Pith. sign in

REVIEW 3 major objections 7 minor 64 references

This paper claims that zero-shot, training-free image morphing can surpass tuning-based methods by adaptively re-injecting cached features from both input images and anchoring the text prompt to a shared semantic concept, and that the field

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

CHIMERA performs zero-shot diffusion image morphing with adaptive multi-scale feature-cache injection (ACI) and VLM-generated semantic anchor prompting (SAP), and proposes a new morphing-quality metric, GLCS.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Solid zero-shot morphing method with a genuine user study win, but the quantitative SOTA headline rests on a self-designed metric tuned in the same feature space the method controls — needs external validation before I'd trust the ranking. the 3 major comments →

arxiv 2512.07155 v6 pith:QNNF4XBU submitted 2025-12-08 cs.CV

CHIMERA: Adaptive Cache Injection and Semantic Anchor Prompting for Zero-shot Image Morphing with Morphing-oriented Metrics

Dahyeon Kye , Jeahun Sung , Minkyu Jeon , Jihyong Oh This is my paper
classification cs.CV
keywords image morphingdiffusion modelszero-shotDDIM inversionfeature reusesemantic anchor promptingevaluation metricGLCS
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that smooth, semantically coherent morphing between very different images does not require per-pair fine-tuning. It proposes CHIMERA, which caches multi-scale features from both inputs during DDIM inversion and re-injects them into the denoising U-Net with a depth- and timestep-aware schedule (ACI), while a vision-language model writes a shared 'anchor' prompt that is injected into cross-attention to keep semantics stable (SAP). On two benchmarks, CHIMERA scores highest on the paper's own morphing-oriented metric (GLCS), and a user study prefers its transitions on all four criteria—smoothness, domain consistency, perceptual quality, and overall quality. The paper also argues that conventional metrics are ill-suited to morphing and proposes GLCS as a quantitative measure that tracks human judgment.

Core claim

The central claim is that image morphing can be formulated as a cached inversion-guided denoising process where the two endpoints contribute not only interpolated latents but also their own multi-scale U-Net features, injected adaptively according to frequency content: low-frequency structural features during early denoising steps and high-frequency detail features later. Text conditioning, previously obtained per-endpoint and then blended, is instead replaced by a VLM-derived anchor prompt that captures the semantic intersection of the two inputs and is concatenated into cross-attention early in denoising. The paper shows that this zero-shot approach achieves the highest Global-Local Consis

What carries the argument

Adaptive Cache Injection (ACI) records down, mid, and up-block features of both inputs during DDIM inversion, blends them via slerp, and adds them as residual guidance using an Inversion-Denoiso Timestep Mapping (IDM) that aligns inversion and denoising steps. Semantic Anchor Prompting (SAP) uses a vision-language model to generate a shared anchor-prompt and two endpoint prompts, then augments cross-attention with the anchor's key-value embeddings. Global-Local Consistency Score (GLCS) is the paper's metric: GCS checks each frame's similarity to the endpoints against a slerp-interpolated expected trend; LCS checks neighbor-based smoothness; the two are combined as sqrt(GCS·LCS) using a DiffS

Load-bearing premise

The quantitative superiority claim rests on the assumption that GLCS—which expects each frame's similarity to the endpoints to follow a slerp-interpolated trend and its neighbor similarities to be smooth—faithfully captures what humans mean by a good morph; if that similarity-space model is wrong, the headline result collapses.

What would settle it

A simple test: collect a set of morphing sequences, ask human raters to rank them by overall quality, and compute GLCS for the same sequences. If any sequence that human raters clearly judge as excellent receives a distinctly lower GLCS than a visibly worse sequence, then GLCS—and the paper's quantitative claims based on it—fail. The paper's own user study is a step, but it covers only 15 sequences; a larger, independent study would settle the question.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If CHIMERA's claims hold, training-free morphing becomes competitive with per-pair fine-tuning, drastically reducing the compute and generalization barriers for morphing across new domains.
  • The ACI mechanism—caching and re-injecting multi-scale inversion features with a timestep mapping—could transfer to other diffusion-based editing or interpolation tasks where structural stability is needed.
  • SAP's use of a VLM-generated shared anchor prompt suggests a general recipe for improving semantic coherence whenever text conditioning is interpolated between two endpoints.
  • If GLCS is adopted by the community, evaluation of morphing will shift from distribution-level metrics to metrics that explicitly model the expected interpolation trajectory and local smoothness.
  • The result that CHIMERA wins no conventional metric but wins on GLCS and the user study implies that current benchmarks may reward the wrong properties, and future morphing research should be re-evaluated with perception-aligned metrics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's cross-backbone claim—'applicable across diverse diffusion backbones without retraining'—is asserted but only demonstrated on Stable Diffusion 2.1; extending ACI/SAP to other backbones is a testable next step, not yet shown.
  • If GLCS becomes standard, it may expose that earlier methods are better than their FID/LPIPS numbers suggest, or worse—depending on how closely DiffSim-based similarity tracks human perception for the specific kinds of structural and semantic changes morphing involves.
  • Applying CHIMERA to video frame interpolation, as the paper itself tries, reveals a fundamental mismatch: morphing generates plausible intermediate states without motion correspondence, so the method would need explicit motion priors to become a practical VFI model—an extension the paper identifies but does not solve.
  • A deeper implication left implicit is that the anchor-prompt's quality is entirely dependent on the VLM's ability to find a true shared concept; failures in that semantic inference would propagate through the entire morphing sequence, suggesting a need for uncertainty-aware anchor selection.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper proposes CHIMERA, a zero-shot diffusion-based image morphing framework with two main components: Adaptive Cache Injection (ACI), which caches multi-scale U-Net features during DDIM inversion and re-injects slerp-interpolated caches into denoising using a timestep mapping (IDM); and Semantic Anchor Prompting (SAP), which uses a VLM to generate a shared anchor prompt and endpoint prompts, augmenting early cross-attention to stabilize semantics. The authors also introduce a new morphing-oriented metric, GLCS, combining a Global Consistency Score (expected DiffSim similarities following a slerp trend) and a Local Consistency Score (neighbor-based smoothness), combined as the geometric mean. Experiments on Morph4Data and MorphBench compare CHIMERA with IMPUS, DiffMorpher, FreeMorph, and latent slerp. In Table 1, CHIMERA wins neither FID_local/FID_global (IMPUS wins) nor LPIPS/PPL (DiffMorpher wins), but achieves the highest GLCS on both datasets. A user study (32 participants, 15 sequences, randomized anonymization) shows CHIMERA preferred on all four criteria with significant Friedman tests. The paper claims a new state of the art in image morphing, with GLCS as the quantitative basis.

