{"id":"034a2af0-f2e6-4e3c-975b-4ef6ebb2403e","arxiv_id":"2504.14509","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"DreamID trains a one-step diffusion face swapper with triplet ID groups (source, pseudo target, real target) to jointly optimize identity transfer and attribute preservation.","lead":"DreamID is a face-swapping model that trains a fast one-step diffusion network using triplet ID groups: a source image, a proxy-generated pseudo target, and a real image that serves as explicit ground truth. The authors report improved identity similarity, attribute preservation, and a 0.6 second inference time at 512x512.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pseudo-target proxy is load-bearing: if FaceDancer's B~ fails to preserve A2's pose/expression/lighting, Lrec regresses toward A2 while conditioning on mismatched B~, and Table 4 shows proxy choice shifts Pose 2.90 to 2.20.","rationale":"Reader and I converge on the same weakest assumption. It is load-bearing because the Triplet ID Group construction replaces the unobserved true swapped image with A2 only under the assumption that B~ exactly captures A2's attributes; otherwise the supervised pair (conditioning on B~, target A2) is internally inconsistent. This is not a cosmetic issue: Table 4 provides a direct empirical lever, changing only the proxy changes Pose by 0.70 and Expression by 0.186. The concrete test isolates the mechanism by filtering triplets based on measured proxy error; if attribute metrics improve, the proxy assumption is the bottleneck, and if they do not, the diffusion model's conditioning is robust to proxy noise. I do not change the reader's CONDITIONAL verdict: the concern is real but addressable and does not by itself invalidate the method; it sharpens the condition under which the central claim holds. The 1-step/4-step reporting issue is also valid but secondary: the supplementary provides the 1-step numbers, so the claim is not unsupported, only awkwardly presented. The critique is on the argument, not the authors, and the proposed check would settle whether the concern actually lands.","tokens_in":12845,"tokens_out":5395,"duration_ms":45557,"concrete_test":"On a held-out subset of the training triplets, run FaceDancer to produce B~ and compute per-sample pose/expression L2 distances (using the same HopeNet and Deep3D metrics as Table 1) between B~ and A2. Retrain DreamID with identical hyperparameters on only the 50% of triplets with the lowest proxy error, then compare Table 1 metrics to the full-data model. If pose, expression, or FID improve significantly, proxy misalignment is an active bias in Lrec; if metrics are statistically unchanged, the framework is robust to proxy error and the assumption is non-load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 constructs the Triplet (A1, B~, A2) with A2 claimed to share attributes with B~ and identity with A1, so Lrec = ||A2 - A_hat||^2_2 in Eq. 4 is supposed to be explicit supervision while the model conditions on B~ through the target latent and landmarks. This pairing is valid only if the GAN proxy (FaceDancer) preserves A2's non-identity attributes exactly in B~. FaceDancer is not a perfect attribute-preserving map: its own metrics in Table 1 are Pose 2.32 and Expression 0.854, i.e., it already has attribute error. Any attribute shift in B~ makes the conditioning input inconsistent with the regression target; gradient descent on Lrec then teaches a compromise that matches neither B~ nor A2, biasing attribute preservation. The paper's Table 4 shows this is not hypothetical: swapping FaceDancer for Inswapper as the proxy worsens Pose from 2.20 to 2.90 and Expression from 0.789 to 0.975, while FID degrades from 4.69 to 5.89. The authors' own interpretation is that the proxy's attribute retention controls the final model's attribute retention. This undercuts the claim in Section 4.1 that using the real image A2 ensures 'the upper bound of the supervisory signal is very high': the effective upper bound is set by the pseudo target's attribute fidelity. The observed improvement over FaceDancer itself (2.20 < 2.32) is suggestive but is not controlled for proxy per-sample error or metric variance, so it does not by itself establish that the method is insensitive to this assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DreamID, a diffusion-based face swapping method. The key idea is a Triplet ID Group learning scheme: from two images of the same identity (A1, A2) and a different identity B, a GAN proxy (FaceDancer) generates a pseudo target B~ by transplanting A2's identity