AceTone: Bridging Words and Colors for Conditional Image Grading

Tianren Ma; Mingxiang Liao; Xijin Zhang; Qixiang Ye

cs.CVApr 1, 2026

AceTone: Bridging Words and Colors for Conditional Image Grading

Authors: Tianren Ma, Mingxiang Liao, Xijin Zhang, Qixiang Ye

Organizations: University of Chinese Academy of Sciences · ByteDance

Abstract

Color affects how we interpret image style and emotion. Previous color grading methods rely on patch-wise recoloring or fixed filter banks, struggling to generalize across creative intents or align with human aesthetic preferences. In this study, we propose AceTone, the first approach that supports multimodal conditioned color grading within a unified framework. AceTone formulates grading as a generative color transformation task, where a model directly produces 3D-LUTs conditioned on text prompts or reference images. We develop a VQ-VAE based tokenizer which compresses a $3\times32^3$ LUT vector to 64 discrete tokens with $ΔE<2$ fidelity. We further build a large-scale dataset, AceTone-800K, and train a vision-language model to predict LUT tokens, followed by reinforcement learning to align outputs with perceptual fidelity and aesthetics. Experiments show that AceTone achieves state-of-the-art performance on both text-guided and reference-guided grading tasks, improving LPIPS by up to 50% over existing methods. Human evaluations confirm that AceTone's results are visually pleasing and stylistically coherent, demonstrating a new pathway toward language-driven, aesthetic-aligned color grading.

Figures & tables

Figure 1: AceTone performs conditional color grading under two paradigms: reference-based (top) and instruction-based (bottom). It accurately captures subtle tonal characteristics, follows user intents, and produces visually coherent color adjustments.

Figure 2: Overview of the AceTone framework. AceTone integrates a vector-quantized tokenizer to represent grading operations. Its phased training includes: (1) vision–language pretraining to learn tone prediction from image pairs, and (2) a post-train stage with SFT and RL (simplified in this figure). During RL training, multiple tone hypotheses are generated and optimized via group relative policy optimization (GRPO), enabling faithful and controllable tone manipulation aligned with user intent.

Benchmarks	Task	#Samples	Ground Truth
DPST [ 24 ]	Transfer	60	none
PST-50 [ 7 ]	Transfer	50	image
MMArt-Bench [ 22 ]	Instruct	200	cropped image
AceTone-Bench	Transfer Instruct	1024 128	image & LUT image & LUT

Table 1: Comparison of existing color grading benchmarks.

Method	Venue	PST-50 (50 samples)				AceTone-Bench $[\text{Transfer}]$ (1024 samples)
Method	Venue	Aes. $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	$\Delta\text{E}\downarrow$	Aes. $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	$\Delta\text{E}\downarrow$
Neural Preset † [ 12 ]	CVPR-23	3.03	21.24	0.15	9.57	–	–	–	–
WCT 2 [ 34 ]	ICCV-19	2.95	19.62	0.18	10.91	2.69	15.12	0.32	18.23
ModFlow [ 17 ]	AAAI-25	3.08	20.13	0.16	10.62	2.96	15.61	0.31	17.79
SA-LUT [ 7 ]	ICCV-25	3.07	21.64	0.16	9.01	3.33	18.14	0.22	13.10
AceTone (ours)		3.29	24.26	0.09	7.26	3.57	22.49	0.11	8.98

Table 2: Comparison on photorealistic style transfer. Aes. stands for the aesthetic score given by DeQA. † : Neural Preset has not been open-sourced, and is only available as a mobile app without support for batch process, therefore we manually evaluate it on PST-50.

Method	Arc.	Aes. $\uparrow$	LPIPS $\downarrow$	Rank	Time
InstructPix2Pix [ 2 ]	Diff.	2.97	0.44	3.68	2s
Qwen-Image-Edit [ 32 ]	Diff.	3.19	0.33	2.37	2min
JarvisArt [ 22 ]	Agent	3.17	0.31	2.72	40s
AceTone (ours)	LUT	3.54	0.22	1.43	1.2s

Table 3: Comparison on instruction-guided grading. Arc. refers to the model architecture, e.g. , diffusion, parameter-based software agent, or color transformation with LUTs.