Significance. If the claims hold, the contributions are meaningful: a training-free feature-reuse mechanism with a timestep-aligned schedule, a VLM-driven semantic anchoring scheme, and a morphing-specific metric. The user study is genuine independent evidence and is a real strength of the paper. The computational cost is lower than tuning-based baselines, and the ablation coverage is reasonably broad. However, the headline quantitative claim rests on GLCS, a metric proposed and validated only within this paper, computed in a diffusion feature space related to the very features ACI manipulates, and used to choose the method's hyperparameters. The user study supports human preference for CHIMERA, but it does not quantitatively validate the GLCS ranking per sequence. Therefore the strongest claim — establishing a new state of the art — is not yet fully supported, though the underlying method appears promising.

major comments (3)
  1. [Sec. 4.3, Table 1, Table 3, Table 4]
  2. [Supp. G and Sec. 4.3 (Eqs. 24–31)]
  3. [Sec. 5.1]
minor comments (7)
  1. [Sec. 4.3]
  2. [Supp. Eq. (24)]
  3. [Sec. 4.1, Eq. (7)]
  4. [Supp. Algorithm 1 and Sec. 4.2]
  5. [Table 1 vs. Table 12]
  6. [Supp. Table 5]
  7. [Supp. Table 10]

Circularity Check

1 steps flagged

CHIMERA's GLCS-based SOTA is weakened by tuning the method on the same self-proposed metric, but the user study provides independent, if limited, support.

specific steps
  1. fitted input called prediction [Sec. 5.3.2 (ACI Ablation – layer weight, Table 3) and Sec. 5.1/Table 1]
    "Although the LPIPS and PPL values are relatively higher, we choose λ_S = 0.4 as the final weight because GLCS offers a more reliable assessment of smoothness."

    The paper's headline quantitative claim is SOTA GLCS in Table 1, but the final model weight λ_S was explicitly selected to maximize the authors' own GLCS in Table 3 (0.4 is the GLCS peak among the tested values). Thus the reported GLCS advantage is not an independent test of the method: a key hyperparameter was tuned on the same metric that later serves as the primary evidence of state-of-the-art performance. This is a partial fit-to-evaluator loop rather than a full by-construction reduction, and the user study provides some external validation.

full rationale

The strongest claim — 'new state-of-the-art in image morphing' — rests mainly on GLCS from Table 1, where CHIMERA does not win FID_local, FID_global, LPIPS, or PPL. GLCS is proposed in this same paper and is computed from a DiffSim-based diffusion-feature similarity (Eq. 19), while CHIMERA's ACI manipulates diffusion U-Net features (Eqs. 6–7). This creates a real risk that the metric rewards exactly the kind of diffusion-feature slerp behavior CHIMERA was designed to produce, though the equations do not make the two identical: GLCS is a similarity-space expectation, not a direct readout of the injected cache. The more concrete circular step is the explicit selection of λ_S=0.4 because it maximizes GLCS, followed by using GLCS as the headline comparison metric. That is a legitimate concern under the 'fitted input called prediction' pattern, since the same benchmark and metric are used for both tuning and final evaluation. However, the paper does include a user study (Tables 12–14) in which CHIMERA is preferred over all baselines on all four criteria, with significance tests; this is independent evidence that the result is not purely an artifact of the self-designed metric. The GLCS–user-study alignment is only shown at the method-aggregate level and the study is small (15 sequences, 32 participants), so it does not fully validate GLCS per sequence. No load-bearing self-citations or uniqueness theorems appear in the paper; the citations to the authors' own prior works are peripheral. Overall, the central method itself is not circular by construction, but the quantitative SOTA claim is partly weakened by tuning and evaluating on the same in-house metric, giving a moderate score of 4 rather than a higher one.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 2 invented entities

The central method is a composition of existing building blocks (DDIM inversion caches, cross-attention control, VLM captioning) plus two paper-specific constructs: the IDM-timed ACI schedule and the GLCS metric. The major loaded assumptions are the frequency-layer correspondence used to justify ACI and the DiffSim/slerp model underlying GLCS. The key free parameters (lambda_S, SAP stage, gamma, alpha schedule) are selected hand-in-hand with the evaluation metric, which is the main epistemic weakness. No new physical entities are introduced; the anchor-prompt is an algorithmic device with in-paper ablative evidence.

