an unexplained 'Mask Encoder c Conditioning' block absent from the text. Together these gaps make the Step-2 ablation and the SOTA claim unreproducible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes ICCLD, a latent diffusion model with long inter-coder skip connections and adversarial alignment of the denoising UNet, for unsupervised domain adaptation in semantic segmentation. Training is done in two steps: first a teacher-student segmentation adaptation on the encoder/decoder, then adversarial domain alignment of the denoising network. The paper reports state-of-the-art mIoU scores of 74.4 on GTA5-to-Cityscapes and 67.2 on Synthia-to-Cityscapes, and presents ablations attributing large gains to the second phase, especially when the inter-coder connection is used.","tokens_in":15050,"tokens_out":10701,"duration_ms":84613,"significance":"If the proposed mechanism were sound, the paper would address a relevant problem and the two-step teacher-student diffusion design would be an interesting direction. The authors are to be credited for tackling difficult benchmarks and for reporting class-wise results. However, the central diffusion objective in Eq. (6) is degenerate as written, the adversarial loss in Eq. (7) is not a well-defined adversarial game, and the ablation numbers in the text disagree with Table 1. These issues prevent the paper from substantiating its main claim, and the reported SOTA margins over HRDA are small enough that repeated-seed evaluation would be needed even if the formulation were corrected.","major_comments":[{"comment":"The diffusion loss in Eq. (6) is degenerate as written. The noisy input is constructed as sqrt(alpha_bar_t) z* + sqrt(1 - alpha_bar_t) epsilon, and the conditioning input is z*_0, which the text identifies as the clean latent extracted from the same source, mixed, or target image. If z* and z*_0 are the same latent, the target noise is a deterministic function of the two inputs, epsilon = (x_t - sqrt(alpha_bar_t) z*_0) / sqrt(1 - alpha_bar_t), so any network can drive L_ldm to zero by implementing this arithmetic; the reverse process then returns the conditioning latent, making the diffusion path an identity. If z* and z*_0 are intended to be different, the distinction is never defined. This collapse is not discussed and undermines the attribution of the Step-2 gains in Table 1 to diffusion-based alignment.","section":"Section 3.4, Eq. (6)"},{"comment":"Equation (7) is not a coherent adversarial objective. The first expectation E[o* log fdis(...)] rewards the discriminator for predicting the true domain, while the KL-to-uniform term rewards a uniform prediction; optimizing the sum drives the two terms in opposite directions. No discriminator classification loss or alternating min-max update is defined, so it is not clear how fdis is trained or why this procedure aligns source and target latents. The paper also does not state the relative weight of L_adv in the total loss, although Table 1 attributes all Step-2 gains to this term.","section":"Section 3.4, Eq. (7)"},{"comment":"The ablation numbers in the text do not match Table 1. Section 5.1 states that the model with the inter-coder connection achieves 69.3 mIoU on Synthia-to-Cityscapes, but Table 1 lists 67.2 for Step 1&2 with the connection and 69.3 for Step 1&2 without the connection on GTA5-to-Cityscapes. Section 5.2 reports improvements of 9.0 and 3.8 mIoU from the second phase without the inter-coder connection, whereas Table 1 implies 11.0 and 15.8 mIoU (58.3 to 69.3 and 42.7 to 58.5). These inconsistencies make the central ablation impossible to verify.","section":"Section 5.1 and Table 1"},{"comment":"The SOTA claim is not statistically supported. The reported margins over HRDA are +0.6 mIoU on GTA5-to-Cityscapes and +1.4 mIoU on Synthia-to-Cityscapes, yet the paper gives no number of independent runs, no standard deviations, and no significance test. In UDA semantic segmentation, margins of this size are often within run-to-run variability, so the claim that ICCLD outperforms existing methods is not established without repeated-seed evaluation.","section":"Section 6, Table 2"},{"comment":"The architecture description is incomplete. Figure 1(b) contains a block labelled 'Mask Encoder c Conditioning' that is never defined in the text, and Section 3.3 does not describe how the conditioning input z*_0 is injected into the denoising UNet (concatenation, cross-attention, or otherwise). Without this information the proposed mechanism cannot be reproduced or compared against alternative explanations of the reported gains.","section":"Figure 1(b) and Section 3.3"}],"minor_comments":[{"comment":"The method name is inconsistent: ICCLD in the title and abstract, 'Conditional and Inter-coder Connected Latent Diffusion (CICLD)' in