{"id":"15e60e40-cccf-4623-9767-d16b1bba6acf","arxiv_id":"2507.01953","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FreeMorph combines spherical interpolation with attention feature blending and a step-wise schedule to produce tuning-free, identity-preserving image morphing in under 30 seconds.","lead":"FreeMorph generates smooth morphing sequences between two arbitrary images without per-image training, by blending the self-attention features of a pretrained diffusion model. It claims a 10x to 50x speedup over existing morphing methods while handling inputs with different semantics and layouts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation 8's high-frequency noise injection is mathematically ill-posed as written: it either leaves the latent unchanged or replaces it with pure Gaussian noise, so the core procedure is underspecified and not reproducible.","rationale":"The reader's verdict is CONDITIONAL, citing missing code, unspecified mask m, and hand-picked hyperparameters. My stress-test identifies a more specific and more severe technical problem: Eq. 8 as printed cannot perform the claimed high-frequency noise injection because IFFT(FFT(z)) is the identity and IFFT(FFT(g)) is pure Gaussian noise; the binary mask m is never applied in the frequency domain. This is a load-bearing flaw because the ablation (Table 3) shows that removing noise injection degrades performance, so the reported state-of-the-art numbers depend on an operation that is not actually defined. However, the flaw is potentially fixable by clarifying the intended mask multiplication and releasing code, so it does not by itself invalidate the empirical claim; it makes the paper currently unverifiable as written. This reinforces the existing CONDITIONAL verdict rather than changing it to a different category. The reader did not notice the full algebraic inconsistency of Eq. 8, only the missing specification of m, hence partial agreement.","tokens_in":13460,"tokens_out":3911,"duration_ms":45015,"concrete_test":"Obtain the official code (or a detailed mask specification) and re-run one MorphBench image pair with the corrected high-frequency noise injection, implementing z = IFFT(m * FFT(z) + (1-m) * FFT(g)) with the authors' exact binary mask m. If the reported outputs and Table 1 metrics cannot be reproduced with this corrected formula, and instead the literal Eq. 8 reading (z unchanged when m=1, z=g when m=0) is what was actually used, then the paper's description is wrong and the claimed benefit of 'high-frequency noise injection' is unsubstantiated. If the corrected formula does reproduce the numbers, then Eq. 8 must be revised in any final version.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of state-of-the-art morphing performance rests on a pipeline whose key high-frequency noise injection step is not correctly specified. In Sec. 3.4, Eq. 8 defines z := IFFT(FFT(z)) if m=1 and z := IFFT(FFT(g)) if m=0, where g is standard Gaussian noise and m is described as a binary high-pass filter mask of the same size as z. Taken literally, this operation is not a high-pass filter: IFFT(FFT(z)) equals z exactly, and IFFT(FFT(g)) equals g, so the mask m has no frequency-selective effect. The if m=1 / if m=0 syntax suggests a global scalar condition, but m is a per-location mask; the equation never multiplies FFT(z) by the mask. A correct high-pass injection would be something like z = IFFT(m * FFT(z) + (1-m) * FFT(g)), with m nonzero only at high frequencies. The paper does not provide this. The ablation in Table 3 shows that removing noise injection changes overall LPIPS from 162.99 to 188.61, so this component materially affects the reported results. Because no code is released and the equation cannot be executed as intended, the results in Table 1 are not independently verifiable. This is not a minor typo: the exact mechanism (mask construction, whether m is elementwise or global, and how g is scaled) determines whether the method performs frequency-domain filtering, identity mapping, or full latent replacement, each of which would produce very different outputs.