{"id":"e406db76-ed80-4d3f-ad43-b42d6a7a2214","arxiv_id":"2504.15723","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A stage-wise latent injection schedule, source latents in early denoising steps and reference latents in later steps, is proposed for zero-shot text and reference guided image editing with Stable Diffusion.","lead":"This paper describes a zero-shot image editing method for diffusion models that combines text prompts and reference images through a two-stage latent injection scheme, with shape anchored in early steps and attributes injected in later steps. The approach is practical and avoids fine-tuning, but the paper's state-of-the-art claim is not supported by its own 12-image evaluation, and the method as written contains unresolved notation conflicts.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 and §3.2–3.3 specify opposite stage orders, and tearly is never reported, so the paper does not define the central stage-wise schedule.","rationale":"The reader's weakest assumption and my concern coincide: the stage-wise timing hypothesis is load-bearing, and it is neither justified nor uniquely specified. I agree with the CONDITIONAL verdict. The sharpest technical problem is not merely that tearly is tuned on the evaluation data or that 'early' and 'late' are informally defined; it is that the prose and the pseudocode describe two different pipelines, and no numeric tearly is supplied. Under the DDIM loop in Algorithm 1, Section 3.2's t ≤ tearly and Algorithm 1's t ≥ tearly cannot both be the intended schedule. The equations also do not connect the attention K/V substitution in Eqs. (2) and (4) to the ϵθ calls in Algorithm 1, so an implementer must guess how injection is applied. This blocks verification of the claimed state-of-the-art result. I do not escalate to REJECT because the defect is a specification gap rather than a demonstrated contradiction with known results; a corrected algorithm, an explicit tearly, and released code could resolve it. Hence UNCHANGED: CONDITIONAL remains the appropriate verdict.","tokens_in":9796,"tokens_out":5704,"duration_ms":54020,"concrete_test":"Run the AFHQ 12-image comparison twice from the same code: (A) follow Algorithm 1 literally with tearly = 30, applying the source-latent branch when t ≥ 30; (B) follow §3.2–3.3 literally, applying shape injection when t ≤ 30. Record Table 1's four metrics plus LPIPS-to-source for each variant, and state which variant reproduces the submitted images. If (A) reproduces the report, the equations describe the wrong method; if (B) reproduces it, Algorithm 1 is wrong; if neither, the results depend on an unspecified tearly. Releasing the code with tearly stated would settle reproducibility.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on an unambiguous \"shape-early, attribute-late\" schedule. The manuscript does not provide one. Section 3.2 assigns shape injection to t ≤ tearly and Section 3.3 assigns attribute injection to t > tearly, whereas Algorithm 1 loops t = T down to 1 and applies the source-latent branch when t ≥ tearly and the reference-latent branch otherwise. Under the standard DDIM convention used in Algorithm 1, high t is the noisy, early stage; so the text labels the late stage as early, while the pseudocode labels the early stage as early. The two specifications cannot both describe the same method. The paper never reports tearly: §4.1 justifies only the guidance scale, and §5.2 discusses injection timing only qualitatively. In addition, Algorithm 1's denoising calls do not match the attention-level K/V redefinition in Eqs. (2) and (4): the pseudocode passes zt with zs*_t or zr*_t to ϵθ without specifying how the injected K/V enter the U-Net. Consequently, at least two different pipelines are consistent with the text, and the claimed comparison in Table 1 cannot be attributed to the proposed stage-wise mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a zero-shot image editing framework for Stable Diffusion that unifies text-guided and reference-guided edits by combining DDIM inversion, null-text inversion (NTI), and a stage-wise attention-injection schedule: source-latent keys/values are injected in early denoising steps for shape preservation, and reference-latent keys/values are injected in later steps for attribute transfer. The claimed contributions are a mask-free, fine-tuning-free pipeline and a new state-of-the-art on AFHQ relative to InjectFusion and DiffuseIT. The central load-bearing idea is the timing hypothesis that early steps control spatial structure and later steps control fine attributes.","tokens_in":9962,"tokens_out":4585,"duration_ms":44492,"significance":"If the stage-wise schedule were precisely defined and substantiated, the framework could be a useful zero-shot editing baseline, since it combines published components (DDIM inversion, NTI, attention injection) in a way that has not been tested in exactly this form. The qualitative figures show coherent edits, and the use of timestep-specific null embeddings is a reasonable direction. However, the manuscript currently does not define a single unambiguous method: the text, equations, and Algorithm 1 conflict on the stage order and on how injection enters the U-Net, and the quantitative evaluation is a 12-image table that does not support the stated state-of-the-art claim. I do