{"id":"4343c4f9-ba78-4277-bd9f-d1bb6076ccb6","arxiv_id":"2505.15822","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Mamba state-space model encoder, MambaStyle, inverts real images into StyleGAN's latent space with fewer parameters and faster inference than prior encoders while keeping competitive reconstruction and editing quality.","lead":"MambaStyle is a new encoder that converts real photos into StyleGAN's editing space using Mamba state-space modules, cutting parameters and inference time versus earlier GAN-inversion encoders. It reports competitive reconstruction and edit quality on faces and cars, suggesting faster real-time image editing is possible.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training only on StyleGAN-synthesized pairs without any real-image supervision leaves the real-image inversion/editing claim unvalidated.","rationale":"The reader's weakest assumption identifies exactly the load-bearing premise of the paper: the central claim concerns real-image inversion and editing, but all training data are generated on the StyleGAN2 manifold. This is the most serious gap because it directly threatens the external validity of the reported real-image metrics. The efficiency and architecture contributions are plausible and supported by Table 1, and the ablation study in Table 3 gives useful internal evidence. However, the absence of any real-image training, fine-tuning, or even a diagnostic comparison between synthetic and real test sets means the headline result may not transfer. The concern is not an internal inconsistency but a missing validation step. If the proposed test shows a large performance gap on synthetic test images, the paper's advertised balance claim would need to be substantially revised, but the conditional verdict remains appropriate until that experiment is run. I therefore recommend keeping the reader's verdict unchanged.","tokens_in":13281,"tokens_out":6623,"duration_ms":72706,"concrete_test":"Evaluate MambaStyle on a held-out set of 1,000 latent codes sampled from the same pretrained StyleGAN2 used to generate the training pairs (z ~ N(0, I), w = M(z), images X = G(w)), and compute the same inversion metrics (LPIPS, L2, MS-SSIM) and edit FIDs as in Table 1. If these synthetic-image metrics are substantially better than the reported real-image results, the encoder has not generalized across the synthetic-to-real domain gap, and the real-image numbers in Table 1 do not support the paper's claims.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is real-image inversion and editing, but every training pair is synthetic: Section 3.2 and Fig. 2 synthesize X = G(w) and X_e = G(w+d) from a random z, so the encoder and Fuser never see a real photograph. At inference (Section 4), the same model is applied to real CelebA-HQ and Stanford Cars images, and the reported LPIPS/L2/MS-SSIM/FID are taken as evidence of real-image performance. This is the load-bearing premise: the VSSM encoder must generalize from the StyleGAN2 manifold (FFHQ/LSUN Cars) to the real-image distribution, which has different backgrounds, sensor noise, and out-of-distribution attributes. The paper provides no analysis of this domain gap and no real-image training or fine-tuning. Because the Fuser injects F_k features into StyleGAN at layer k=9 (Eq. 5, Section 3.2), the encoder can achieve low L2/MS-SSIM by copying image content into features rather than learning a meaningful W+ code; the edit metrics (FID for Smile/Glasses/Old) are the only evidence that editing actually works on real images, and those are worse than SFE on two of three edits. Without a control experiment on synthetic test images or real-image supervision, the abstract's claim of 'superior inversion and editing results' on real images is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"MambaStyle proposes a single-stage, encoder-based StyleGAN inversion and editing framework that uses vision state-space model (VSSM) modules in both a multi-scale encoder and a novel 'Fuser' module. The encoder outputs a W+ latent code and feature maps; the Fuser combines the editing direction with the encoded features and injects them into StyleGAN2 at layer k=9. Training uses paired synthetic images generated from noise with a pretrained StyleGAN2 generator and editing directions, and the model is evaluated on real CelebA-HQ and Stanford Cars test images. The paper claims a superior balance of inversion quality, editing quality, and computational efficiency, reporting 