cs.CVOct 23, 2026

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

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

Organizations: University of North Carolina at Chapel Hill · Industrial Light & Magic · Fayetteville State University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

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 unintended changes to source content 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 composition and content of the original shot. To ensure KINESYS can produce a diverse set of edits that preserve the source, 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 preservation; on effect transfer, it remains competitive with Nano Banana 2 and outperforms open-weight editors.

Figures & tables

Figure 1: We present Kinesys , a test-time adaptation framework for reference-based visual effects editing in images. (Top) The artist supplies a source, effect reference, optional mask, editing prompt, and verification questions, and Kinesys generates multiple samples with diverse styles. (Bottom) Kinesys transfers puddle and boat-wake effects while preserving source composition and content; baselines change the scene, add unwanted content, or miss effect properties.
Figure 2: Overview of Kinesys . Reference-grounded VQA and source-aware EditReward provide complementary feedback. Each of KK independent runs has its own LoRA parameters ϕk\phi_{k} and NN fixed-noise candidates. Within a run, αi\alpha_{i} weights candidate ii ’s objective, while the decaying term λdiv(t)D\lambda_{\mathrm{div}}(t)\mathcal{D} discourages mode collapse. The axis from 00 to TT marks optimization steps; less promising runs stop at pruning points, while survivors continue to later checkpoints.
Table 1: Comparison on 156 shared cases from 81 requests and two seeds. Open-weight methods are ordered by Gemini; bold and underline mark first and second. †\dagger : EditReward guides Kinesys tuning and selection (not independent). ‡\ddagger : Nano Banana 2 and the judge share the Gemini family.
Figure 4: Six-effect comparison. Columns: reference, source, Kinesys , Nano Banana 2, Step1X-Edit v1.2. For snow, Kinesys retains the source house and viewpoint while Nano Banana 2 changes both. For fire, Kinesys adds flames to trash cans without Step1X-Edit’s scene-wide blaze.
Figure 5: Feedback ablations on explosion and wake edits. The “w/o” labels denote removal of the named gradient; red boxes mark unintended source changes without EditReward. Without VQA, effect form departs from the reference. Full Kinesys preserves source content and transfers the effect.
Figure 6: Fresh-noise variations in smoke shape, snow intensity, fire extent, and cast shadows. Each pair uses the same source, reference, instruction, and selected LoRA; only the noise changes.
Figure 7: Spatially controlled effect-reference edits. Columns show the source, pink edit region, effect reference, and Kinesys result.
Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A.1: Qualitative comparisons, rows 1 to 10. Columns: effect reference, source, Kinesys , Nano Banana 2, Step1X-Edit v1.2, Dual-Process, FLUX.2 Klein 4B, and Qwen-Image-Edit.
Figure A.2: Qualitative comparisons, rows 11 to 20. Columns: effect reference, source, Kinesys , Nano Banana 2, Step1X-Edit v1.2, Dual-Process, FLUX.2 Klein 4B, and Qwen-Image-Edit.
Figure B.1: The 81 requests by effect family. Counts are requests, not unique photographs.
Effect familyRequestsReferencesFamily-specific checks
Boat wake1010Placement behind the boat; turbulence and spread
Vehicle dent84Body-panel location; size and depth
Explosion108Location; scale and intensity
Fire118Flames on the subject; size and intensity
Puddle87Foreground placement; reflection clarity
Cast shadow118Position beneath the subject; sharpness, direction, and length
Appendix
Table B.1: Composition of the 81 requests. Counts are requests and distinct reference images. The listed properties summarize location and questions specific to each effect.
Effect familyRequest 1Request 2
Boat wakeboatwake.4boatwake.10
Vehicle dentdent.7dent.8
Explosionexplosion.3explosion.9
Firefire.8fire.9
Puddlepuddle.2puddle.5
Cast shadowshadow.5shadow.11
Appendix
Table C.1: Requests in the search configuration analysis, each evaluated under both seeds.
Figure E.1: Preference by effect family for source preservation and effect transfer against Nano Banana 2, using 41 responses. Purple favors Kinesys ; gray denotes a tie or unclear choice; red favors the baseline.
Figure E.2: Preference by effect family for source preservation and effect transfer against Step1X-Edit v1.2, using 40 responses. Colors follow Figure E.1 .
Figure E.3: Example survey question for the boat-wake effect, showing the video and original wording. The first question corresponds to source preservation.
More in this topic

Image Editing

Image editing changes selected visual content while preserving the parts of an image that should remain unchanged.

See all 313 papers

Explore similar work

CardsList

  1. Refinement Is Inherently Editable: Training-Free Prompt-to-Prompt Image Editing with Generative Refinement Network

    Sep 17, 2026Yulong Chen, Ziqian Zhang, Haoyu Zhang +4Text-Guided Image EditingTraining-Free Image Editing

    City University of Hong Kong (Dongguan), Guangdong, China · City University of Hong Kong, Hong Kong, China · Mohamed bin Zayed University of Artificial Intelligence, Masdar, Abu Dhabi

  2. GRNEdit: Efficient General Video Editing from a New Binary-Evidence Perspective in Generative Refinement Networks

    Aug 17, 2026Feng Xie, Jiagao Hu, Fuhao Li +5Video EditingInstruction-Based Video Editing

    Xidian University · MiLM Plus, Xiaomi Inc.

  3. Mind the RefGAP: Correcting Reference Attention in Diffusion-Based Visual Editing

    Sep 28, 2026Yanan Wang, Shengcai Liao, Guangyi Liu +1Video Diffusion ModelsAttention Control in Diffusion Models

    Mohamed bin Zayed University of Artificial Intelligence · Institute of Foundation Models · United Arab Emirates University