{"id":"4975cda9-4a6e-43ff-8e1a-9d72af7cd11e","arxiv_id":"2509.11213","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Beyond Sliders augments Concept Sliders with perceptual, adversarial, and an undefined triplet loss, claiming better real-world edits, but the evidence is weak and the derivation is not valid.","lead":"Beyond Sliders combines LoRA adapters, a GAN discriminator, and perceptual loss to extend Concept Sliders to real-world photo editing. The paper claims better realism and consistency than Concept Sliders, but provides no code, no training details, and an invalid mathematical derivation.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central 'conceptual triplet loss' is never defined; Eq. (8) and the training update in Eq. (11) are therefore underspecified, so the claimed improvement over Concept Sliders cannot be reproduced.","rationale":"The strongest claim is that Beyond Sliders produces 'a marked enhancement in image quality and realism' over Concept Sliders. To evaluate that claim, the method must be specified, and the text fails at the mathematical core. The reader's weakest assumption—that the perceptual loss (Eq. 7) pairs each generated image with an arbitrary, unaligned LFW face—is a serious design flaw. But even if that pairing were resolved, the method still would not be implementable because L_triplet is never defined and Eq. (11) prescribes an unjustified gradient update. This is not a typographical nit: the triplet loss is listed in Section II.I as a key differentiator, and it appears in the total loss without a formula. The experimental tables also contain internal inconsistencies (e.g., Table I reports LPIPS around 0.03–0.08, while Table II reports 0.56–0.64 for what should be similar tasks), and no code or training details are provided. The paper's own Section V admits limited generalization to other image categories. Taken together, the central empirical claim is unsupported. The reader's verdict of REJECT is appropriate; no adjustment is needed.","tokens_in":7399,"tokens_out":4289,"duration_ms":55312,"concrete_test":"Attempt to reimplement the method exactly as published: write an explicit definition of L_triplet consistent with Eq. (5) (e.g., a negative log-likelihood-ratio loss), substitute it into Eq. (8), and derive ∂L_total/∂ϵθ through the SDXL denoising loop and VAE decoder. If L_triplet cannot be defined unambiguously or the gradient is not computable as written, the central algorithm is underspecified. Alternatively, ask the authors to supply the missing L_triplet formula and the derivation of Eq. (11); absent that, the method is not reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B introduces the method's key 'conceptual triplet loss,' but no definition of L_triplet is ever given. Eq. (5) is a Bayesian re-weighting of Pθ(X|ct), not a loss; Eq. (6) is a malformed gradient expression with mismatched parentheses and undefined terms; and L_triplet then appears in Eq. (8) as part of L_total without any formula. Eq. (11) instructs updating the noise prediction ϵθ by η·∂L_total/∂ϵθ, but L_total is computed on decoded images through adversarial, perceptual, and the unspecified triplet losses. The paper never derives how these image-space losses differentiate through the VAE decoder and the diffusion denoising steps. Since the triplet loss is advertised in Section II.I as the mechanism that 'explicitly align[s] latent directions with target attributes,' the central training objective is underspecified. Without a concrete L_triplet and a valid gradient derivation, the method cannot be implemented or tested, and the claimed 'marked enhancement in image quality and realism' is not reproducible.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes 'Beyond Sliders', an image-editing framework built on Stable Diffusion XL with LoRA adapters, trained with a combination of a 'conceptual triplet loss', perceptual loss, and adversarial training. The stated goal is to improve on Concept Sliders for editing in-the-wild/non-AIGC images, with claims of better realism, identity consistency, and generalization across image categories. The paper presents a method section with equations for the losses, qualitative figures, a small quantitative table comparing CLIP/LPIPS to Concept Sliders, and an ablation study of the perceptual and adversarial components.","tokens_in":7751,"tokens_out":3656,"duration_ms":44869,"significance":"If the method were fully specified and validated, the idea of using a discriminator and perceptual losses to improve diffusion-based concept sliders for real-world images would be a useful contribution. The paper correctly identifies an important limitation of Concept Sliders on in-the-wild images and provides qualitative evidence that the proposed pipeline can produce plausible edits. However, the