104.56M total parameters, 157.82 GMACs, and 0.023 s inference time.","tokens_in":13521,"tokens_out":3502,"duration_ms":35198,"significance":"If the efficiency and real-image generalization claims hold, the paper provides a useful demonstration that state-space models can serve as efficient encoders for GAN inversion, potentially enabling more interactive applications. The architecture and the efficiency numbers are a genuine contribution. However, the central claim of 'superior inversion and editing results' is contradicted by the paper's own Table 1: SFE achieves better LPIPS, FID, and two of three editing FID scores, while MambaStyle's wins are concentrated in L2, MS-SSIM, and one edit. Moreover, the model is trained exclusively on synthetic StyleGAN images and then applied to real photos, leaving an unverified domain gap. The paper does not provide code or error bars, which limits reproducibility and statistical assessment, but the efficiency advantage is clearly quantified.","major_comments":[{"comment":"The claim that MambaStyle achieves 'superior inversion and editing results' is not supported by the paper's own numbers. In Table 1, SFE outperforms MambaStyle on LPIPS (0.019 vs. 0.025), FID (3.535 vs. 7.575), and on two of the three editing metrics (Smile(-): 24.388 vs. 27.149; Old(+): 41.677 vs. 42.819). MambaStyle wins on L2, MS-SSIM, and Glasses(+), but these are pixel-level or structural metrics that can be inflated by feature copying through the Fuser. The paper should reframe its contribution as efficiency at competitive quality, and temper the abstract and conclusion accordingly.","section":"Abstract; Section 4.2, Table 1"},{"comment":"The model is trained exclusively on synthetic pairs (X, Xe) generated from random noise via pretrained StyleGAN2 (Fig. 2, and the text 'generated from noise vectors z without directly using the original images'), yet all quantitative evaluations are performed on real CelebA-HQ and Stanford Cars images. The encoder never observes a real photograph during training, so the reported real-image inversion and editing metrics rest on an unverified assumption of generalization from the StyleGAN manifold to the real-image distribution. The authors need to provide a control experiment on held-out synthetic test images, add real-image fine-tuning, or analyze the distribution shift (for example, reporting FID between the generated training images and the real test set). Without this, the central real-image claims are not established.","section":"Section 3.2 and Section 4 (Dataset paragraph)"},{"comment":"No error bars, confidence intervals, or statistical significance tests are reported for any metric. This is particularly important because MambaStyle trails SFE on several metrics and only marginally leads on others (e.g., L2 0.001 vs. 0.002; Glasses(+) 72.518 vs. 73.098). The reader cannot determine whether the differences between methods are meaningful. The authors should report means and standard deviations over multiple runs, or at least bootstrap confidence intervals, for the main comparisons.","section":"Section 4.2, Tables 1 and 2"},{"comment":"The 'W/o Fuser' ablation removes the Fuser entirely, which causes a large performance drop; this is expected because the Fuser provides the feature injection at layer k=9. However, the paper does not isolate the effect of the Fuser conditional on the same latent code quality. Since the encoder and Fuser are trained jointly, the ablation conflates the Fuser's contribution with changes in the encoder's training dynamics. A more informative ablation would keep the encoder fixed and add/remove only the Fuser path during training, or evaluate the encoder-only output with a frozen Fuser. This is a secondary point, but it would strengthen the causal claim about the Fuser.","section":"Section 4.3, W/o Fuser ablation"}],"minor_comments":[{"comment":"The text says 'limiting its flexibility for editions' — 'editions' should be 'editing' or 'edits'.","section":"Section 3, paragraph after Eq. (5)"},{"comment":"The sentence 'our method effectively preserves the background of images both in inversion and after editing' is followed by a comparison to HyperStyle, but the grammar is incomplete: 'such as HyperStyle, which often fail' should be 'which often fails' or 'which often fail to preserve them'.","section":"Section 4.1, second paragraph"},{"comment":"The inference time measurements in Table 1 and Table 3 lack a stated hardware configuration. Please specify the GPU model and any warm-up or batch-size details so that times are comparable across the paper and against the literature.","section":"Section 4, Metrics paragraph"},{"comment":"The values of the weighting coefficients λ1...