free parameters (5)
  • ACI injection weight lambda_S = 0.4
    Sec. 5.3.2 / Table 3: tuned on GLCS and FID; LPIPS and PPL are better at lambda=0.7, but 0.4 is kept because GLCS is deemed more reliable. The choice is entangled with the evaluation metric.
  • GLCS sharpening exponent gamma = unspecified
    Eq. 26 (Supplement): gamma>=1 sharpens GCS sensitivity; no value is reported anywhere, so all GLCS numbers depend on an undisclosed knob.
  • SAP activation stage = early stage only
    Sec. 5.3.3 / Table 4: empirically chosen; stage1-only is adopted because stage2 or both degrade FID/GLCS.
  • GLCS expected-trend interpolation weights alpha_k = (k+1)/(K+1)
    Eq. 20 (and inconsistently, Algorithm 2): the 'ideal mixing ratio' used in GCS is a modeling choice about how similarity to endpoints should progress; different schedules would change the score.
  • Sampling configuration = N_inv=N_dng=50, guidance 0.75, 768x768
    Sec. 5 / Implementation Detail: inherited from FreeMorph [7]; results depend on these choices.
axioms (6)
  • domain assumption DDIM inversion of SD2.1 produces latents whose slerp interpolation traces a perceptually meaningful morphing path
    Sec. 4, Eqs. 2: the entire pipeline assumes the morphing can be seeded by interpolating inverted latents of the two endpoints.
  • domain assumption U-Net mid-block features are low-frequency and up-block features are high-frequency; early denoising timesteps are low-frequency and later ones high-frequency
    Sec. 3, 'Observation' and Fig. 3: measured only on the base model over Morph4Data; the frequency-matching injection schedule of ACI depends on this.
  • domain assumption Linear IDM mapping phi(t) correctly aligns cached inversion timesteps with denoising timesteps
    Eq. 5: a simple linear/rounding map is asserted to match features of similar frequency characteristics; ablations (Tables 6-7) support but do not prove it.
  • domain assumption CLIP text embedding space is approximately locally linear, so correlated captions and anchor prompts blend stably
    Sec. 4.2, cited to [5, 32, 52]; underpins the SAP interpolation reasoning.
  • domain assumption Qwen2.5-VL reliably extracts a shared semantic/layout anchor from an arbitrary endpoint pair with a fixed prompt template
    Sec. 4.2 and Supplement F: no failure analysis of the VLM output beyond CLIP similarity of anchor to captions (Table 10).
  • ad hoc to paper DiffSim-based bounded similarity s(X,Y) in [-1,1], with per-frame similarity trends modeled by slerp of four endpoint similarities, is a valid model of perceptual morphing progression
    Supplement E, Eqs. 19-31: the entire GLCS construct rests on this; it is validated only against the paper's own user study.
invented entities (2)
  • Shared semantic anchor-prompt (text_anc) independent evidence
    purpose: Injected into cross-attention key/value pairs to stabilize semantics and prevent drift in heterogeneous morphing pairs
    Biased-anchor ablations (Anchor=A, Anchor=B, Anchor=Irrelevant, Fig. 20, Table 11) show degradation, and CLIP cosine similarity with both captions is symmetric (~0.91, Table 10), giving a falsifiable handle within the paper.
  • GLCS metric (GCS x LCS) no independent evidence
    purpose: Proposed morphing-oriented quantitative evaluation score
    GLCS is validated only inside this paper, against its own user study and its own qualitative examples; no external benchmark or independent evaluator confirms it. It is falsifiable in principle but has no evidence outside the paper.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CHIMERA: Adaptive Cache Injection and Semantic Anchor Prompting for Zero-shot Image Morphing with Morphing-oriented Metrics." pith.science (2026). https://pith.science/paper/QNNF4XBU

@misc{pith2026251207155,
  author       = {Pith},
  title        = {Pith review of: CHIMERA: Adaptive Cache Injection and Semantic Anchor Prompting for Zero-shot Image Morphing with Morphing-oriented Metrics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QNNF4XBU}},
  note         = {Machine review of arXiv:2512.07155}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent diffusion-based image morphing methods typically interpolate inverted latents and reuse limited conditioning signals, which often yields unstable intermediates for heterogeneous endpoint pairs. In particular, (i) feature reuse is usually partial or non-adaptive, leading to abrupt structural changes or over-smoothing, and (ii) text conditions are commonly obtained independently per endpoint and then interpolated, which can introduce incompatible semantics. We present CHIMERA, a novel zero-shot diffusion morphing framework that addresses both issues via inversion-guided denoising with complementary feature reuse and text conditioning. Adaptive Cache Injection (ACI) caches a broader set of multi-scale diffusion features beyond Key-Value-only reuse during DDIM inversion, and re-injects them with layer- and timestep-aware scheduling to stabilize denoising and enable gradual fusion. Semantic Anchor Prompting (SAP) uses a VLM to generate a shared anchor-prompt and anchor-conditioned endpoint prompts, and injects the anchor into cross-attention to improve intermediate semantic coherence. Finally, we propose Global-Local Consistency Score (GLCS), a morphing-oriented metric that jointly captures global domain harmonization and local transition smoothness. Extensive experiments and a user study show that CHIMERA produces smoother and more semantically consistent morphing results than prior methods, while remaining efficient and applicable across diverse diffusion backbones without retraining.