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FreeMorph, a tuning-free image morphing method built on Stable Diffusion. The method combines spherical interpolation in latent space, guidance-aware modifications to self-attention key/value features, a step-oriented variation trend, and a high-frequency noise injection step, with a staged forward/reverse denoising schedule controlled by hyperparameters lambda_1 to lambda_4. The authors evaluate on the existing MorphBench dataset and on a newly introduced Morph4Data dataset, reporting lower LPIPS, FID, and PPL than IMPUS, DiffMorpher, and a spherical-interpolation baseline, together with a user study and qualitative comparisons. The central claims are that FreeMorph is the first tuning-free method to handle inputs with different semantics or layouts, produces high-fidelity transitions in under 30 seconds, and outperforms trained baselines.","tokens_in":13971,"tokens_out":5653,"duration_ms":73668,"significance":"If the central claims hold, the contribution is practically significant: a tuning-free diffusion-based image morphing method that works across semantic and layout gaps would be a useful advance, and the reported speed-up over trained baselines is substantial. The paper has concrete strengths: the component-level ablations in Table 3 are informative, the qualitative figures show visible differences from the baselines, and the newly collected Morph4Data attempts to cover a regime that existing benchmarks under-represent. However, several load-bearing details are either incorrect or underspecified as written, and the evaluation has reproducibility limitations, so I cannot recommend acceptance in the current form.","major_comments":[{"comment":"The high-frequency Gaussian noise injection step is mathematically ill-posed as written. For m=1, z := IFFT(FFT(z)) is exactly the identity transform; for m=0, z := IFFT(FFT(g)) equals g if g is a latent-space noise vector. Since m is described as a binary high-pass mask of the same size as z, the equation never multiplies the spectrum by m and has no frequency-selective effect. The ablation in Table 3 shows that removing noise injection changes overall LPIPS from 162.99 to 188.61, so this component materially affects the reported numbers. Please replace Eq. (8) with a correct masking operation such as z = IFFT(m * FFT(z) + (1-m) * FFT(g)), specify how m is constructed and how g is scaled, and state whether the implementation follows the corrected formula.","section":"Sec. 3.4, Eq. (8)"},{"comment":"The spherical interpolation formula for z_{0-j} is inconsistent with the stated range j in [1,J]. With j=1, the coefficient sin((1-j)*phi)/sin(phi) equals zero, so the first intermediate latent would not be close to the left input; with j=J, sin((1-J)*phi) is generally not sin(phi). The likely intended schedule involves a normalized interpolation parameter such as j/(J+1) inside the spherical-interpolation angle. Please correct Eq. (4) or clearly define the interpolation parameter, because this formula defines the initialization on which the entire pipeline depends.","section":"Sec. 3.2, Eq. (4)"},{"comment":"All quantitative results are reported as point estimates without standard deviations, confidence intervals, or significance tests. FreeMorph has stochastic components, and the hyperparameters lambda_1 to lambda_4 are hand-picked, so the observed margins over baselines could be within run-to-run variability. Please report multiple runs or seeds with error bars, add statistical significance tests for the main comparisons, and include a sensitivity analysis for lambda_1, lambda_2, lambda_3, and lambda_4.","section":"Tables 1-3"},{"comment":"Morph4Data is a newly curated evaluation dataset, but it is not released and the paper does not provide detailed curation criteria or source breakdowns beyond the four class descriptions. Without access to the dataset or a precise construction protocol, the Morph4Data numbers in Tables 1 and 3 cannot be reproduced or extended by other researchers. Please release the dataset or provide a detailed protocol and a public download link.","section":"Sec. 4, Morph4Data"},{"comment":"The main quantitative comparison omits AID and Smooth Diffusion, which are discussed only qualitatively in the appendix. If these methods are considered baselines for the morphing task, the state-of-the-art claim is incomplete. Please include their quantitative results in Table 1 or clearly justify why they are excluded from the metric-based comparison.","section":"Sec. 4.1 and Appendix B.1"}],"minor_comments":[{"comment":"The sentence 'Note that z_{t-j}, z_{t-left} and z_{t-right} are derived based on Eq. 3' is unclear because Eq. 3 is the attention formula, not the DDIM inversion or interpolation formula; please correct the cross-reference.","section":"Sec. 3.2"},{"comment":"The notation IFFT(FFT(g)) implicitly