λ5 are deferred to the supplementary material, but no supplementary file is present in the submission. Please include these values, or make the supplementary material available, since the reported trade-offs depend on them.","section":"Section 3.3, Eq. (8)"},{"comment":"The caption is extremely long and essentially duplicates the method description in Section 3. Please shorten it to a summary and refer the reader to Section 3 for details.","section":"Figure 2 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper has a credible efficiency story and a well-executed architecture comparison, but the title, abstract, and conclusion overclaim quality superiority relative to Table 1. The synthetic-only training and lack of real-image supervision is a deeper issue that, in my view, warrants a major revision rather than outright rejection. The authors could make this a solid contribution by repositioning it as an efficient encoder that is competitive with SFE on quality, adding a synthetic-vs-real generalization analysis, and reporting uncertainty estimates."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: MambaStyle is a real engineering contribution. The idea of using vision state-space models in a GAN inversion encoder is new, at least in this subfield, and their Fuser module—conditioning reconstruction features on an edit direction—is a sensible design. The efficiency numbers hold up: 104.56M total parameters versus SFE's 429M, 157.82 GMACs versus 397.27, and 0.023s inference versus 0.070s. That is the strongest part of the paper and it alone makes the work worth a look.\n\nWhere the paper hurts itself is in the abstract and conclusion, which claim 'superior inversion and editing results.' Table 1 says otherwise. SFE beats them on LPIPS (0.019 vs 0.025), FID (3.535 vs 7.575), and two of the three edit FIDs (Smile 24.388 vs 27.149; Old 41.677 vs 42.819). MambaStyle wins on L2, MS-SSIM, and Glasses, and the efficiency metrics. That is a 'competitive with efficiency advantage' story, not a 'superior' story. The writing should be fixed.\n\nThe train-on-synthetic/evaluate-on-real concern raised in the stress-test note is real but not novel to this paper. pSp, e4e, HFGI, and SFE all train on StyleGAN-generated pairs and then run on real photos. It is a field-wide convention. MambaStyle should at least mention the domain gap and note that no real-image supervision is used, but I would not call it a fatal flaw. The more serious weakness is that no error bars or code are provided, and the car-domain result is one table with no editing metrics. Given that the efficiency claim is the main contribution, releasing code and checkpoints matters.\n\nThe ablation study is mostly sensible. Replacing VSSM with ViT at similar parameter count gives close performance (ViT: LPIPS 0.037 vs Final 0.025; FID 11.724 vs 7.575) with more GMACs, which supports the efficiency claim. The Fuser and Le ablations show clear drops when removed.\n\nVerdict: this is a paper worth taking seriously. An editor should send it to review, not desk-reject it. The reviewer's job will be to push for honest framing, confidence intervals, and code. I'd bring it to a reading group focused on efficient generative models.","headline":"Efficient VSSM-based GAN inversion encoder with real speed and parameter gains; quality claims are overstated relative to its own table.","tokens_in":14087,"tokens_out":2541,"would_cite":false,"duration_ms":24848,"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":"MambaStyle uses vision state-space models to invert real images into StyleGAN's latent space in 0.023 seconds with a 104.56M-parameter encoder, matching the inversion quality of much larger encoders while enabling fast attribute edits.","keywords":["GAN inversion","StyleGAN2","state-space models","vision state-space models","image editing","encoder-based inversion","Mamba","real-time image editing"],"falsifier":"Retrain the same MambaStyle architecture on synthetic pairs drawn from a different base distribution or with an added distribution shift such as severe compression or out-of-domain faces, then evaluate inversion on the standard real-image test sets; if LPIPS and FID on real photos degrade substantially relative to the reported 7.575 FID on CelebA-HQ, the synthetic-only