Figures

Figures reproduced from arXiv: 2512.07155 by Dahyeon Kye, Jeahun Sung, Jihyong Oh, Minkyu Jeon.

Figure 1
Figure 1. Figure 1: Key challenges in morphing and a user study with our morphing-oriented metric (GLCS). Existing methods struggle with smoothness, domain consistency, and perceptual quality (red arrows), while our approach (CHIMERA) produces coherent transitions across all three. Standard metrics (FID, LPIPS [57], PPL [22]) fail to reflect true morphing quality, whereas user study results on two datasets [7, 55] align close… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative result of smoothness of morphing transi￾tion (Smooth), heterogeneous-aware domain consistency (Do￾main Consistency), and perceptual quality (Perceptual Qual￾ity). Here, ✗ indicates cases that fail for most pairs, ▲ represents cases that fail for some pairs, and ✓ denotes cases that succeed for most pairs. To overcome these limitations, we propose CHIMERA (Adaptive CacHe Injection and SeMantic A… view at source ↗
Figure 3
Figure 3. Figure 3: Frequency analysis of diffusion features and denois￾ing timesteps. Low- (blue) and high-frequency (orange) compo￾nents across (a) U-Net feature layers and (b) DDIM denoising timesteps are measured for the base model without CHIMERA’s ACI and SAP on Morph4Data [7]. Values are obtained by apply￾ing FFT with masked frequency bands and averaging the resulting magnitudes. 2.2. Diffusion Latents and Feature Reus… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the CHIMERA framework. (a) DDIM Inversion: Inputs A and B are inverted while caching multi–scale U-Net features from the down, mid, and up blocks. The cached features are interpolated via slerp, forming morphing-aligned latents. (b) Denoising: The interpolated caches are re-injected through ACI, which aligns inversion and denoising timesteps via the proposed IDM. ACI injects mid-block features … view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results of the proposed GLCS. Given in￾put image pairs (1) and (3), different methods produce morphing sequences shown in (2), which highlight cases where GLCS suc￾cessfully reflects differences in global–local consistency that are not fully captured by conventional metrics. pairs (KA, VA), (KB, VB), and (Kanc, Vanc). The anchor projection is concatenated with each endpoint branch as fol￾lows: … view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparisons with existing SOTA meth￾ods. (1)–(2) denote the input image pairs. (a)–(d) show qualitative results for each model on the Morph4Data dataset. 5.2. Qualitative Evaluations To demonstrate the effectiveness of our proposed CHIMERA, we provide a qualitative comparison with existing methods in [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative results based on the types of features cached in ACI. (i) and (ii) represent the input image pair, while D, M, and U denote the down, mid, and up features, respectively. 5.3. Ablation Studies 5.3.1. Caching Feature Type on ACI When only the down or down–mid features are provided, as shown in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison showing the results of generating five morphing images. Panels (1)–(4) denote the input images, and panels (a)–(d) correspond to IMPUS, DiffMorpher, FreeMorph, and CHIMERA (Ours), respectively. The same convention applies to panels (e)–(h). (b) DiffMorpher (c) FreeMorph (d) CHIMERA (a) IMPUS (1) Input A (2) Input B (f) DiffMorpher (g) FreeMorph (h) CHIMERA (e) IMPUS (3) Input A (4) I… view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparison showing the results of generating five morphing images. Panels (1)–(4) denote the input images, and panels (a)–(d) correspond to IMPUS, DiffMorpher, FreeMorph, and CHIMERA (Ours), respectively. The same convention applies to panels (e)–(h). 2 [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison showing the results of challenging 14-image morphing (compared to 5-image morphing). Panels (1)–(4) denote the input images, and panels (a)–(d) correspond to IMPUS, DiffMorpher, FreeMorph, and CHIMERA (Ours), respectively. The same convention applies to panels (e)–(h). Please zoom in for better visualization. 3 [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative comparison showing the results of challenging 14-image morphing (compared to 5-image morphing). Panels (1)–(4) denote the input images, and panels (a)–(d) correspond to IMPUS, DiffMorpher, FreeMorph, and CHIMERA (Ours), respectively. The same convention applies to panels (e)–(h). Please zoom in for better visualization. 4 [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative results for different injection weights of the cached ACI features in the denoising process. (i) and (ii) denote the input image pair, and (a)–(d) show the results for λS values of 0.1, 0.4, 0.7, and 1.0, respectively. D. Evaluation Metric This section provides detailed explanations of the metrics introduced in [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative results when the inversion timesteps are fixed. Panels (b) Early, (c) Mid, and (d) Late correspond to states with high noise, medium noise, and no noise, respectively. Panel (a) represents our model with the IDM applied. (b) Early (c) Mid (d) Late (a) Ours (1) Input A (2) Input B [PITH_FULL_IMAGE:figures/full_fig_p017_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative results when the denoising timesteps are fixed. Panels (b) Early, (c) Mid, and (d) Late correspond to states with high noise, medium noise, and no noise, respectively. Panel (a) represents our model with the IDM applied. D.1. Frechet Inception Distance (FID)-Based Met- ´ rics Local FID. We use a local variant, FIDlocal, to measure distribution gaps between the input image pair {A, B} and the m… view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative examples showing how the GCS component of GLCS aligns with human perception. Blue arrows indicate frames where the domains of A and B are properly mixed according to the perceived interpolation ratio, while red arrows indicate frames where the two domain cues are not well reflected given the same interpolation ratio. 