assumes g is defined in the same spatial or latent domain and has the same size as z; please state the domain, size, and normalization of g explicitly.","section":"Sec. 3.4, Eq. (8)"},{"comment":"The instruction 'Add text-conditioned features' is vague; the mechanism by which text embeddings from LLaVA captions are injected into the denoising process, such as classifier-free guidance with averaged or interpolated text embeddings, should be specified.","section":"Algorithm 1, step 7"},{"comment":"The user study reports only preference percentages; please provide the number of pairwise comparisons per participant, the randomization protocol, and inter-subject agreement or a related measure of reliability.","section":"Sec. 4, User studies"},{"comment":"The section on GAN-based morphing ends with 'Additional evaluations and discussions will be included in the revised version,' which indicates that the comparison with Neural Crossbreed is incomplete as submitted; please either complete this section or remove the claim that it supports.","section":"Appendix B.3"},{"comment":"The paper claims a 10x to 50x speed-up but reports only 'under 30 seconds' on an A100 GPU; please include the exact runtime of each baseline and the timing protocol so that the speed comparison is verifiable.","section":"Sec. 4, Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the qualitative results are often convincing, but the incorrect formulation of Eq. (8) and the ambiguity in Eq. (4) are load-bearing reproducibility issues. I would be willing to reconsider after a thorough revision that fixes these equations, adds sensitivity and significance analysis, and releases the dataset and code. Given the absence of code and the unresolved equation, I cannot currently verify the reported numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: FreeMorph has a genuinely new combination—spherical latent interpolation plus K/V attention blending from the two inputs and a step-wise schedule—and it seems to produce morphs that are visually smoother than DiffMorpher/IMPUS in the examples shown. The efficiency claim (under 30s vs 30 min) is believable. The ablations are reasonably thorough and the failure cases are honestly noted.\n\nThe problem is that the paper as written does not actually specify the high-frequency noise injection in Eq. 8. Taken literally, IFFT(FFT(z)) = z and IFFT(FFT(g)) = g, so the binary mask m has no effect at all; the equation either leaves the latent unchanged or replaces it with pure Gaussian noise. The ablation says removing this component changes LPIPS from 163 to 189, so the step matters. Without a corrected formula and a clear description of how m is built, the method is not reproducible. That's not a cosmetic gripe.\n\nAlso soft around the edges: no code, no released dataset (Morph4Data), no error bars or significance tests, and lambda1-4 are hand-picked without sensitivity analysis. The quantitative comparison omits AID and Smooth Diffusion, even though they appear in the appendix qualitatively. The claim of 'first tuning-free generalized morphing' is plausible but depends on the reader granting that the implementation details are right, which I can't do with Eq. 8 in its current form.\n\nThat said, nothing here is fatal conceptually. The core pipeline makes sense and the components are standard enough that an attentive reader could probably reconstruct a working version. The paper is worth a serious referee—not a desk reject—but the referee should demand code, a fixed Eq. 8, error bars, and a sensitivity analysis on the lambdas before publication. If those arrive, the SOTA claim becomes credible.","headline":"A promising tuning-free morphing pipeline with a genuine new combination of attention blending tricks, but Eq. 8 is underspecified to the point of non-reproducibility, and the missing code/data keep the SOTA claim from being fully credible.","tokens_in":14325,"tokens_out":2111,"would_cite":false,"duration_ms":21959,"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":"FreeMorph makes image morphing tuning-free, handling different semantics and layouts, with transitions in under 30 seconds.","keywords":["image morphing","diffusion models","tuning-free","self-attention control","spherical interpolation","DDIM","Stable Diffusion","image editing"],"falsifier":"A concrete test: fix the hyperparameters $\\lambda_1$ through $\\lambda_4$ and run FreeMorph on pairs ranging from near-identical to semantically unrelated. If the