Figure 3: Visualization of the tokenizer’s compression quality. We apply the original and reconstructed LUT to the image, and report the corresponding $\Delta\text{E}$ metric.

Figure 4: A Qualitative Visualization of AceTone. Top: Style transfer comparison. AceTone effectively captures the target color style while avoiding color banding or unnatural hue shifts. Bottom: Instruction-guided grading comparison (from AceTone-Bench $[\text{Instruct}]$ ). AceTone faithfully follows user intent and produces visually appealing adjustments. Best viewed in color and zoomed in for full details.

Reward Metrics	PSNR $\uparrow$	LPIPS $\downarrow$	$\Delta\text{E}\downarrow$	Aes. $\uparrow$
$\Delta\text{E}$	22.51	0.11	8.93	3.29
$\Delta\text{E}$ & LPIPS	22.24	0.10	9.05	3.28
$\Delta\text{E}$ & PSNR	22.58	0.11	9.01	3.31
$\Delta\text{E}$ & Aes.	22.49	0.11	8.98	3.57

Table 4: Ablation on Reward Designs.

Figure 5: Performance Curve during Training.

Method	Task	Pref. (%)	Rank
Neural Preset [ 12 ]	PST	30.1	2.03
SA-LUT [ 7 ]		28.5	2.22
AceTone (ours)		41.4	1.75
JarvisArt [ 22 ]	IGG	18.3	2.34
Qwen-Image-Edit [ 32 ]		34.8	1.97
AceTone (ours)		46.9	1.69

Table 5: User Study. The correlated statistics are: Friedman

\chi^{2}(2)=58.14,p<0.001

and inter-rater agreement

\kappa=0.213

Image Editing

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

See all 313 papers

Multimodal Image Colorization: Quantifying the Impact of Text-Conditioned Guidance on Grayscale-to-Color Translation
Jun 16, 2026Colten Reissmann, Hugo Garrido-Lestache BelinchonImage-to-Image Translation Conditional Image Generation
Department of Computer Science and Software Engineering[cs.GR] Milwaukee School of Engineering Milwaukee, WI US
LumiVideo: An Intelligent Agentic System for Video Color Grading
Apr 2, 2026Yuchen Guo, Junli Gong, Hongmin Cai +2Agentic Video Editing Automated Grading
Northwestern University, Evanston, USA · Northeastern University, Seattle, USA · South China University of Technology, Guangzhou, China +2
Adaptive Color Grading
Sep 18, 2026Trevor D. Canham, Abhijith Punnappurath, Michael S. BrownImage Editing Image Enhancement
York University, Canada

Jun 16, 2026J/K move · Enter open · S save

Multimodal Image Colorization: Quantifying the Impact of Text-Conditioned Guidance on Grayscale-to-Color Translation

Colten Reissmann, Hugo Garrido-Lestache Belinchon

Department of Computer Science and Software Engineering[cs.GR] Milwaukee School of Engineering Milwaukee, WI US

Image-to-Image Translation Conditional Image Generation

Open paper AskPDF arXiv

Grayscale images are commonly found in historical photography restoration, medical imaging, and artistic media. However, automatically applying color to these images remains a significant challenge in computer vision because many plausible colorizations can correspond to the same grayscale input. In this work, we quantify the effect of text conditioning on pixel-level and perceptual metrics for grayscale-to-color image models. Specifically, we compare two architectures, a U-Net and Stable Diffusion 1.5, each tested with and without CLIP text conditioning while holding all other variables constant. Our results show that text conditioning improves PSNR by 5.6%, SSIM by 1.2%, and colorfulness by 36.6%, while reducing LPIPS by 7.6% in the U-Net tier. In the Stable Diffusion tier, text conditioning improves PSNR by 5.8%, SSIM by 1.5%, and colorfulness by 0.6%, while reducing LPIPS by 11.3%. These results indicate that text conditioning provides consistent, measurable improvements to colorization quality across both architecture scales.

Abstract

Image Editing

Explore similar work

Multimodal Image Colorization: Quantifying the Impact of Text-Conditioned Guidance on Grayscale-to-Color Translation

LumiVideo: An Intelligent Agentic System for Video Color Grading

Adaptive Color Grading