central mathematical derivation is incomplete and, as written, not meaningful: key losses are undefined, the gradient update is not justified, and several free parameters are unspecified. The quantitative evaluation is too thin to support the claimed 'marked enhancement'. No code or reproducibility details are provided. Because the core algorithm cannot be implemented or checked from the manuscript, the current contribution cannot be assessed rigorously.","major_comments":[{"comment":"The central derivation is not mathematically meaningful. Eq. (5) is a heuristic reweighting of Pθ(X|ct), and Eq. (6) is not a valid consequence: it has mismatched parentheses, contains the expression log(Pθ(X|c+)−Pθ(X|c−)) rather than a log-ratio, and does not follow from the stated Bayesian expansion. Since this is the foundation of the triplet-based control, the method's main mechanism is not defined.","section":"Section III-B, Eqs. (5)-(6)"},{"comment":"L_triplet is never defined. The paper refers to a 'conceptual triplet loss' and includes it in L_total, but no formula or explanation is given for how the triplet loss operates on diffusion latent directions. Citations [26]-[28] describe embedding-space triplet losses, not a loss for adjusting noise predictions. Consequently, Eq. (11) is underspecified and the claimed training procedure cannot be reproduced.","section":"Section III-B, Eqs. (8) and (11)"},{"comment":"Updating εθ(X, ct, t) by η·∂L_total/∂εθ is not justified. L_total is computed on decoded images through the VAE decoder and discriminator/perceptual networks; the manuscript does not derive how these image-space losses differentiate through the decoder and the diffusion denoising steps to produce a gradient with respect to the noise prediction. This is a load-bearing step: without a correct and explicit gradient derivation, the optimization is not well defined.","section":"Section III-B, Eq. (11)"},{"comment":"The 'corresponding real image' X_real,i is undefined. The perceptual loss requires a paired reference image, but no pairing or alignment mechanism is described for arbitrary input photos. In the experiments, real images are drawn from the LFW dataset, yet the method is applied to 3D characters and other in-the-wild inputs; it is unclear which X_real,i is used there. This is a serious gap because both the perceptual loss and the discriminator rely on LFW faces, making the claimed generalization unsupported.","section":"Section III-B, Eq. (7)"},{"comment":"The quantitative comparison is insufficient to support the central claim. No dataset sizes, standard deviations, confidence intervals, or significance tests are reported, and the CLIP differences are tiny (e.g., 29.03 vs. 29.06). LPIPS values for the 'age' row are inconsistent with the ablation table (0.027 vs. 0.557), suggesting different evaluation protocols that are not explained. The reported metrics alone do not establish a 'marked enhancement in image quality and realism'.","section":"Section IV-C, Table I"},{"comment":"Multiple free parameters are left unspecified: t0 and k in the schedule, the step size η, the loss weights λ_triplet and λ_perp, and the LoRA rank/scale α. Without these, the method cannot be reimplemented. Additionally, Table II's CLIP/LPIPS values (34.89/0.557 for full set) are on a completely different scale from Table I's values, and the text quotes CLIP 34.9 vs. table value 34.89 while describing changes; this internal inconsistency further weakens confidence in the empirical claims.","section":"Sections III-A/III-B and Table II"}],"minor_comments":[{"comment":"The forward process notation is loose: the variance schedule βt is not defined, and the arrow assignment xt ← ... is not standard. Eq. (2) omits the expectation and uses ∥...∥2 instead of ∥...∥^2.","section":"Section III-A, Eq. (1)"},{"comment":"There are numerous typographical and grammatical issues, e.g., 'precedant' in Fig. 1, 'we utilize Stable Diffusion XL' without specifying SDXL version, and inconsistent capitalization of 'Beyond Sliders'. Figure captions and tables would benefit from clearer formatting.","section":"General"},{"comment":"The paper compares only to Concept Sliders. Given the claim of state-of-the-art performance, comparisons to other recent editing methods (e.g., Reversion, TexSliders, or direct editing baselines) and a user study would be expected.","section":"Section IV"},{"comment":"The Limitation section states that performance on 'other image categories' has not been fully validated, which directly contradicts the abstract's claim of generalization 'across diverse image categories'. This tension should be addressed explicitly.","section":"Section V"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an early draft. The primary issues are not stylistic: the core triplet loss is