not see a circularity problem in the derivation; the issues are under-specification, internal inconsistency, and insufficient evidence. The significance is therefore conditional on a substantial revision.","major_comments":[{"comment":"The stage schedule is specified contradictorily. The text assigns shape injection to t ≤ tearly (Eq. 2) and attribute injection to t > tearly (Eq. 4), while Algorithm 1 loops t = T down to 1 and applies the source-latent branch when t ≥ tearly and the reference-latent branch otherwise. Under the standard DDIM convention used in the loop, high t is the early/noisy stage, so the algorithm's source branch is the early stage; the text's inequality therefore either reverses the stages or, depending on the unreported value of tearly, labels the late stage as early. Since the stage-wise order is the paper's central contribution, this conflict makes the method undefined. In addition, tearly is never reported in §4.1 or §5.2, so the key hyperparameter of the central mechanism is absent.","section":"§3.2–3.3 and Algorithm 1"},{"comment":"Algorithm 1 does not implement the attention-level injection described in Eqs. (2) and (4). The pseudocode calls εθ(zt, zs*_t, ...) or εθ(zt, zr*_t, ...), but the equations define injection by replacing the keys and values inside the U-Net's self-attention, which is not expressible as an extra argument unless the internal mechanism is specified. The algorithm also computes zt−1 from εθ and then computes a separate ˜εt that is never used to update zt−1, and it introduces a mixing coefficient α that is not defined anywhere; §4.1 only reports a guidance scale m = 7. As written, the pseudocode is not executable and at least two different pipelines are consistent with the text, so the reported comparisons cannot be attributed to a specific stage-wise injection mechanism.","section":"Algorithm 1 and Eqs. (2)–(5)"},{"comment":"The quantitative evidence does not support the claim that the method 'outperforms InjectFusion and DiffuseIT.' Table 1 is based on 12 images with no error bars, no per-image statistics, and no significance test. On the two Palette metrics, the proposed method is worse than DiffuseIT (Palette src 0.4846 vs 0.3425; Palette ref 1.0237 vs 0.7242) and worse than both baselines on Palette ref. The paper's own text concedes the 'marginally higher discrepancy' but then concludes an 'optimal trade-off' and state-of-the-art performance, which is not warranted by the two CLIP margins alone. Please report the full distribution, error bars, and a statistical test, or substantially temper the conclusion.","section":"Table 1 and §4.4"},{"comment":"The central timing hypothesis is only qualitatively justified. Section 4.1 states that settings were chosen 'after extensive empirical evaluation' but does not report tearly, α, or the search range; Section 5.2's timing ablation is purely qualitative, with example images and no quantitative metrics for structural or semantic consistency. Since the contribution is the stage-wise schedule itself, the absence of a quantitative timing study (e.g., a sweep over tearly with Palette/CLIP scores) leaves the main claim unsupported, especially because the hyperparameters appear to be tuned without a held-out protocol.","section":"§4.1 and §5.2"}],"minor_comments":[{"comment":"The word 'connectivity' should be 'concatenation,' and the operator ⊕ is not defined in the text.","section":"Eq. (4)"},{"comment":"The guidance-scale notation is inconsistent: §3.1 and Eq. (3) use w and state w = 7.5, while §4.1 reports m = 7. Please unify the symbol and value.","section":"§3.1, §3.2, §4.1"},{"comment":"The figure contains the undefined abbreviation 'NPI' and labels that are difficult to read; please clarify.","section":"Fig. 3"},{"comment":"The dataset list includes Oxford-IIIT Pets, DreamBooth, AFHQ, CelebA, COCO, and Custom Diffusion, but the quantitative evaluation uses only 12 AFHQ images; the paper should state which dataset and images support each qualitative and quantitative result.","section":"§4.2"},{"comment":"The caption mentions 'adding a hat,' but the displayed text prompts show blue hair, tiger, dog, and cat transformations; the caption should match the actual prompts.","section":"Fig. 6 caption"},{"comment":"The supplementary URLs contain line breaks and spaces and should be formatted as a proper reference list.","section":"§7"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a plausible core idea but is not currently reproducible or self-consistent. The conflicts between the textual description, the equations, and Algorithm 1 are load-bearing, not cosmetic. The quantitative evaluation is far too weak for the claimed state-of-the-art. I would encourage the editor to invite a revision that resolves the schedule definition, reports tearly and α, fixes the pseudocode, and replaces the 12-image table with a statistically grounded evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is a stage-wise schedule for combining source and reference latents during denoising: early steps inject source latents into self-attention K/V to hold shape, later steps switch to reference latents to transfer attributes. That split is a real, if modest, extension of NTI plus RIVAL-style attention injection, and the qualitative figures do show coherent edits. The ablation in Fig. 7 supports that both the null-text inversion and the injected schedule matter, which is more than many papers in this space bother to check.