onto B. The triplet (A1, B~, A2) provides a real image A2 that shares identity with A1 and attributes with B~, enabling explicit pixel-level supervision through an ID loss, a reconstruction loss, and a diffusion loss. The method uses SD Turbo to enable one-step inference and proposes a three-part architecture (SwapNet, FaceNet, and ID Adapter). Experiments on FFHQ report state-of-the-art identity similarity, pose/expression preservation, and FID, with inference time of 0.6 seconds at 512×512. The supplementary includes a 4-step variant and reports its metrics separately.","tokens_in":13171,"tokens_out":5430,"duration_ms":46188,"significance":"The explicit-supervision idea is a genuine and potentially useful contribution: constructing paired training data for face swapping through a proxy model and a real target image is a practical workaround to the lack of ground-truth swapped images. The reported quantitative results are strong, the architecture components are reasonable, and the proxy-model ablation is informative. The fast single-step inference is practically valuable. However, the manuscript in its current form contains a major reporting inconsistency between the headline numbers and the speed claim, and a training detail (fixed t=999 diffusion loss) that is not adequately justified. These issues are fixable, but they currently prevent the central claim from being assessed fairly.","major_comments":[{"comment":"The main quantitative results, shown in the table on p. 6 (called Table 2 in the text but numbered Table 1 in the PDF), are obtained from the 4-step model, while the 0.6 s speed claim in the Abstract and §5.2 corresponds to the 1-step model. Supplementary Table 5 shows that the 1-step model has FID 5.08, Pose 2.31, Expression 0.790, which are worse than the headline FID 4.69, Pose 2.20, Expression 0.789. The manuscript should disclose this discrepancy in the main text and clearly distinguish which model produced each reported number; otherwise the statement that DreamID outperforms previous methods in just 0.6 seconds is misleading.","section":"§5.2 and Supplementary Table 5"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any metric. The differences that support the central claims, such as ID similarity 0.71 vs. CSCS 0.68 or Pose 2.20 vs. FaceDancer 2.32, may be within run-to-run or bootstrap variance on the 1,000-image test set. The authors should report standard deviations or confidence intervals, ideally over multiple training runs or at least via bootstrap resampling of the evaluation set, before claiming consistent superiority.","section":"§5.1, Table 1"},{"comment":"The diffusion loss is defined with a fixed timestep t = 999, with the justification 'we use one step property of SD Turbo for diffusion loss calculating'. This is not self-evident: standard diffusion training samples t uniformly from the noise schedule, and a fixed maximum-noise timestep would, in a standard diffusion model, only train the model to denoise from pure noise. Since SD Turbo is a distilled adversarial model, the authors need to explain why fixing t = 999 is appropriate and how this loss interacts with the SD Turbo objective. As written, this detail is load-bearing for the training procedure and must be justified.","section":"§4.1, Eq. (3)"},{"comment":"The claim that using the real image A2 as the learning target 'ensures that the upper bound of the supervisory signal is very high' is overstated. Because the pseudo target B~ is generated by a GAN proxy, any attribute error in B~ makes the conditioning input inconsistent with the regression target A2; the effective upper bound is therefore limited by the proxy's attribute fidelity. The authors' own Table 4 demonstrates this: replacing FaceDancer with Inswapper as the proxy degrades Pose from 2.20 to 2.90 and Expression from 0.789 to 0.975. The observed improvement over FaceDancer itself (2.20 < 2.32) is suggestive but not controlled for metric variance. The claim should be tempered and the proxy-dependence explicitly acknowledged as a limitation of the framework.","section":"§4.1, Triplet ID Group Construction"}],"minor_comments":[{"comment":"The text refers to the quantitative comparison as 'Table 2', but the PDF numbering labels it Table 1; the inference-speed table becomes Table 2, and the following tables are misaligned. Renumber the tables consistently.","section":"Section 5.2"},{"comment":"The inference-time measurement is reported as '0.6s' without details on the hardware configuration (e.g., batch size, whether VAE encoding/decoding is included, or whether the 1-step model