Section 7, and 'Conditional and Long skip-connected Diffusion Model (CLDM)' in Figure 2.","section":"Title, Abstract, Section 7, Figure 2"},{"comment":"The phrase 'prepared target label ys' appears to be a typo for the source label; as written it is unclear which labels are used for the mixed image.","section":"Section 3.4, Step 1"},{"comment":"The sentence claiming that class-wise performance differences are 'generally small—often below 0.2' is contradicted by Table 2, where several differences exceed 1 mIoU (e.g., Traffic Sign and Terrain on GTA5-to-Cityscapes).","section":"Section 6, Table 2"},{"comment":"The notation in Eq. (6) is ambiguous: z* is used both as a generic latent variable and as a placeholder for zs, zs+t, zt, while z*_0 is defined only verbally; a precise index convention would help.","section":"Equation (6)"},{"comment":"Implementation details are incomplete: the paper does not specify the number of diffusion timesteps T used during training, the loss weight for L_adv, or the exact architecture of the discriminator output head, and no code is provided.","section":"Section 4, Implementation details"},{"comment":"The qualitative comparison refers to 'green-coloured dotted boxes', but the figure as printed does not clearly show these boxes; please adjust the figure or the caption.","section":"Figure 4"}],"recommendation":"reject","confidential_remarks":"I recommend rejection. The central technical contribution, namely the second-phase diffusion objective and adversarial alignment, is not well-defined as presented, and the ablation numbers are internally inconsistent. A resubmission would need to fix the conditioning scheme, specify a proper adversarial game, and provide repeated-seed evaluations with error bars."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper reports plausible-looking SOTA numbers on two standard UDA benchmarks, but the described method has a load-bearing flaw—the diffusion loss is degenerate—and the internal numbers do not add up. I would not send this to referees in its current form.\n\nWhat's new: applying latent diffusion to UDA semantic segmentation, with adversarial alignment on the denoising UNet, is a reasonable combination I have not seen exactly in this form. The inter-coder skip connection (long skip from encoder to decoder) is a sensible architectural tweak, and the related work coverage is adequate.\n\nThe problem is that the method as written cannot do what it claims. Equation 6 conditions the denoiser on the clean latent z0* that was also used to construct the noisy input. For any t>0, the target noise is a deterministic function of those two inputs; the network can drive the loss to zero by recovering the noise arithmetically. So the diffusion path is effectively identity, and the +6.0/+6.5 mIoU gain attributed to Step 2 in the ablation cannot be credited to diffusion-based alignment. The paper never rules out this shortcut. Equation 7 is also not a coherent adversarial objective: the cross-entropy term and the KL-to-uniform term pull in opposite directions, and no discriminator update or min-max procedure is defined.\n\nThere are smaller but telling problems. Figure 1(b) shows a 'Mask Encoder c Conditioning' block that is never explained in the text. The model name is inconsistent (ICCLD vs CLDM vs CICLD across abstract, figures, and conclusion). Section 5.2 claims improvements of 9.0 and 3.8, which match neither the with-intercoder gains (6.0, 6.5) nor the without-intercoder gains (11.0, 15.8) in Table 1. The reported margins over HRDA are +0.6 and +1.4 mIoU with no error bars, no seeds, and no code.\n\nWhat is salvageable: if the authors can fix the conditioning to be non-circular (e.g., condition on a different modality or a frozen feature), give a proper adversarial formulation, and correct the numbers, the idea might be worth revisiting. As is, the paper is not reproducible and the central claim is unsupported.\n\nRecommendation: desk reject, or at most return for major revision before any referee work. The topic is hot, but the current draft needs substantial reworking before it merits referee time.","headline":"Plausible SOTA numbers undermined by a circular diffusion loss and internal inconsistencies; not referee-ready.","tokens_in":15538,"tokens_out":6007,"would_cite":false,"duration_ms":46297,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ICCLD claims that a latent diffusion segmentation model with inter-coder skip connections and adversarial denoising achieves state-of-the-art unsupervised domain adaptation, with mIoU 74.4 on GTA5→Cityscapes and 67.2 on Synthia→Cityscapes.","keywords":["unsupervised domain adaptation","semantic segmentation","latent diffusion models","adversarial learning","inter-coder connections","denoising UNet","GTA5 to Cityscapes","Synthia