SpiralDiff: Spiral Diffusion with LoRA for RGB-to-RAW Conversion Across Cameras

Huanjing Yue; Shangbin Xie; Cong Cao; Qian Wu; Lei Zhang; Lei Zhao; Jingyu Yang

cs.CVMar 16, 2026

SpiralDiff: Spiral Diffusion with LoRA for RGB-to-RAW Conversion Across Cameras

Authors: Huanjing Yue, Shangbin Xie, Cong Cao, Qian Wu, Lei Zhang, Lei Zhao, Jingyu Yang

Organizations: School of Electrical and Information Engineering, Tianjin University · State key laboratory of Smart Power Distribution Equipment and System, Tianjin University · Individual

Abstract

RAW images preserve superior fidelity and rich scene information compared to RGB, making them essential for tasks in challenging imaging conditions. To alleviate the high cost of data collection, recent RGB-to-RAW conversion methods aim to synthesize RAW images from RGB. However, they overlook two key challenges: (i) the reconstruction difficulty varies with pixel intensity, and (ii) multi-camera conversion requires camera-specific adaptation. To address these issues, we propose SpiralDiff, a diffusion-based framework tailored for RGB-to-RAW conversion with a signal-dependent noise weighting strategy that adapts reconstruction fidelity across intensity levels. In addition, we introduce CamLoRA, a camera-aware lightweight adaptation module that enables a unified model to adapt to different camera-specific ISP characteristics. Extensive experiments on four benchmark datasets demonstrate the superiority of SpiralDiff in RGB-to-RAW conversion quality and its downstream benefits in RAW-based object detection. Our code and model are available at https://github.com/Chuancy-TJU/SpiralDiff.

Figures & tables

Figure 1: Comparison of noise schedule (top) and visualization of the noisy $\bm{x}_{T}$ (bottom) in diffusion: ResShift [ 45 ] uses uniform Gaussian noise, whereas our SpiralDiff introduces a signal-dependent noise schedule based on pixel intensity ( $\bm{w}_{t}$ ).

Figure 2: Relationship between RGB pixel intensity and residual magnitude between RGB and RAW images across channels on the FiveK Nikon dataset. Colored lines and shaded regions represent the mean and standard deviation, respectively.

Figure 3: Overview of the proposed SpiralDiff with CamLoRA. (a) SpiralDiff introduces a signal-dependent noise schedule via a weight map set $\{\bm{w}_{t}\}_{t=1}^{T}$ that aligns with the RAW-to-RGB conversion process. The noise level depends on local pixel intensity and diffusion timestep $t$ : darker regions ( $\triangle$ ) receive less noise, and brighter regions ( $\bigcirc$ ) receive more. (b) The spiral structure visualizes how noise scales with $\bm{w}_{t}$ for different pixel intensities. (c) shows the framework of SpiralDiff. The noisy image $\bm{x}_{t}$ , RGB image $\bm{y}_{0}$ and camera label are fed into denoising U-Net, which iteratively samples to refine the RAW output. (d) The camera label selects the camera-specific LoRA layer in CamLoRA, enhancing adaptation to each camera’s characteristic.

Methods	Separate				Combined
Methods	FiveK Canon	FiveK Nikon	NOD Nikon	NOD Sony	FiveK Canon	FiveK Nikon	NOD Nikon	NOD Sony
CycleISP [ 47 ]	37.93 / 0.9913	40.18 / 0.9920	50.11 / 0.9985	46.57 / 0.9975	38.06 / 0.9911	38.86 / 0.9918	48.68 / 0.9984	46.12 / 0.9971
InvISP [ 40 ]	36.81 / 0.9814	34.30 / 0.9163	48.29 / 0.9954	44.76 / 0.9922	35.37 / 0.9828	36.60 / 0.9850	47.73 / 0.9948	44.43 / 0.9907
ReRAW [ 3 ]	35.08 / 0.9608	33.09 / 0.9265	48.25 / 0.9906	45.50 / 0.9788	32.70 / 0.9684	33.24 / 0.9648	46.39 / 0.9715	45.51 / 0.9817
RAW-Diffusion [ 33 ]	39.96 / 0.9890	39.68 / 0.9866	50.52 / 0.9954	47.31 / 0.9908	40.63 / 0.9889	40.35 / 0.9916	49.60 / 0.9950	46.80 / 0.9897
SpiralDiff	42.82 / 0.9936	41.72 / 0.9925	53.64 / 0.9990	50.46 / 0.9980	41.99 / 0.9931	43.19 / 0.9947	52.06 / 0.9987	49.52 / 0.9974

Table 1: Performance comparison of different methods trained on separate and combined datasets, measured by PSNR(

\uparrow

) and SSIM(

\uparrow

). +CamLoRA denotes that CamLoRA is applied to the combined training setting. Best results under each setting are highlighted in bold .