is used). Specify the measurement protocol so the speed claim is reproducible.","section":"Section 5.2, Inference Speed"},{"comment":"The forward diffusion formula is garbled: 'p(1−αtϵ' should be '√(1−αt) ϵ'. The paper should be proofread for such typographical errors.","section":"Section 3, Eq. (1)"},{"comment":"The notation 'E(ID ˜B)' is ambiguous because 'ID B~' is used to denote both an image and an identity. Use a consistent notation, e.g., write the image as I_{B~} and the identity as IDB.","section":"Section 4.2, SwapNet description"},{"comment":"Several references are URLs or incomplete (e.g., [12] and [13] are GitHub links, [16] is a software citation without a venue). Please format them according to the journal's reference style.","section":"References"},{"comment":"The caption says 'replacing the source face (top-left) with the target face (bottom-left)', which is confusing because in face swapping the source provides identity and the target provides attributes. Clarify which image is the source and which is the target.","section":"Figure 1 caption"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the discrepancy between the 4-step and 1-step results: the main table reports the 4-step numbers while the abstract and conclusion claim the 0.6 s speed of the 1-step model. This looks like a versioning error in the manuscript preparation, but the authors must resolve it before acceptance. The fixed t=999 diffusion loss is technically questionable and should be explained or removed. The proxy-model sensitivity is a genuine limitation that should be discussed more carefully, but it is not a fatal flaw."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe real news here is the training data construction. DreamID takes two same-identity images (A1, A2), runs a GAN proxy (FaceDancer) to swap a different identity B onto A2's attributes, producing a pseudo target B~, and then trains a diffusion model to map (A1, B~) to the real image A2. That turns face swapping into a supervised reconstruction problem with a real pixel target, which is a genuine departure from the implicit-loss diffusion face swappers. It's a clean trick, and the results suggest it works: ID similarity 0.71 vs. 0.68 for the next best, pose/expression better than FaceDancer itself.\n\nWhat's good: the architecture (SwapNet, FaceNet, ID Adapter) is standard AnimateAnyone/IP-Adapter fare, but the paper doesn't oversell it. The ablations are informative, especially the proxy-model comparison: swapping FaceDancer for Inswapper degrades pose from 2.20 to 2.90, which is honest evidence that the proxy's attribute fidelity is load-bearing. The writing has typos but the logic is clear.\n\nThe main problem is a reporting inconsistency. Table 1's headline numbers are for the 4-step model (FID 4.69, pose 2.20, 1.3s), while the abstract and intro claim 0.6s, which corresponds to the 1-step model. The supplement shows the 1-step model is slightly worse: FID 5.08, pose 2.31, expression 0.790. That still looks competitive, but the paper shouldn't pair the 4-step metrics with the 1-step speed. The authors even explain they chose the 1-step model for speed, but then Table 1 doesn't say which one it is. That needs fixing.\n\nThe stress-test concern about proxy bias is real but not fatal. If the pseudo target B~ misaligns with A2 in pose/lighting/expression, the reconstruction loss pushes the model to output something that matches A2 while conditioning on a mismatched B~. The paper's own ablation shows proxy choice matters, so the claim that using A2 gives 'a very high upper bound' is overstated. It would be better to acknowledge that the effective supervision quality is capped by the proxy. That's a limitation to discuss, not a reason to reject.\n\nAlso missing: error bars, code release, and any justification for fixing t=999 in the diffusion loss. For a method that's meant to be practical, these matter.