to Cityscapes"],"falsifier":"Run the full GTA5→Cityscapes pipeline with the noise-prediction loss of Equation (6) inactivated while keeping Equation (7); if mIoU stays at 74.4, the diffusion objective is not load-bearing. Alternatively, on the validation set, compute the correlation between the sampled noise ε and the predicted noise εθ: near-zero correlation would show the denoiser is not actually solving the stated denoising task.","tokens_in":14492,"feed_emoji":"🚗","tokens_out":7894,"duration_ms":60186,"temperature":0.7,"pith_summary":"This paper proposes Inter-Coder Connected Latent Diffusion (ICCLD), a latent diffusion segmentation model that adds long skip connections between its encoder and decoder and trains its denoising network with an adversarial domain-alignment loss. The goal is to transfer semantic segmentation from labeled synthetic images to unlabeled real images, avoiding expensive manual pixel annotations. The paper reports state-of-the-art results on the two standard benchmarks: 74.4 mIoU for GTA5→Cityscapes and 67.2 mIoU for Synthia→Cityscapes, surpassing prior methods by 0.6 and 1.4 mIoU. A two-stage recipe is central: first adapt the encoder and decoder on segmentation with mixed source-target images, then align latent distributions by training the denoiser to both predict noise and confuse a domain discriminator. If correct, the method shows that diffusion dynamics themselves can serve as a domain-alignment mechanism rather than only as a generative prior.","feed_headline":"Adversarial latent diffusion tops UDA segmentation at 74.4 mIoU","feed_subtitle":"Inter-coder skip connections plus adversarial denoising set new records on GTA5 and Synthia transfers.","key_machinery":"The load-bearing mechanism is ICCLD, an LDM-style encoder-denoise-decoder stack. Its distinctive structural piece is the inter-coder connection: long skip connections that feed the encoder's multi-scale features directly into corresponding decoder blocks, preserving fine spatial detail that a latent bottleneck would lose. Its distinctive training piece is the second phase, where the denoising UNet ϵθ is conditioned on the clean latent z0 (Eq. 6) and optimized with a DDPM noise-prediction loss plus an adversarial KL-divergence loss (Eq. 7); the discriminator fdis decides whether a predicted noise came from the source, target, or mixed domain, and aligning those predicted-noise distributions is what the paper credits for closing the domain gap.","core_discovery":"The paper's central claim is that inter-coder connected latent diffusion plus adversarial alignment sets a new state of the art for synthetic-to-real semantic segmentation without target labels. Training is split into two phases: first, the encoder and decoder are domain-adapted on segmentation using ClassMix-generated mixed images and pseudo-labels; second, with the encoder frozen, the denoising UNet is trained to predict the diffusion noise and to fool a discriminator that classifies predicted noise as source, target, or mixed. Across the two benchmarks the method reaches 74.4 mIoU for GTA5→Cityscapes and 67.2 mIoU for Synthia→Cityscapes, outperforming the previously published best results from DAFormer and HRDA. The ablation data support both design choices: the inter-coder connection improves boundary-level accuracy, and the adversarial second phase reduces false positives and produces cleaner masks.","pith_inferences":["A direct test of the mechanism would be to train the second phase with Equation (6) removed and only the adversarial loss active; comparable mIoU would indicate the diffusion objective is not the source of the gains.","Because the reported margins over HRDA are 0.6 and 1.4 mIoU, a repeated-seed study with standard deviations would determine whether the method reliably surpasses the prior state of the art.","The three-domain discriminator over predicted noise is portable: the same adversarial alignment could be applied to any feature-space UDA pipeline without requiring a diffusion model around it.","If the conditioning shortcut is absent in this architecture, the recipe suggests broad use of latent diffusion in dense prediction; if the shortcut is present, other conditional diffusion segmentation models that condition on the same clean latent would need re-examination."],"forward_implications":["Adding the adversarial second phase moves Step-1 performance from 68.4 to 74.4 mIoU on GTA5→Cityscapes and from 60.7 to 67.2 mIoU on Synthia→Cityscapes when the inter-coder connection is present.","The inter-coder connection alone lifts Step-1 performance from 58.3 to 68.4 mIoU on GTA5→Cityscapes and from 42.7 to 60.7 mIoU on Synthia→Cityscapes.","A diffusion-based UDA model can outperform non-diffusion state-of-the-art methods on both standard benchmarks, setting