9 [PITH_FULL_IMAGE:figures/full_fig_p020_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative examples showing how the LCS component of GLCS aligns with human perception. Blue arrows indicate cases that are judged as similar by human observers, while red arrows indicate cases with abrupt perceptual changes. 10 [PITH_FULL_IMAGE:figures/full_fig_p021_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Qualitative comparisons between FIDlocal and GCS, which is a component of our proposed metric. Panels (a) and (b) present qualitative results for two different cases. LPIPS↓: 1.9848 / PPL↓: 0.3308 / LCS↑: 82.0715 LPIPS↓: 2.1391 / PPL↓: 0.3565 / LCS↑: 85.5664 LPIPS↓: 1.9956 / PPL↓: 0.3326 / LCS↑: 83.8628 LPIPS↓: 2.1526 / PPL↓: 0.3588 / LCS↑: 90.8176 (a) LPIPS & PPL vs. LCS case 1 (b) LPIPS & PPL vs. LCS ca… view at source ↗
Figure 18
Figure 18. Figure 18: Qualitative comparisons between LPIPS, PPL, and LCS, which is a component of our proposed metric. Panels (a) and (b) present qualitative results for two different cases. 11 [PITH_FULL_IMAGE:figures/full_fig_p022_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Examples of VLM-generated captions and anchor prompts. Given two endpoint images, VLM produces per-image captions (textA, textB) and a shared anchor-prompt (textanc), which is used by SAP to enforce semantic alignment during the denoising process. mized for morphing rather than mere description. To disentangle the effect of this morphing-oriented tex￾tual interface from architectural changes, we conduct a… view at source ↗
Figure 20
Figure 20. Figure 20: illustrates distinct failure modes for each variant. When Anchor=A, the transition is heavily skewed toward the source, with attributes specific to Input A (e.g., tousled hair) persisting unnaturally into later images. In contrast, when Anchor=B, target-specific attributes (e.g., black ar￾mor and a red-glowing eye) appear too early, causing the facial skin to take on a plastic, armor-like appearance pre￾m… view at source ↗
Figure 21
Figure 21. Figure 21: User study interface and questionnaire form. this shared formulation through the VLM prompting strat￾egy, SAP receives a balanced textual anchor that maintains semantic symmetry over the sequence, directly contributing to the GLCS and fidelity gains reported in [PITH_FULL_IMAGE:figures/full_fig_p026_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: User study win–tie–loss ratios of CHIMERA against each baseline Method step1inv [s] step2denoise [s] Total [s] Params (B) IMPUS 32.92 18.44 478.91 1.93 DiffMorpher 3.57 60.13 64.92 1.30 FreeMorph 20.42 8.79 30.66 1.29 CHIMERA (Ours) 4.90 9.59 14.49 1.29 [PITH_FULL_IMAGE:figures/full_fig_p028_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Qualitative VFI results on Vimeo90K-septuplet. Panels (a)–(d) correspond to IMPUS, DiffMorpher, FreeMorph, and CHIMERA (Ours), respectively. For each sequence, red arrows mark representative artifacts such as unrealistic limb configurations or duplicated local structures in the interpolated frames. (1) Input A GT (2) Input B (a) IMPUS (b) DiffMorpher (c) FreeMorph (d) CHIMERA (1) Input A GT (2) Input B (a… view at source ↗
Figure 24
Figure 24. Figure 24: Qualitative VFI results on DAVIS. Panels (a)–(d) correspond to IMPUS, DiffMorpher, FreeMorph, and CHIMERA (Ours), respectively. The red arrows highlight severe failure cases where the interpolated results exhibit non-physical human bodies, including truncated or distorted arms and legs. tions between disparate visual concepts. Given two images that serve as keyframes, the framework generates a tempo￾rally… view at source ↗
Figure 25
Figure 25. Figure 25: Failure cases on images with prominent text. When the endpoint images contain different words or textual layouts, all compared methods, including CHIMERA, often produce broken or unreadable characters and occasional abrupt changes in the rendered text. anism itself, but from the inherent inductive biases of the underlying pre-trained diffusion backbones [9, 37]. Stan￾dard text-to-image models are known to… view at source ↗
Figure 26
Figure 26. Figure 26: Additional qualitative results for 5-frame morphing. (a) IMPUS (b) DiffMorpher (c) FreeMorph (d) CHIMERA (1) Image A (2) Image B (3) Image A (4) Image B (e) IMPUS (f) DiffMorpher (g) FreeMorph (h) CHIMERA Generated Morphing Images Generated Morphing Images [PITH_FULL_IMAGE:figures/full_fig_p031_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Additional qualitative results for 5-frame morphing. 20 [PITH_FULL_IMAGE:figures/full_fig_p031_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: Additional qualitative results for 14-frame morphing. 21 [PITH_FULL_IMAGE:figures/full_fig_p032_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: Additional qualitative results for 14-frame morphing. 22 [PITH_FULL_IMAGE:figures/full_fig_p033_29.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

64 extracted references · 8 linked inside Pith

  1. [1]

    Image morphing techniques: A re- view.Technium, 9, 2023

    Alyaa Qusay Aloraibi. Image morphing techniques: A re- view.Technium, 9, 2023. 3

    2023

  2. [2]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923, 2025. 2, 3, 5, 13, 15