\n\nOverall, this is a solid contribution with a novel supervision idea. It deserves a serious referee. I'd recommend conditional acceptance pending clarification of the step-count reporting, error bars, and a more careful statement about proxy dependence. For a reading group, it's a good example of how to construct pseudo ground truth.","headline":"A genuinely new supervision trick for diffusion face swapping, but the paper mixes 1-step and 4-step numbers and leans on a proxy whose attribute errors propagate.","tokens_in":13782,"tokens_out":3549,"would_cite":true,"duration_ms":28548,"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":"The paper claims that a Triplet ID Group—two photos of one identity plus a GAN-proxy pseudo target—gives explicit pixel-level supervision, enabling a one-step diffusion model to swap faces at 512×512 in 0.6 seconds with top identity…","keywords":["face swapping","diffusion models","explicit supervision","identity preservation","attribute preservation","one-step inference","triplet learning"],"falsifier":"Train the same pipeline with a proxy model that deliberately distorts the target's lighting or pose during the swap; if attribute preservation on the test set stays flat or the generated images reproduce the distorted lighting, the reconstruction loss is anchored to the pseudo target rather than to the real image A2, and the claimed supervisory ceiling breaks.","tokens_in":12593,"feed_emoji":"🎭","tokens_out":9110,"duration_ms":76682,"temperature":0.7,"pith_summary":"Face swapping has lacked real ground-truth pairs: for a given source and target, no \"true\" swapped image exists, so prior diffusion methods train on implicit identity and reconstruction losses. This paper claims that explicit supervision is possible by building a Triplet ID Group: two photos of the same person (A1, A2) and a different identity B, where a GAN proxy swaps A2's face onto B to make a pseudo target. Because A2 is a real image sharing identity with A1 and all attributes with the pseudo target, A2 becomes the exact ground-truth output for swapping A1 into the pseudo target. To make pixel-level losses affordable, the denoiser is initialized from an accelerated one-step diffusion model, so the identity loss, reconstruction loss, and diffusion loss are computed on a single denoising step. The paper reports that the resulting model beats previous methods on identity similarity, pose and expression preservation, and fidelity, and runs at 512×512 in 0.6 seconds.","feed_headline":"Triplet trick gives face swapping real ground truth in 0.6 seconds","feed_subtitle":"A pseudo target links source identity to target attributes, turning the task into supervised image generation.","key_machinery":"The load-bearing object is the Triplet ID Group $(A_1, \\tilde{B}, A_2)$, where $\\tilde{B}$ is made by a GAN proxy from A2 and B; A2 gives exact supervision because it shares identity with the source and all non-identity attributes with the pseudo target. The second piece is one-step accelerated diffusion, a distilled latent diffusion model that denoises in one iteration, making the reconstruction and identity losses tractable and the inference fast. The third is the architecture: SwapNet, the base U-Net initialized from the accelerated model; FaceNet, a U-Net encoder that injects pixel-level identity features into self-attention; and an ID Adapter that injects semantic-level identity embeddings into cross-attention. Pose and expression conditioning enters through landmarks from a 3D face reconstruction model.","core_discovery":"The central claim is that the missing ground truth in face swapping can be manufactured. Given two images A1 and A2 of the same identity and one image B of a different identity, a GAN proxy model swaps A2's identity onto B, producing pseudo target $\\tilde{B}$; then $(A_1, \\tilde{B}, A_2)$ forms a triplet in which A2 is the exact target for swapping A1 into $\\tilde{B}$. The pseudo target is deliberately not the learning target—the real image is—so the supervisory signal is not limited by the proxy's output quality. Because the base model is an accelerated one-step diffusion model, the training can apply an L2 reconstruction loss between the generated image and A2, a cosine identity loss against A1, and the standard diffusion noise-prediction loss, all in a single step. The paper's experiments claim this yields top quantitative results, including an identity similarity of 0.71, an FID of 4.69, and improved pose and expression preservation, and that attribute-specific control follows by editing the triplet data, for example removing glasses from the pseudo target or altering its face shape.","pith_inferences":["The same triplet construction should transfer to other image-to-image tasks without ground truth, such as relighting or expression transfer: generate a pseudo target with a proxy, then use a real image sharing the target's attributes as supervision.","Because the real target A2 anchors the reconstruction, the method's ceiling is bounded by the quality of A2 itself; corrupting or lowering the resolution of A2 should degrade fidelity even if the proxy is perfect, which is a testable prediction.","A natural extension is to bootstrap: use DreamID's own outputs, rather than a GAN proxy, to