optimal schedule drifts with pair type, or if re-running with different random high-frequency masks $m$ changes LPIPS or PPL by more than the reported gap to DiffMorpher, the tuning-free generalization claim is weaker than stated.","tokens_in":13278,"feed_emoji":"🎞️","tokens_out":6139,"duration_ms":59617,"temperature":0.7,"pith_summary":"FreeMorph claims to be the first image morphing method that needs no per-pair training and still works when the two inputs differ in subject matter or layout. The paper's central argument is that the real obstacle is not interpolation itself but the lack of guidance inside the diffusion model's attention: if the self-attention keys and values come from the input images at the right stages of denoising, the transition becomes directional and identity-preserving. On that basis the authors report state-of-the-art scores on LPIPS, FID, and PPL on the existing MorphBench and on a new Morph4Data benchmark, while running 10 to 50 times faster than trained alternatives. If right, this turns morphing from a per-pair fine-tuning job into a sub-30-second operation usable in animation and editing pipelines.","feed_headline":"Morph any two images in under 30 seconds, no training","feed_subtitle":"New diffusion method blends attention features to morph images with different semantics or layouts.","key_machinery":"The load-bearing object is the self-attention module of the pre-trained U-Net. At each denoising step FreeMorph computes the key and value features of the two input-image latents and then replaces the generative attention with one of three blends: an equal average of the two inputs' attention outputs (Eq. 5), an average over all interpolated latents (Eq. 6), or a linearly weighted blend with weight $\\alpha_j = j/(J+1)$ (Eq. 7). Four thresholds $\\lambda_1, \\lambda_2, \\lambda_3, \\lambda_4$ schedule which replacement is active across the 50 forward-diffusion and reverse-denoising steps. Identity preservation, directional drift, and smoothness each come from which attention replacement is active at which stage.","core_discovery":"The paper's discovery is a tuning-free recipe built on Stable Diffusion and DDIM: caption both inputs with a vision-language model, invert both images to latents, spherically interpolate the latents, and then overwrite the self-attention of the denoiser so that its keys and values are derived from the two inputs rather than from the current noisy latent. During forward diffusion, a prior-driven attention averages over all interpolated latents to keep the noise from collapsing into near-identical copies; during reverse denoising, a step-oriented variation trend with weight $\\alpha_j = j/(J+1)$ shifts the attention blend from the source to the target so the sequence moves smoothly from one identity to the other. A high-frequency Gaussian noise injection adds flexibility. The authors claim this outperforms trained baselines on LPIPS, FID, and PPL on both MorphBench and their new Morph4Data, and completes a morph in under 30 seconds.","pith_inferences":["If attention-feature blending is the operative mechanism, the same schedule could be carried over to video diffusion or multi-image interpolation, where the same 'variation trend' problem appears.","The reported failure at human limbs suggests the ceiling is set by Stable Diffusion's prior, so swapping the backbone for a differently trained diffusion model may relocate the failure modes without changing FreeMorph's structure.","The binary high-pass mask $m$ in Eq. (8) is left unspecified, so the exact noise injection is not reproducible from the text alone; a sensitivity study over $m$ would clarify how much of the smoothness comes from that step.","Since text-guided editing is demonstrated as a special case, FreeMorph could serve as a zero-shot editing baseline that does not require inversion-specific machinery."],"forward_implications":["Image morphing no longer needs per-pair fine-tuning; a morph sequence is produced in under 30 seconds on a single A100 GPU.","Inputs with different semantics or layouts become tractable, not just near-identical pairs, because the guidance comes from attention features rather than from per-pair LoRA weights.","The method establishes new state-of-the-art numbers on MorphBench and on the newly introduced Morph4Data on LPIPS, FID, and PPL.","Text-guided image editing follows as a special case in which the same image is used as both endpoints and different text prompts drive the transition.","Because