\n\nNow the soft spots, and one of them is load-bearing. The paper never actually pins down the schedule. Section 3.2 says shape injection runs for t ≤ tearly and attribute injection for t > tearly; Algorithm 1 does the opposite, applying the source branch for t ≥ tearly. Under the DDIM convention where t=T is noise and t=1 is the final clean image, those are mutually exclusive. The paper also never reports the value of tearly, and §5.2 only discusses timing qualitatively. So the central mechanism—the one thing that distinguishes this from RIVAL or InjectFusion—is not defined well enough to reproduce.\n\nThe quantitative support is also weak. Table 1 reports 12 images with no error bars, and the proposed method is worse than DiffuseIT on both Palette metrics. Claiming state-of-the-art on that basis is a stretch. No code is provided, and the pseudocode does not specify how the injected K/V from the source or reference latent actually replace the U-Net's self-attention, beyond the equations in §3.2–3.3.\n\nCircularity is not a real concern: nothing is derived from its own outputs. But guidance and timing hyperparameters appear to be tuned on the evaluation set, which is a minor issue.\n\nBottom line: the underlying idea is plausible and the qualitative results are suggestive, so this deserves a careful referee rather than a desk reject. But as submitted it needs major revision: specify the schedule unambiguously, report tearly, supply code, and add a proper evaluation with more images, error bars, and a fair treatment of metrics.","headline":"A plausible stage-wise attention-injection heuristic for diffusion editing, undermined by an internally contradictory schedule specification and a thin 12-image quantitative evaluation.","tokens_in":10558,"tokens_out":3355,"would_cite":false,"duration_ms":31159,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a two-stage attention-injection schedule lets a frozen diffusion model edit images without masks, fine-tuning, or user constraints.","keywords":["zero-shot image editing","diffusion models","DDIM inversion","null-text inversion","attention injection","reference-guided editing","text-guided editing","structure preservation"],"falsifier":"On a fixed set of source-reference pairs, sweep the switching time $t_{\\mathrm{early}}$ across all 50 DDIM steps and plot structure-preservation and attribute-transfer metrics; if the best result does not fall in the later half of the schedule, or if the quality curve is flat, the stage-wise timing claim is not load-bearing. A second decisive check is to reverse the schedule so attributes are injected early and shape late; the method should degrade markedly if the timing hypothesis is correct.","tokens_in":9500,"feed_emoji":"🖼️","tokens_out":9646,"duration_ms":75305,"temperature":0.7,"pith_summary":"The paper is trying to establish that a single, training-free diffusion pipeline can perform both text-guided and reference-guided image editing without masks, fine-tuning, or user constraints. The key move is to partition the denoising process into an early shape-injection stage, where attention keys and values come from the source image's inverted latent (its compressed internal representation), and a later attribute-injection stage, where they come from the reference image's inverted latent. If this holds, structural preservation and attribute transfer become two phases of one process rather than competing objectives, and null-text inversion, which was limited to text prompts, gains a reference-guided counterpart.","feed_headline":"Two-stage latent injection edits images with no masks or fine-tuning","feed_subtitle":"Early denoising steps freeze the source layout; later steps pull texture and expression from a reference image.","key_machinery":"The central mechanism is a stage-wise latent injection schedule inside the denoising U-Net. In the shape-injection stage (early timesteps) the self-attention keys and values are the source image's inverted latents, which pins down geometry; in the attribute-injection stage (later timesteps) the keys and values become the reference image's inverted latents concatenated with the current latent, which imports fine-grained appearance. The schedule is carried by the attention formula $softmax(QK^\\top/\\sqrt{d})V$ and by timestep-specific null-text embeddings that serve as a structural anchor for classifier-free guidance.","core_discovery":"The central claim is that the denoising trajectory carries structure early and attributes late, and that this ordering can be exploited by swapping which latents feed the self-attention layers. In the early timesteps, the query attends to keys and values taken from the source image's DDIM-inverted latent $z^{s*}_t$, so the edited image inherits the source layout, pose, and object shape; after a transition, the keys and values are taken from the reference image's inverted latents $z^{r*}_t$ concatenated with the current latent, so texture, expression, and style migrate onto the already-preserved layout. The method also replaces the ordinary empty prompt with timestep-specific null-text embeddings optimized for the source image, which anchor reconstruction during classifier-free guidance. The paper reports that