training assumption is falsified.","tokens_in":13053,"feed_emoji":"🖼️","tokens_out":8776,"duration_ms":82334,"temperature":0.7,"pith_summary":"Real-image editing through StyleGAN has been held back by a three-way trade-off: optimization-based inversion is accurate but slow, encoder-based inversion is fast but loses detail, and the most detailed encoders are heavy. This paper claims a single-stage encoder built on vision state-space models (VSSMs) breaks that trade-off, reconstructing and editing images while using far fewer parameters and less compute than the top-quality prior encoders. It reports inversion with L2=0.001 and MS-SSIM=0.986 on CelebA-HQ, second-best perceptual scores after a much larger method, and the fastest inference (0.023 s per image) among compared encoders. The same network both inverts and edits, conditioned on an edit direction d, and the whole design is trained on paired images synthesized from pretrained StyleGAN2 noise samples rather than on real photographs.","feed_headline":"StyleGAN editing runs in 0.023 seconds per image","feed_subtitle":"A lightweight state-space encoder matches big-encoder quality at interactive speed.","key_machinery":"The load-bearing object is the Vision State-Space Module (VSSM), built on the 2D selective scan (SS2D) operation: it expands an image patch into four directional sequences, processes each with the input-dependent S6 state-space update from Mamba, and recombines the results, giving linear complexity in sequence length. Three such modules refine the multi-scale features of the encoder, and the Fuser combines those features with the editing direction d using 1D SSM, 2D convolutions, and a VSSM module before passing through modulated convolution blocks to produce the fused feature map that the pretrained StyleGAN2 generator consumes. These two modules together convert the problem of editing a real image into a learned, single-pass mapping from (image, edit direction) to (latent code, feature map).","core_discovery":"The paper's central claim is that an encoder built from vision state-space models can map a real image into both a StyleGAN2 style code in W+ space and a spatial feature map at layer k=9, so the pretrained generator reconstructs the image or produces an edited version from the same representation. When the editing direction d is zero the pipeline inverts; when d is nonzero it generates an edited image. The contribution is the architecture: a multi-scale IResNet-based encoder whose hierarchical features are refined by VSSM blocks, plus a Fuser module that injects the edit direction into the feature map via 1D SSM, convolutional, and modulated-convolution layers. With 104.56M total parameters and 157.82 GMACs, the paper reports the best L2 and MS-SSIM inversion results on CelebA-HQ among the compared encoders, second-best LPIPS and FID behind a much heavier method, and faster per-image inference than every compared baseline, with a new state of the art on the car domain.","pith_inferences":["The authors do not test temporal consistency; if inversion at 0.023 s is applied to video frames, edited attributes could flicker, so measuring per-frame drift would tell whether real-time video editing is actually achievable.","Because training uses only StyleGAN2-generated pairs, the encoder's generalization is bounded by the generator's support; replacing the synthetic source with a different generator or mixing in real images is a direct test of whether the single-stage scheme transfers.","The Fuser injects the edit direction directly into spatial feature maps, so the same architecture could plausibly take text or reference-image embeddings as the conditioning input, turning the framework into a general conditioning mechanism rather than a direction-only editor."],"forward_implications":["Interactive image editing becomes practical: the full inversion-and-edit pipeline runs in about 0.023 s per image, faster than every compared encoder and orders of magnitude faster than optimization-based PTI (124 s).","The same trained model handles both tasks: setting the edit direction to zero reconstructs the image, and setting it to a StyleCLIP, GANSpace, or InterfaceGAN direction edits the attribute, with an explicit loss term enforcing that the latent-code change equals the requested direction.","High inversion fidelity no longer requires a heavy encoder: with 104.56M parameters and 157.82 GMACs, MambaStyle matches or beats the inversion quality of encoders two to four times