CHIMERA: Adaptive CacHe Injection and SeMantic Anchor Prompting for ZERo-shot ImAge Morphing with Morphing-oriented Metrics

Dahyeon Kye ^*,1

Jeahun Sung ^*,1

Minkyu Jeon ²

Jihyong Oh ^†,1

arXiv 2026

^*equal contribution, ^†corresponding author, ¹CMLab, Chung-ang University, ²Princeton University

Creative Vision and Multimedia Lab (CMLab)

Chung-Ang University

Princeton University

{rpekgus, jhseong, jihyongoh}@cau.ac.kr
mj7341@princeton.edu

Code arXiv

TL;DR

CHIMERA enables smooth and semantically consistent zero-shot image morphing through Adaptive Cache Injection (ACI) and Semantic Anchor Prompting (SAP), along with GLCS, a new morphing-oriented metric for evaluating transition quality.

Results of CHIMERA

Figure 1. Qualitative Result of Our Method

Abstract

Recent diffusion-based image morphing methods typically interpolate inverted latents and reuse limited conditioning signals, often yielding 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 the 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.

Overview

Motivation & Observation

Observation on diffusion feature reuse across KV, QKV, and ResNet features — Figure 3. Observation on diffusion feature reuse. (a) reuses only attention-layer Key/Value features (KV), (b) reuses attention features including Query (Attn; QKV), and (c) further includes ResNet features (Attn, Res). (i) shows qualitative results, (ii) reports domain consistency, and (iii) reports smoothness. The results in (ii) and (iii) are averaged over the Morph4Data dataset.

Frequency analysis showing LF bias — Figure 4. Frequency analysis of each feature in the diffusion U-Net and across different denoising timesteps.

Frequency analysis of the diffusion U-Net and the denoising timesteps

Diffusion features tend to contain more low-frequency information in the mid layers and more high-frequency information in the up layers. In addition, early denoising timesteps mainly encode low-frequency information, while late timesteps contain more high-frequency information. Based on these properties, ACI injects diffusion features that match the characteristics of each denoising timestep.

Proposed Method

Diagram of the transformer deep learning architecture. — Figure 5. ACI corrects the timestep mismatch between inversion and denoising via the proposed IDM and reinjects multi-scale cached features (low-frequency structures early and high-frequency details later) to guide consistent morphing. SAP introduces a VLM-derived anchor prompt into early cross-attention layers, stabilizing semantics and reducing drift for heterogeneous input pairs.

Algorithm for the CHIMERA method (ACI and SAP) — Algorithm 1. Algorithm for the full CHIMERA pipeline, combining Adaptive Cache Injection (ACI) and Semantic Anchor Prompting (SAP).

Proposed Metric

Effect of GCS — Figure 6. Qualitative examples demonstrating the effectiveness of GLCS. GLCS consists of GCS and LCS, and the qualitative results illustrate how well each component aligns with human perception.

Effect of LCS — Figure 6. Qualitative examples demonstrating the effectiveness of GLCS. GLCS consists of GCS and LCS, and the qualitative results illustrate how well each component aligns with human perception.

Algorithm 2. Algorithm for the full computation of GLCS, which consists of GCS and LCS.

Quantitative Results

Qualitative Results

5-frame Qualitative Result 1 — Figure 7. IMPUS [ICLR’24] shows good domain consistency with the input image pair, but it contains abrupt transitions and therefore lacks smoothness. DiffMorpher [CVPR’24] provides smoother transitions, but its domain consistency is weak, with objects disappearing or becoming unstable. FreeMorph [ICCV’25] produces overly saturated colors, which are common artifacts in diffusion-based generation. In contrast, the proposed CHIMERA maintains both smoothness and domain consistency.

5-frame Qualitative Result 2 — Figure 7. IMPUS [ICLR’24] shows good domain consistency with the input image pair, but it contains abrupt transitions and therefore lacks smoothness. DiffMorpher [CVPR’24] provides smoother transitions, but its domain consistency is weak, with objects disappearing or becoming unstable. FreeMorph [ICCV’25] produces overly saturated colors, which are common artifacts in diffusion-based generation. In contrast, the proposed CHIMERA maintains both smoothness and domain consistency.

14-frame Qualitative Result 1 — Figure 8. This qualitative evaluation presents the more challenging 14-image morphing results. Consistent with Fig. 7, CHIMERA maintains both smoothness and domain consistency in this extended setting.

14-frame Qualitative Result 2 — Figure 8. This qualitative evaluation presents the more challenging 14-image morphing results. Consistent with Fig. 7, CHIMERA maintains both smoothness and domain consistency in this extended setting.

BibTeX citation

@article{kye2025chimera,
  title={CHIMERA: Adaptive Cache Injection and Semantic Anchor Prompting for Zero-shot Image Morphing with Morphing-oriented Metrics},
  author={Kye, Dahyeon and Sung, Jeahun and Jeon, Minkyu and Oh, Jihyong},
  journal={arXiv preprint arXiv:2512.07155},
  year={2025}
}