build pseudo targets for the next training round; if the claim about a high supervisory upper bound is right, this should improve attribute preservation further."],"forward_implications":["Face swapping becomes a supervised image-to-image task rather than an implicitly supervised generation problem, so identity and attribute objectives no longer need to be balanced through proxy losses alone.","Diffusion-based face swapping can run at interactive speed: a single 512×512 swap in 0.6 seconds, compared with several seconds for prior diffusion methods.","Attribute control reduces to editing the triplet data: removing glasses from the pseudo target teaches the model to keep glasses from the source, and reshaping the pseudo target teaches face-shape transfer.","The trained model can surpass the proxy's attribute-preservation quality, because supervision comes from the real A2 rather than from the proxy's output.","The explicit-supervision framework extends beyond real-photo swapping, since the paper shows plausible results on stylized targets such as sketches, oil paintings, and cartoons."],"supporting_citations":[{"why":"Supplies the distilled one-step latent diffusion backbone that makes pixel-space losses tractable and enables 0.6-second inference.","marker":"[30]"},{"why":"The GAN proxy used to build pseudo targets; its attribute-preservation quality sets the ceiling for the reconstruction supervision.","marker":"[28]"},{"why":"Provides the frozen face-recognition encoder used to compute the cosine identity loss between the generated image and the source.","marker":"[12]"},{"why":"Defines the latent diffusion objective adopted as the diffusion loss in the total training loss.","marker":"[27]"},{"why":"Provides the 3D face landmarks and pose/expression coefficients used to condition the base U-Net.","marker":"[33]"},{"why":"Describes the adapter mechanism for injecting extra key-value identity embeddings, used by the ID Adapter module.","marker":"[36]"}],"fun_headline_variants":["Pseudo targets give face swapping explicit supervision in 0.6s","Triplet ID groups turn face swapping into supervised learning","One-step diffusion model swaps faces in 0.6 seconds with high fidelity","Pseudo-target triplets give face swapping ground truth","0.6-second face swap via one-step diffusion with triplet supervision"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the proxy model that builds the pseudo target faithfully preserving the target image's pose, lighting, expression, background, and makeup while only changing identity.","fun_headline_variants_meta":{"raw":{"variants":["Pseudo targets give face swapping explicit supervision in 0.6s","Triplet ID groups turn face swapping into supervised learning","One-step diffusion model swaps faces in 0.6 seconds with high fidelity","Pseudo-target triplets give face swapping ground truth","0.6-second face swap via one-step diffusion with triplet supervision"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000553,"raw_usage":{"total_tokens":2687,"prompt_tokens":1047,"completion_tokens":1640,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":1552}},"tokens_in":663,"tokens_out":1640,"duration_ms":11537,"temperature":1.0,"reasoning_tokens":1552,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:46:49.189805+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same pipeline with a proxy model that deliberately distorts the target's lighting or pose during the swap; if attribute preservation on the test set stays flat or the generated images reproduce the distorted lighting, the reconstruction loss is anchored to the pseudo target rather than to the real image A2, and the claimed supervisory ceiling breaks.","supporting_citations":[{"cited_title":"Facedancer: Pose-and occlusion-aware high fidelity face swapping","cited_arxiv_id":null,"evidence_quote":"The GAN proxy used to build pseudo targets; its attribute-preservation quality sets the ceiling for the reconstruction supervision."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the frozen face-recognition encoder used to compute the cosine identity loss between the generated image and the source."},{"cited_title":"High-resolution image synthesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Defines the latent diffusion objective adopted as the diffusion loss in the total training loss."},{"cited_title":"3d face reconstruction with the geometric guidance of facial part segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the 3D face landmarks and pose/expression coefficients used to condition the base U-Net."}],"review_version":1}