    Pith/arXiv arXiv 2025

  3. [3]

    Feature-based image metamorphosis.ACM SIGGRAPH Computer Graphics, 26 (2):35–42, 1992

    Thaddeus Beier and Shawn Neely. Feature-based image metamorphosis.ACM SIGGRAPH Computer Graphics, 26 (2):35–42, 1992. 3

    1992

  4. [4]

    Comparative study of triangulation based and feature based image morphing.Signal & Image Process- ing, 2(4):235, 2011

    Bhumika G Bhatt. Comparative study of triangulation based and feature based image morphing.Signal & Image Process- ing, 2(4):235, 2011. 2

    2011

  5. [5]

    Sega: Instructing text-to-image models using semantic guidance

    Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems, 36: 25365–25389, 2023. 5

    2023

  6. [6]

    High accuracy optical flow estimation based on a theory for warping

    Thomas Brox, Andres Bruhn, Nils Papenberg, and Joachim Weickert. High accuracy optical flow estimation based on a theory for warping. InEuropean Conference on Computer Vision (ECCV), pages 25–36, 2004. 3

    2004

  7. [7]

    Freemorph: Tuning-free generalized im- age morphing with diffusion models.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    Jiayi Cao, Xinyu Lin, Yichen Xu, Jing Xu, Zhen Zhang, and Zongwei Li. Freemorph: Tuning-free generalized im- age morphing with diffusion models.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 3, 5, 6, 7, 8, 13, 14, 15, 17, 18

    2025

  8. [8]

    Attention interpolation for text-to-image diffusion models.Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023

    Hila Chefer, Ron Mokady, Or Lang, Yuval Alaluf, Gal Chechik, and Daniel Cohen-Or. Attention interpolation for text-to-image diffusion models.Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3

    2023

  9. [9]

    Textdiffuser: Diffusion models as text painters.Advances in Neural Information Processing Sys- tems, 36:9353–9387, 2023

    Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters.Advances in Neural Information Processing Sys- tems, 36:9353–9387, 2023. 19

    2023

  10. [10]

    Repurposing pre-trained video diffusion models for event-based video interpolation

    Jingxi Chen, Brandon Y Feng, Haoming Cai, Tianfu Wang, Levi Burner, Dehao Yuan, Cornelia Fermuller, Christopher A Metzler, and Yiannis Aloimonos. Repurposing pre-trained video diffusion models for event-based video interpolation. InProceedings of the Computer Vision and Pattern Recogni- tion Conference, pages 12456–12466, 2025. 17

    2025

  11. [11]

    Improv- ing diffusion models with self-conditioning.Proceedings of the International Conference on Machine Learning (ICML),

    Ting Chen, Ruixiang Zhang, and Martin Arjovsky. Improv- ing diffusion models with self-conditioning.Proceedings of the International Conference on Machine Learning (ICML),

  12. [12]

    Multiple video frame interpolation via enhanced deformable separable con- volution.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 44(10):7029–7045, 2021

    Xianhang Cheng and Zhenzhong Chen. Multiple video frame interpolation via enhanced deformable separable con- volution.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 44(10):7029–7045, 2021. 17

    2021

  13. [13]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. InAdvances in Neural Information Processing Systems (NeurIPS), 2021. 3

    2021

  14. [14]

    Wave: Warping ddim inversion features for zero-shot text-to-video editing

    Yutang Feng, Sicheng Gao, Yuxiang Bao, Xiaodi Wang, Shumin Han, Juan Zhang, Baochang Zhang, and Angela Yao. Wave: Warping ddim inversion features for zero-shot text-to-video editing. InEuropean Conference on Computer Vision, pages 38–55. Springer, 2024. 2

    2024

  15. [15]

    Omnitext: A training-free generalist for controllable text-image manipulation.arXiv preprint arXiv:2510.24093, 2025

    Agus Gunawan, Samuel Teodoro, Yun Chen, Soo Ye Kim, Jihyong Oh, and Munchurl Kim. Omnitext: A training-free generalist for controllable text-image manipulation.arXiv preprint arXiv:2510.24093, 2025. 19

    arXiv 2025

  16. [16]

    Tenenbaum, An- tonio Torralba, and Ariel Shamir

    Amir Hertz, Ron Mokady, Jonathan B. Tenenbaum, An- tonio Torralba, and Ariel Shamir. Prompt-to-prompt im- age editing with cross-attention control.arXiv preprint arXiv:2208.01626, 2022. 3

    Pith/arXiv arXiv 2022

  17. [17]

    Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter

    M. Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time- scale update rule converge to a local nash equilibrium. In Neural Information Processing Systems, 2017. 6, 7

    2017

  18. [18]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. InAdvances in Neural Information Processing Systems (NeurIPS), 2022. 3

    2022

  19. [19]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. InAdvances in Neural Informa- tion Processing Systems (NeurIPS), 2020. 3

    2020

  20. [20]

    On exact inversion of dpm-solvers

    Seongmin Hong, Kyeonghyun Lee, Suh Yoon Jeon, Hyewon Bae, and Se Young Chun. On exact inversion of dpm-solvers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7069–7078, 2024. 2

    2024

  21. [21]

    Horn and Brian G

    Berthold K.P. Horn and Brian G. Schunck. Determining op- tical flow.Artificial Intelligence, 17(1-3):185–203, 1981. 3