on AFHQ dog images this schedule outperforms InjectFusion and DiffuseIT in semantic consistency and structural coherence while preserving the source's facial features, ear shapes, and pose.","pith_inferences":["Editorial inference: because the paper gives no value for the switching time $t_{\\mathrm{early}}$, the method is likely sensitive to that hyperparameter; a content-aware switching rule (for example, switching when the source layout has stabilized) is a natural test the paper does not run.","Editorial inference: the time-axis split suggests a layer-wise analogue — if early timesteps encode shape and later timesteps encode appearance, different U-Net layers may also encode them, so selective layer swapping could give finer localization than a global time threshold.","Editorial inference: the reported CLIP-ref gain is what the mechanism predicts, since later attention keys and values come directly from the reference latent; a per-step attribution study would separate the attention swap's contribution from null-text inversion's contribution."],"forward_implications":["The same Stable Diffusion checkpoint handles both text-guided edits (hair color, species change) and reference-guided edits (texture, expression, style) with no per-task training.","Editing quality is governed by injection timing: injecting attributes too early warps the layout, and injecting too late leaves the semantic change incomplete.","Removing Null-Text Inversion introduces distortions, and removing the two-stage injection produces inconsistent attribute blending, so both components are cast as load-bearing.","On the AFHQ subset, the method achieves the highest CLIP agreement with both source and reference among InjectFusion, DiffuseIT, and itself, which the paper reads as evidence of the structure-attribute balance."],"supporting_citations":[{"why":"Supplies the Stable Diffusion V1.5 backbone on which inversion, injection, and generation run.","marker":"[26]"},{"why":"Provides the DDIM inversion used to build the source and reference latent trajectories.","marker":"[29]"},{"why":"Contributes null-text inversion, the source of the timestep-specific null embeddings that anchor structure.","marker":"[21]"},{"why":"Introduces the inversion-chain idea and RIVAL baseline that motivates injecting latents into attention layers.","marker":"[34]"},{"why":"Defines the attention operation the injection blocks use.","marker":"[31]"},{"why":"Defines classifier-free guidance, the balancing mechanism between structure and semantics.","marker":"[11]"},{"why":"Provides the InjectFusion baseline compared in the AFHQ experiments.","marker":"[14]"},{"why":"Provides the DiffuseIT baseline for reference-guided translation comparison.","marker":"[17]"},{"why":"Supplies the AFHQ dataset used in the quantitative and qualitative evaluation.","marker":"[4]"}],"fun_headline_variants":["Stage-wise latent swap keeps structure, swaps style","Early structure, late attributes: zero-shot diffusion edit","No masks, no fine-tuning: staged latent injection","Shape first, texture later: zero-shot diffusion edit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that early denoising timesteps encode spatial layout and later timesteps encode fine attributes, so switching attention keys and values from the source latent to the reference latent at a fixed threshold preserves structure while transferring style; the paper never quantifies that threshold, and its prose and pseudocode disagree about whether shape injection runs on early or late timesteps.","fun_headline_variants_meta":{"raw":{"variants":["Stage-wise latent swap keeps structure, swaps style","Early structure, late attributes: zero-shot diffusion edit","No masks, no fine-tuning: staged latent injection","Shape first, texture later: zero-shot diffusion edit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1512,"prompt_tokens":847,"completion_tokens":665,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":603}},"tokens_in":463,"tokens_out":665,"duration_ms":6163,"temperature":1.0,"reasoning_tokens":603,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:19:14.064408+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a fixed set of source-reference pairs, sweep the switching time $t_{\\mathrm{early}}$ across all 50 DDIM steps and plot structure-preservation and attribute-transfer metrics; if the best result does not fall in the later half of the schedule, or if the quality curve is flat, the stage-wise timing claim is not load-bearing. A second decisive check is to reverse the schedule so attributes are injected early and shape late; the method should degrade markedly if the timing hypothesis is correct.","supporting_citations":[{"cited_title":"Null-text inversion for editing real im- ages using guided diffusion models","cited_arxiv_id":null,"evidence_quote":"Contributes null-text inversion, the source of the timestep-specific null embeddings that anchor structure."},{"cited_title":"Training- free content injection using h-space in diffusion models","cited_arxiv_id":null,"evidence_quote":"Provides the InjectFusion baseline compared in the AFHQ experiments."},{"cited_title":"Stargan v2: Diverse image synthesis for multiple domains","cited_arxiv_id":null,"evidence_quote":"Supplies the AFHQ dataset used in the quantitative and qualitative evaluation."}],"review_version":1}