KINESYS: Test-Time Adaptation for Reference-Based Effect Editing in Images

Frahm, Noah; Qi, Luchao; Li, Haoze; Goldman, Dan B; Bhattacharya, Sambit; Sengupta, Roni

KINESYSTest-Time Adaptation for
Reference-Based Effect Editing in Images

Noah Frahm¹, Luchao Qi¹, Haoze Li¹, Dan B Goldman², Sambit Bhattacharya³, Roni Sengupta¹

¹ University of North Carolina at Chapel Hill ² Industrial Light & Magic ³ Fayetteville State University

arXiv PDF

Transfer and Preserve

Colorful smoke editing: an effect reference, source racecar, mask, and editing prompt produce two results. The illustrated user prefers the first result because its smoke trail is fuller and longer.

Abstract

Visual effects are often easier to demonstrate via reference than to describe in words. We study effect-reference editing, where a reference image shows a desired visual effect and language identifies which properties should guide the edit. Existing single-pass editing techniques often fail to make fine-grained edits, resulting in subtle identity distortions or deviations from the reference effect. We introduce KINESYS, a test-time adaptation framework for effect-reference editing that combines reference-aware and source-aware feedback to transfer the desired effect while preserving the identity of the original shot. To ensure KINESYS can produce a diverse set of identity-preserving edits, we develop a robust optimization strategy in which the generative model is fine-tuned across multiple noise samples and optimization runs. Evaluation with VLM judges, source-preservation metrics (LPIPS, TPIPS), and a user study shows that KINESYS outperforms state-of-the-art image editors in source identity preservation; on effect transfer, it remains competitive with Nano Banana 2 and outperforms open-weight editors.

Effect Comparison

Loading effect comparisons…

How KINESYS works

KINESYS method diagram: source, reference, optional mask, prompt, and questions enter a fine-tuning loop. Reference-grounded VQA and source-aware EditReward evaluate candidate edits. Independent LoRA runs are pruned before a checkpoint is selected.

Editing quality and source preservation

Comparison on 156 shared request–seed cases (81 requests). Gemini, EditScore, and EditReward assess editing quality; LPIPS and TPIPS measure source similarity. ↑ higher is better; ↓ lower is better. Bold and underline mark first and second. † EditReward was used for tuning and selection and is not an independent evaluation. ‡ Generator–judge family overlap. Scores are shown to three decimal places.
Method	Automated judges			Identity preservation
Method	Gemini ↑	EditScore ↑	EditReward ↑	LPIPS ↓	TPIPS ↑
Open-weight methods
OmniGen2	4.950	7.530	−0.018	0.753	0.599
ACE++	4.980	5.950	−0.503	0.795	0.500
ICEdit	5.930	6.910	−0.097	0.429	0.632
Dual-Process	7.090	7.930	0.212	0.577	0.615
Qwen-Image-Edit	7.670	8.310	0.435	0.645	0.599
FLUX.2 Klein 4B	7.900	8.680	0.501	0.579	0.640
Step1X-Edit	8.220	8.800	0.666	0.449	0.724
Proprietary methods
Nano Banana 2	8.910^‡	9.060	0.738	0.429	0.696
KINESYS (Ours)	8.830	8.950	1.007^†	0.377	0.731

Blind human evaluation

User study

vs. Nano Banana 2

Identity preservation

KINESYS: 45.3%
Tie / unclear: 30.0%
Baseline: 24.7%

Effect transfer

KINESYS: 48.2%
Tie / unclear: 9.8%
Baseline: 42.1%

vs. Step1X-Edit v1.2

Identity preservation

KINESYS: 53.6%
Tie / unclear: 25.5%
Baseline: 20.9%

Effect transfer

KINESYS: 69.8%
Tie / unclear: 8.1%
Baseline: 22.0%

Blind paired comparisons evaluated identity preservation and reference-effect transfer separately across 16 source–reference cases, two from each of eight effect families. Participants saw anonymous A/B results in synchronized source-to-edit videos, with method positions counterbalanced and item order shuffled; most respondents were not VFX experts. Bars show preferences, including “Hard to tell / No clear winner.” Percentages are reproduced from the paper to one decimal place and may not sum to 100% due to rounding.

BibTeX

@misc{frahm2026kinesystesttimeadaptationreferencebased,
      title={KINESYS: Test-Time Adaptation for Reference-Based Effect Editing in Images},
      author={Noah Frahm and Luchao Qi and Haoze Li and Dan B Goldman and Sambit Bhattacharya and Roni Sengupta},
      year={2026},
      eprint={2610.33292},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.33292},
}

Acknowledgments

This work was partly supported by NCInnovation under the project “KINESYS+ATLAS: AI and Synthetic Data for Critical Applications”.