    1981

  22. [22]

    Laine, M

    Tero Karras, S. Laine, M. Aittala, Janne Hellsten, J. Lehti- nen, and Timo Aila. Analyzing and improving the image quality of stylegan. InComputer Vision and Pattern Recog- nition, 2019. 1, 6, 7

    2019

  23. [23]

    Probability density geodesics in image diffusion latent space.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    Jonghyun Kim, Jaehyung Park, Seungho Yang, and Dongy- oon Han. Probability density geodesics in image diffusion latent space.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3

    2024

  24. [24]

    Acevfi: A comprehensive survey of advances in video frame interpolation.arXiv preprint arXiv:2506.01061, 2025

    Dahyeon Kye, Changhyun Roh, Sukhun Ko, Chanho Eom, and Jihyong Oh. Acevfi: A comprehensive survey of advances in video frame interpolation.arXiv preprint arXiv:2506.01061, 2025. 17

    arXiv 2025

  25. [25]

    Flow-based image morph- ing.IEEE Transactions on Image Processing, 21(2):820– 833, 2012

    Sangho Lee and Hanseok Kim. Flow-based image morph- ing.IEEE Transactions on Image Processing, 21(2):820– 833, 2012. 3

    2012

  26. [26]

    Layer control: Revisiting layer-wise feature modulation for diffusion models.arXiv preprint arXiv:2404.12217, 2024

    Wenqi Li, Jingyu Zhao, Yifan Zhang, and Liwei Wang. Layer control: Revisiting layer-wise feature modulation for diffusion models.arXiv preprint arXiv:2404.12217, 2024. 3

    Pith/arXiv arXiv 2024

  27. [27]

    Sparse global matching for video frame interpolation with large motion

    Chunxu Liu, Guozhen Zhang, Rui Zhao, and Limin Wang. Sparse global matching for video frame interpolation with large motion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19125– 19134, 2024. 17

    2024

  28. [28]

    Neural image morphing for cross-domain transitions

    Haoran Liu, Xiaolong Wang, and Lei Zhang. Neural image morphing for cross-domain transitions. InEuropean Con- ference on Computer Vision (ECCV), pages 401–418, 2022. 3

    2022

  29. [29]

    Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023. 3, 13 9

    2023

  30. [30]

    Null-text inversion for editing real im- ages using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real im- ages using guided diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6038–6047, 2023. 2, 3, 13

    2023

  31. [31]

    Softmax splatting for video frame interpolation

    Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 5437–5446, 2020. 17

    2020

  32. [32]

    Styleclip: Text-driven manipulation of stylegan imagery

    Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 2085–2094,

    2085

  33. [33]

    The 2017 davis challenge on video object segmentation.arXiv preprint arXiv:1704.00675, 2017

    Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Ar- bel´aez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation.arXiv preprint arXiv:1704.00675, 2017. 17

    Pith/arXiv arXiv 2017

  34. [34]

    Barron, and Ben Milden- hall

    Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion. InAd- vances in Neural Information Processing Systems (NeurIPS),

  35. [35]

    Learning transferable visual models from natural language supervision.Proceedings of the International Conference on Machine Learning (ICML),

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision.Proceedings of the International Conference on Machine Learning (ICML),

  36. [36]

    Riemannian morph- ing on manifolds

    Aleksa Rajkovi ´c and Laurent Younes. Riemannian morph- ing on manifolds. InInternational Conference on Computer Vision (ICCV), pages 1–11, 2023. 3

    2023

  37. [37]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2, 4, 5, 6, 19

    2022

  38. [38]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. InInternational Conference on Medical image com- puting and computer-assisted intervention, pages 234–241. Springer, 2015. 2, 3

    2015

  39. [39]

    Berg, and Li Fei-Fei

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, San- jeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Chal- lenge.International Journal of Computer Vision (IJCV), 115 (3):211–252, 2015. 6

    2015

  40. [40]

    Bim- vfi: Bidirectional motion field-guided frame interpolation for video with non-uniform motions

    Wonyong Seo, Jihyong Oh, and Munchurl Kim. Bim- vfi: Bidirectional motion field-guided frame interpolation for video with non-uniform motions. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 7244–7253, 2025. 17

    2025

  41. [41]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 2, 3, 4

    Pith/arXiv arXiv 2010

  42. [42]

    Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions.Proceedings of the International Conference on Learn- ing Representations (ICLR), 2021. 3

    2021

  43. [43]

    Diffsim: Taming diffusion models for evaluating visual similarity

    Yiren Song, Xiaokang Liu, and Mike Zheng Shou. Diffsim: Taming diffusion models for evaluating visual similarity. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16904–16915, 2025. 6, 7, 8

    2025

  44. [44]

    T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models

    Shengding Tang, Wei Wu, Yixin Zhang, Yichong Jiang, Xingjun Li, Chen Lin, Jiajun Wang, Shujian Huang, Kun Zhou, Dahua Lin, and Ping Luo. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR),

  45. [45]

    Diffusion-based image interpolation via denoising trajectory alignment

    Tian Wang, Polina Golland, and Joshua Tenenbaum. Diffusion-based image interpolation via denoising trajectory alignment. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. 2

    2024

  46. [46]

    Generative inbetweening: Adapting image- to-video models for keyframe interpolation.arXiv preprint arXiv:2408.15239, 2024