undefined, the gradient update is not derived, and the experimental evaluation is too limited to support the claims. If the authors can provide a complete mathematical formulation and rigorous experiments with reproducible details, a fresh submission would be more appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this paper is not close to being a reproducible method. The core 'conceptual triplet loss' L_triplet appears in the total loss (Eq. 8) but is never defined anywhere. Eq. (6) is a malformed gradient expression with mismatched parentheses, and Eq. (11) tells you to update the noise prediction by the gradient of a loss computed on decoded images, with no derivation of how that gradient flows through the VAE decoder and the denoising steps. On its own terms, the math doesn't work.\n\nWhat's genuinely useful here is the motivation. Combining LoRA sliders with adversarial training and a perceptual loss to improve realism on in-the-wild photos is a reasonable research direction, and the paper's qualitative examples (if they hold up) suggest the combination can be effective. The authors also list honest limitations: no experiments with flow-matching models, and limited generalization beyond faces.\n\nBut the execution is too raw. The perceptual loss in Eq. (7) requires 'the corresponding real image' X_real,i, which is never defined. The real images come from LFW (faces), yet the method claims to work on 3D characters and arbitrary natural scenes, so it's unclear what X_real,i is for a non-face input. The triplet loss is the centerpiece of the method, and it's absent. The ablation numbers in Table II (CLIP 34.89, LPIPS 0.557) are on a totally different scale from the main results in Table I (CLIP ~25-29, LPIPS 0.03-0.08), which makes the quantitative evaluation suspect. There are no error bars, no code, no training details, and no specification of the free parameters (t0, k, eta, lambda weights, LoRA rank).\n\nThis paper is for a reader who wants a hint of a research direction, not for anyone who wants to build on the method. It deserved to be shared as a workshop idea or tech report, not as a serious claim of a new framework.\n\nRecommendation: desk reject. Not because the idea is bad, but because the paper doesn't define its method. If the authors provide the missing definitions, derivations, and consistent experiments, it could be worth another look, but in its current form, referees would be reviewing a set of equations that don't parse.","headline":"The paper's central 'conceptual triplet loss' is never defined, and the equations that matter don't parse; despite a sensible motivation, the method is not reproducible.","tokens_in":8161,"tokens_out":3170,"would_cite":false,"duration_ms":33634,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that adding GAN-style supervision to diffusion-based sliders makes image edits on real-world photos more realistic and faithful than prior concept-slider methods.","keywords":["diffusion models","image editing","concept sliders","LoRA","adversarial training","perceptual loss","triplet loss","in-the-wild images"],"falsifier":"Run editable real-world photos from non-face categories (for example, animals, vehicles, or indoor scenes) through Beyond Sliders and Concept Sliders, and compare CLIP and LPIPS: if the margin seen on faces disappears or reverses, the claimed generalization to arbitrary image categories fails. Also, replace the reference images in the perceptual loss with randomly chosen unrelated faces; if output quality is unchanged, the perceptual term is not doing the identity-preserving work attributed to it.","tokens_in":7336,"feed_emoji":"🎛️","tokens_out":4215,"duration_ms":44145,"temperature":0.7,"pith_summary":"The paper claims that Concept Sliders-style diffusion editing underperforms on non-AIGC and real-world images because it optimizes mostly for text alignment, and that adding GAN-style supervision fixes this. Beyond Sliders wraps a Stable Diffusion XL generator with LoRA adapters in adversarial training, a perceptual loss, and a conceptual triplet loss to push edited images toward realism. The authors report that their edits score higher on CLIP and lower on LPIPS than Concept Sliders across text-guided and in-the-wild tasks, and that the method can compose multiple sliders and repair distorted hands or room layouts. If the claim holds, diffusion-model editing can be applied to ordinary photos, 3D characters, and celebrity images without sacrificing identity or visual fidelity.","feed_headline":"GAN-style losses sharpen diffusion edits on real photos","feed_subtitle":"Adding adversarial and perceptual training to LoRA sliders keeps edits close to the original while matching prompts more closely.","key_machinery":"The central mechanism is a LoRA low-rank adapter on Stable Diffusion