reference scores of 74.4 and 67.2 mIoU.","Class-wise, the method wins 11 of 19 classes on GTA5→Cityscapes and 12 of 16 on Synthia→Cityscapes, with the largest improvements concentrated in boundary-heavy classes such as Wall, Fence, and Rider."],"supporting_citations":[{"why":"Supplies the DDPM forward and reverse processes and the simplified noise-prediction loss used in Equation (4).","marker":"[11]"},{"why":"Provides the latent diffusion model backbone with encoder, denoising UNet, and decoder that ICCLD modifies.","marker":"[26]"},{"why":"Supplies ClassMix, the source-target image and label mixing strategy used in the first adaptation phase.","marker":"[22]"},{"why":"Provides the teacher-student EMA training scheme and the DAFormer baseline that must be surpassed.","marker":"[15]"},{"why":"Provides HRDA, the strongest baseline, with 73.8 and 65.8 mIoU respectively, and the main comparison for the state-of-the-art claim.","marker":"[16]"},{"why":"Provides the evaluation protocol and the student-teacher framework referenced for the unsupervised domain adaptation setup.","marker":"[40]"},{"why":"Supplies the DDIM scheduler and the 50-step inference procedure used to produce final denoised latents.","marker":"[32]"},{"why":"Defines the Cityscapes dataset that is the target domain for both benchmark transfers.","marker":"[4]"},{"why":"Defines the GTA5 synthetic dataset used as the labeled source domain in the first benchmark.","marker":"[25]"},{"why":"Defines the Synthia synthetic dataset used as the labeled source domain in the second benchmark.","marker":"[28]"}],"fun_headline_variants":["ICCLD: adversarial latent diffusion sets new UDA segmentation record","Adversarial latent diffusion hits 74.4 mIoU on GTA5→Cityscapes","Latent diffusion with adversarial domain adaptation beats UDA SOTA","Inter-coder connected latent diffusion sets UDA segmentation SOTA","ICCLD adversarial latent diffusion improves synthetic-to-real segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach depends on the assumption that conditioning the denoising network on the same clean latent used to build its noisy input leaves a genuine denoising task rather than an identity shortcut, so the second-phase gains can be attributed to diffusion-based alignment.","fun_headline_variants_meta":{"raw":{"variants":["ICCLD: adversarial latent diffusion sets new UDA segmentation record","Adversarial latent diffusion hits 74.4 mIoU on GTA5→Cityscapes","Latent diffusion with adversarial domain adaptation beats UDA SOTA","Inter-coder connected latent diffusion sets UDA segmentation SOTA","ICCLD adversarial latent diffusion improves synthetic-to-real segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00067,"raw_usage":{"total_tokens":3048,"prompt_tokens":935,"completion_tokens":2113,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":2019}},"tokens_in":551,"tokens_out":2113,"duration_ms":13387,"temperature":1.0,"reasoning_tokens":2019,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:15:08.915681+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full GTA5→Cityscapes pipeline with the noise-prediction loss of Equation (6) inactivated while keeping Equation (7); if mIoU stays at 74.4, the diffusion objective is not load-bearing. Alternatively, on the validation set, compute the correlation between the sampled noise ε and the predicted noise εθ: near-zero correlation would show the denoiser is not actually solving the stated denoising task.","supporting_citations":[{"cited_title":"Denoising diffu- sion probabilistic models","cited_arxiv_id":null,"evidence_quote":"Supplies the DDPM forward and reverse processes and the simplified noise-prediction loss used in Equation (4)."},{"cited_title":"High-resolution image synthesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Provides the latent diffusion model backbone with encoder, denoising UNet, and decoder that ICCLD modifies."},{"cited_title":"Classmix: Segmentation-based data aug- mentation for semi-supervised learning","cited_arxiv_id":null,"evidence_quote":"Supplies ClassMix, the source-target image and label mixing strategy used in the first adaptation phase."},{"cited_title":"The cityscapes dataset for semantic urban scene understanding","cited_arxiv_id":null,"evidence_quote":"Defines the Cityscapes dataset that is the target domain for both benchmark transfers."},{"cited_title":"Playing for data: Ground truth from computer games","cited_arxiv_id":null,"evidence_quote":"Defines the GTA5 synthetic dataset used as the labeled source domain in the first benchmark."},{"cited_title":"The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes","cited_arxiv_id":null,"evidence_quote":"Defines the Synthia synthetic dataset used as the labeled source domain in the second benchmark."}],"review_version":1}