    Xiaojuan Wang, Boyang Zhou, Brian Curless, Ira Kemelmacher-Shlizerman, Aleksander Holynski, and Steven M Seitz. Generative inbetweening: Adapting image- to-video models for keyframe interpolation.arXiv preprint arXiv:2408.15239, 2024. 17

    Pith/arXiv arXiv 2024

  47. [47]

    Interpolat- ing between images with diffusion models.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    Yuxin Wang, Wenguan Wang, and Yi Yang. Interpolat- ing between images with diffusion models.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3

    2023

  48. [48]

    Image morphing: A survey.The Visual Computer, 14(8-9):360–372, 1998

    George Wolberg. Image morphing: A survey.The Visual Computer, 14(8-9):360–372, 1998. 3

    1998

  49. [49]

    Perception-oriented video frame interpolation via asymmetric blending

    Guangyang Wu, Xin Tao, Changlin Li, Wenyi Wang, Xiao- hong Liu, and Qingqing Zheng. Perception-oriented video frame interpolation via asymmetric blending. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2753–2762, 2024. 17

    2024

  50. [50]

    Inversion-free image editing with natural language

    Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. arXiv preprint arXiv:2312.04965, 2023. 5

    Pith/arXiv arXiv 2023

  51. [51]

    Video enhancement with task-oriented flow.International Journal of Computer Vision, 127(8): 1106–1125, 2019

    Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow.International Journal of Computer Vision, 127(8): 1106–1125, 2019. 17

    2019

  52. [52]

    Impus: Image morphing with perceptually- uniform sampling using diffusion models.arXiv preprint arXiv:2311.06792, 2023

    Zhaoyuan Yang, Zhengyang Yu, Zhiwei Xu, Jaskirat Singh, Jing Zhang, Dylan Campbell, Peter Tu, and Richard Hartley. Impus: Image morphing with perceptually- uniform sampling using diffusion models.arXiv preprint arXiv:2311.06792, 2023. 2, 3, 5, 7, 13, 15, 17, 18

    Pith/arXiv arXiv 2023

  53. [53]

    Textctrl: Diffusion-based scene text editing with prior guidance control.Advances in Neural Information Pro- cessing Systems, 37:138569–138594, 2024

    Weichao Zeng, Yan Shu, Zhenhang Li, Dongbao Yang, and Yu Zhou. Textctrl: Diffusion-based scene text editing with prior guidance control.Advances in Neural Information Pro- cessing Systems, 37:138569–138594, 2024. 19

    2024

  54. [54]

    Free-lunch color-texture disentanglement for stylized image generation.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    Hanting Zhang, Yifan Deng, Yixiao Zhang, Xinyu Chen, Zhaoxi Li, Kaili Chen, Yifan Li, Pan Lu, Ping Luo, and Dengxin Dai. Free-lunch color-texture disentanglement for stylized image generation.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 10

    2023

  55. [55]

    Diffmorpher: Unleashing the capability of dif- fusion models for image morphing

    Kaiwen Zhang, Yifan Zhou, Xudong Xu, Bo Dai, and Xin- gang Pan. Diffmorpher: Unleashing the capability of dif- fusion models for image morphing. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7912–7921, 2024. 1, 2, 3, 5, 6, 7, 13, 15, 17, 18

    2024

  56. [56]

    Adding conditional control to text-to-image diffusion mod- els

    Lvmin Zhang, Maneesh Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion mod- els. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023. 3

    2023

  57. [57]

    Efros, Eli Shecht- man, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition,

  58. [58]

    Eden: Enhanced diffusion for high-quality large-motion video frame interpo- lation

    Zihao Zhang, Haoran Chen, Haoyu Zhao, Guansong Lu, Yanwei Fu, Hang Xu, and Zuxuan Wu. Eden: Enhanced diffusion for high-quality large-motion video frame interpo- lation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 2105–2115, 2025. 17

    2025

  59. [59]

    Layout-agnostic scene text image synthesis with diffusion models

    Qilong Zhangli, Jindong Jiang, Di Liu, Licheng Yu, Xi- aoliang Dai, Ankit Ramchandani, Guan Pang, Dimitris N Metaxas, and Praveen Krishnan. Layout-agnostic scene text image synthesis with diffusion models. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7496–7506. IEEE Computer Society, 2024. 19

    2024

  60. [60]

    A survey of morphing techniques.International Journal of Advanced Engineering, Management and Science, 3(2):239773, 2017

    Bhushan Zope and Soniya B Zope. A survey of morphing techniques.International Journal of Advanced Engineering, Management and Science, 3(2):239773, 2017. 3 11 CHIMERA: Adaptive Cache Injection and Semantic Anchor Prompting for Zero-shot Image Morphing with Morphing-oriented Metrics Supplementary Material A. Extended Experiment Results This section provide...

    arXiv 2017

  61. [61]

    First, carefully identify their shared semantic concept, the main subject, action, or event that connects both images

  62. [62]

    Next, identify their shared layout structure, the spatial arrange- ment or composition of major elements (e.g., background, perspective, subject position) that remains partially consistent between both

  63. [63]

    Summarize the shared theme (semantic and/or layout) in one short compact phrase

  64. [64]

    beautiful

    Then, write short but precise captions for each image, ensuring that both captions naturally include the shared semantic meaning and layout structure. Use this exact format strictly: Anchor-prompt:[compact phrase capturing shared seman- tic or layout aspect] Caption A:[short factual description of image1 including the shared theme] Caption B:[short factua...

    1901

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.