the pre-trained diffusion model is used as-is, its generalization is preserved instead of being narrowed by fine-tuning."],"supporting_citations":[{"why":"The trained baseline it must beat and the source of MorphBench plus the LPIPS/FID/PPL evaluation protocol.","marker":"[49]"},{"why":"The other trained baseline, IMPUS, whose per-pair fine-tuning and semantic/layout limits motivate the tuning-free target.","marker":"[47]"},{"why":"Earlier CLIP-latent interpolation approach and source of the Class-A image pairs in Morph4Data.","marker":"[43]"},{"why":"The Stable Diffusion backbone whose self-attention modules FreeMorph modifies.","marker":"[39]"},{"why":"The DDIM inversion and denoising schedule used for both forward and reverse passes.","marker":"[38]"},{"why":"MasaCtrl's tuning-free mutual self-attention control, the template for replacing key/value features while keeping query features.","marker":"[5]"},{"why":"Plug-and-play feature injection that motivates using features of the input images to guide the denoiser.","marker":"[41]"},{"why":"Prompt-to-prompt attention control whose attention-modification idea underlies the designed attention replacements.","marker":"[13]"}],"fun_headline_variants":["Morph any images without training, just attention blending","Tuning-free morphing handles different objects and layouts","Spherical interpolation plus attention gives smooth morphs","Fastest morphing yet: 10-50x speedup and no training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that mixing the self-attention key and value features of the two inputs at particular hand-chosen points in the denoising schedule produces genuine intermediates that preserve both identities, an empirical assumption demonstrated only on the authors' evaluation pairs.","fun_headline_variants_meta":{"raw":{"variants":["Morph any images without training, just attention blending","Tuning-free morphing handles different objects and layouts","Spherical interpolation plus attention gives smooth morphs","Fastest morphing yet: 10-50x speedup and no training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000555,"raw_usage":{"total_tokens":2638,"prompt_tokens":937,"completion_tokens":1701,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":1633}},"tokens_in":553,"tokens_out":1701,"duration_ms":12017,"temperature":1.0,"reasoning_tokens":1633,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:38:53.291360+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: fix the hyperparameters $\\lambda_1$ through $\\lambda_4$ and run FreeMorph on pairs ranging from near-identical to semantically unrelated. If the optimal schedule drifts with pair type, or if re-running with different random high-frequency masks $m$ changes LPIPS or PPL by more than the reported gap to DiffMorpher, the tuning-free generalization claim is weaker than stated.","supporting_citations":[{"cited_title":"Diffmorpher: Unleashing the capability of diffusion models for image morphing","cited_arxiv_id":null,"evidence_quote":"The trained baseline it must beat and the source of MorphBench plus the LPIPS/FID/PPL evaluation protocol."},{"cited_title":"Impus: Image morphing with perceptually-uniform sampling using diffusion models","cited_arxiv_id":null,"evidence_quote":"The other trained baseline, IMPUS, whose per-pair fine-tuning and semantic/layout limits motivate the tuning-free target."},{"cited_title":"Stable diffusion","cited_arxiv_id":null,"evidence_quote":"The Stable Diffusion backbone whose self-attention modules FreeMorph modifies."},{"cited_title":"Denoising diffusion implicit models","cited_arxiv_id":null,"evidence_quote":"The DDIM inversion and denoising schedule used for both forward and reverse passes."},{"cited_title":"Masactrl: Tuning-free mu- tual self-attention control for consistent image synthesis and editing","cited_arxiv_id":null,"evidence_quote":"MasaCtrl's tuning-free mutual self-attention control, the template for replacing key/value features while keeping query features."},{"cited_title":"Plug-and-play diffusion features for text-driven image-to- image translation","cited_arxiv_id":null,"evidence_quote":"Plug-and-play feature injection that motivates using features of the input images to guide the denoiser."},{"cited_title":"Prompt-to-prompt image editing with cross-attention control","cited_arxiv_id":null,"evidence_quote":"Prompt-to-prompt attention control whose attention-modification idea underlies the designed attention replacements."}],"review_version":1}