XL trained with a weighted sum of three losses: a conceptual triplet loss that raises the likelihood of a positive attribute and lowers a negative one around a neutral concept; a perceptual loss that matches deep features between the edit and a reference real image; and an adversarial loss from a discriminator trained on real face images. The relative weights of perceptual and triplet losses shift during training, with the perceptual weight large early and the triplet weight increasing later, so the model first learns realism and then sharpens semantic control.","core_discovery":"Beyond Sliders' central discovery is that low-rank concept directions learned by sliders can be steered by a total loss combining triplet, perceptual, and adversarial terms, and that this combination is what transfers editing from AIGC images to uncontrolled real photos. The generator remains Stable Diffusion XL plus LoRA; a discriminator sees real images from a face dataset, and the perceptual term compares generated edits with reference real images at the feature level. The paper argues this joint optimization yields edits that are semantically aligned and visually consistent, and it demonstrates higher CLIP and lower LPIPS than the baseline on age, chubby, long hair, and muscular editing","pith_inferences":["The paper's stated limitations—no tests on Flow-Matching models and unvalidated performance outside faces—suggest a natural next test: swapping the SDXL backbone for a rectified-flow model and measuring whether the same losses transfer.","If the perceptual loss is the main driver of the reported gains, then the method should work even without the triplet loss; a head-to-head ablation removing triplet while keeping adversarial and perceptual terms would isolate which component carries the improvement.","Because the perceptual term references unpaired real images, its benefit may be mainly a global realism prior rather than identity preservation; testing with random versus matched references would show whether pairing matters.","A direct extension would replace the face-specific discriminator with a category-agnostic patch discriminator, which may close the generalization gap the paper concedes for non-face categories."],"forward_implications":["Edits of real photographs and non-AIGC images become more natural than Concept Sliders, with higher CLIP alignment and lower LPIPS distortion.","Multiple sliders can be composed additively on the same image, so attributes like chubby, age, and smiling stack without breaking identity.","The same framework can repair flawed generations, such as malformed hands and physically implausible room layouts, as slider strength increases.","Because LoRA keeps base weights frozen, the method inherits cheap per-concept adapters and adjustable edit strength at inference via scaling."],"fun_headline_variants":["Diffusion edits get real with GAN losses","Beyond sliders: real-photo edits via adversarial tuning","Adversarial losses make diffusion sliders work on real images","Gan-boosted diffusion edits transfer to uncontrolled real photos","From AIGC to real: diffusion editing via triplet and GAN losses"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that comparing an edited image against an arbitrary real face image, with no pairing or alignment, teaches the model to make any input photograph look real; the paper's own limitation section concedes that performance outside facial editing is not fully validated.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion edits get real with GAN losses","Beyond sliders: real-photo edits via adversarial tuning","Adversarial losses make diffusion sliders work on real images","Gan-boosted diffusion edits transfer to uncontrolled real photos","From AIGC to real: diffusion editing via triplet and GAN losses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000315,"raw_usage":{"total_tokens":1564,"prompt_tokens":624,"completion_tokens":940,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":368,"completion_tokens_details":{"reasoning_tokens":865}},"tokens_in":368,"tokens_out":940,"duration_ms":8624,"temperature":1.0,"reasoning_tokens":865,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:53:01.346433+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run editable real-world photos from non-face categories (for example, animals, vehicles, or indoor scenes) through Beyond Sliders and Concept Sliders, and compare CLIP and LPIPS: if the margin seen on faces disappears or reverses, the claimed generalization to arbitrary image categories fails. Also, replace the reference images in the perceptual loss with randomly chosen unrelated faces; if output quality is unchanged, the perceptual term is not doing the identity-preserving work attributed to it.","supporting_citations":[],"review_version":1}