larger, so the accuracy gains of feature-space inversion can be deployed in lower-budget settings.","The single-stage training recipe uses only synthetic pairs generated from pretrained StyleGAN2, avoiding the multi-phase training and pretrained-encoder initialization used by predecessors such as SFE and e4e."],"supporting_citations":[{"why":"Supplies the S6 selective state-space update that is the core computation of the VSSM modules.","marker":"[13]"},{"why":"Provides the 2D selective scan (SS2D) operation used to adapt S6 to image data.","marker":"[26]"},{"why":"Defines the pretrained StyleGAN2 generator whose W+ and feature spaces are being inverted and edited.","marker":"[21]"},{"why":"The strongest prior encoder for detail-rich inversion; its quality and parameter count are the main comparison.","marker":"[7]"},{"why":"Established the W+ encoder paradigm and the distortion-editability trade-off that MambaStyle targets.","marker":"[35]"},{"why":"A high-inversion-quality residual encoder used as a baseline for quality and cost comparisons.","marker":"[29]"},{"why":"Supplies the InterfaceGAN semantic directions used to evaluate editing quality.","marker":"[33]"},{"why":"Represents the widely used pSp encoder baseline for inversion and editing comparison.","marker":"[31]"}],"fun_headline_variants":["State-space model makes StyleGAN inversion 0.023s, lighter","MambaStyle: match big-encoder quality at 10x speed","Real-time StyleGAN editing: state-space encoder cuts cost","0.023s per image: VSSM encoder matches heavy GAN inversion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that paired training images generated entirely from random noise with a pretrained StyleGAN2 are a sufficient stand-in for real photographs, so an encoder that never sees real images during training will generalize to real test photos at inference.","fun_headline_variants_meta":{"raw":{"variants":["State-space model makes StyleGAN inversion 0.023s, lighter","MambaStyle: match big-encoder quality at 10x speed","Real-time StyleGAN editing: state-space encoder cuts cost","0.023s per image: VSSM encoder matches heavy GAN inversion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000338,"raw_usage":{"total_tokens":1840,"prompt_tokens":891,"completion_tokens":949,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":871}},"tokens_in":507,"tokens_out":949,"duration_ms":8572,"temperature":1.0,"reasoning_tokens":871,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:41:53.455282+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same MambaStyle architecture on synthetic pairs drawn from a different base distribution or with an added distribution shift such as severe compression or out-of-domain faces, then evaluate inversion on the standard real-image test sets; if LPIPS and FID on real photos degrade substantially relative to the reported 7.575 FID on CelebA-HQ, the synthetic-only training assumption is falsified.","supporting_citations":[{"cited_title":"Vmamba: Visual state space model","cited_arxiv_id":null,"evidence_quote":"Provides the 2D selective scan (SS2D) operation used to adapt S6 to image data."},{"cited_title":"Analyzing and improv- ing the image quality of stylegan","cited_arxiv_id":null,"evidence_quote":"Defines the pretrained StyleGAN2 generator whose W+ and feature spaces are being inverted and edited."},{"cited_title":"The devil is in the details: Stylefeatureeditor for detail-rich stylegan inversion and high quality image editing","cited_arxiv_id":null,"evidence_quote":"The strongest prior encoder for detail-rich inversion; its quality and parameter count are the main comparison."},{"cited_title":"Designing an encoder for stylegan image manipulation","cited_arxiv_id":null,"evidence_quote":"Established the W+ encoder paradigm and the distortion-editability trade-off that MambaStyle targets."},{"cited_title":"Styleres: Transforming the residuals for real image editing with stylegan","cited_arxiv_id":null,"evidence_quote":"A high-inversion-quality residual encoder used as a baseline for quality and cost comparisons."},{"cited_title":"In- terpreting the latent space of gans for semantic face editing","cited_arxiv_id":null,"evidence_quote":"Supplies the InterfaceGAN semantic directions used to evaluate editing quality."},{"cited_title":"Encoding in style: a stylegan encoder for image-to-image translation","cited_arxiv_id":null,"evidence_quote":"Represents the widely used pSp encoder baseline for inversion and editing comparison."}],"review_version":1}