Methods	FiveK Canon		NOD Nikon
Methods	PSNR	SSIM	PSNR	SSIM
CycleISP [ 47 ]	27.75	0.9833	35.86	0.9957
InvISP [ 40 ]	25.15	0.9699	36.66	0.9905
ReRAW [ 3 ]	26.36	0.9664	38.56	0.9691
RAW-Diffusion [ 33 ]	30.60	0.9739	40.05	0.9926
SpiralDiff	31.10	0.9856	40.79	0.9973

Table 2: Quantitative comparison on the over-exposed test set.

Figure 4: Qualitative comparison with state-of-the-art RGB-to-RAW conversion methods on the FiveK dataset (top two rows) and the NOD dataset (bottom two rows). For each result, the left half is the predicted RAW, and the right half is the error map. The proposed SpiralDiff shows better conversion results, especially in bright regions.

Variant	FiveK Canon		NOD Nikon
Variant	PSNR	SSIM	PSNR	SSIM
Baseline	41.40	0.9906	53.48	0.9989
$\bm{y}_{0}$	40.06	0.9912	53.42	0.9990
$\bm{w}_{t}$	42.82	0.9936	53.64	0.9990

Table 3: Ablations on the noise weighting strategies. The baseline uses spatially uniform noise.

\bm{y}_{0}

(

\bm{w}_{t}

) means the noise is weighted according to the intensity of

\bm{y}_{0}

(

\bm{w}_{t}

Variant	FiveK Canon		NOD Nikon
Variant	PSNR	SSIM	PSNR	SSIM
Uncond.	41.99	0.9931	52.06	0.9987
Embed.	40.97	0.9917	51.44	0.9984
CamLoRA	42.46	0.9934	52.62	0.9988

Table 4: Ablations on three conditioning strategies: Uncond. (no camera info), Embed. (camera embedding), and CamLoRA .

Figure 5: Residual-intensity relationship on rawpy ISP (FiveK Nikon) and real ISP (iPhone and Samsung).

Variant	1-shot		5-shot
Variant	PSNR	SSIM	PSNR	SSIM
Scratch	39.83	0.9895	43.40	0.9951
CamLoRA	42.85	0.9942	43.92	0.9954

Table 5: Few-shot RGB-to-RAW conversion on the SID Sony dataset. We compare training from scratch with CamLoRA adaptation under 1-shot and 5-shot settings.

Method	iPhone	Samsung
RAW-Diffusion (DDPM)	22.74 / 0.7503	30.28 / 0.9140
RAW-Diffusion (SpiralDiff)	30.11 / 0.8810	34.07 / 0.9322

Table 6: Plug-in experiment on the real-ISP dataset.

Dataset	NOD Nikon			NOD Sony
Dataset	AP	AP 50	AP 75	AP	AP 50	AP 75
RGB-only	19.1	37.0	18.5	19.7	38.3	18.0
RAW-only	18.4	35.1	17.0	17.6	35.7	15.1
RAW + CS-RAW	24.5	46.2	23.5	25.9	49.3	24.5
RAW + BDD-RAW	26.7	49.5	25.1	29.0	55.1	26.5

Table 7: Comparison of object detection results on the NOD test set with different training datasets.

Figure 13

Method	FiveK Canon		FiveK Nikon		NOD Nikon		NOD Sony
Method	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
CycleISP [ 47 ]	27.75	0.9833	30.19	0.9872	35.86	0.9957	32.52	0.9890
InvISP [ 40 ]	25.15	0.9699	29.11	0.9764	36.66	0.9905	30.85	0.9782
ReRAW [ 3 ]	26.36	0.9664	27.51	0.9520	38.56	0.9691	34.96	0.9713
RAW-Diffusion [ 33 ]	30.60	0.9739	32.95	0.9850	40.05	0.9926	35.58	0.9792
SpiralDiff	31.10	0.9856	34.42	0.9908	40.79	0.9973	36.11	0.9919

Table 8: Quantitative comparison results on the over-exposed test set.

Rank	FiveK Canon		FiveK Nikon		NOD Nikon		NOD Sony
Rank	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
$r=0$	41.99	0.9931	43.19	0.9947	52.06	0.9987	49.52	0.9974
$r=2$	42.64	0.9928	43.52	0.9948	52.23	0.9987	49.78	0.9977
$r=4$	42.34	0.9931	43.95	0.9948	52.47	0.9988	49.82	0.9977
$r=8$	42.46	0.9934	43.82	0.9950	52.62	0.9988	50.08	0.9977
$r=16$	42.00	0.9932	42.82	0.9945	52.36	0.9989	49.60	0.9978

Table 9: Ablation study on the LoRA rank

r

. We set

r=8

as the default setting for all experiments.

Variant	FiveK Canon		FiveK Nikon		NOD Nikon		NOD Sony
Variant	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
RAW-Diffusion (DDPM)	39.96	0.9890	39.68	0.9866	50.52	0.9954	47.31	0.9908
RAW-Diffusion (SpiralDiff)	41.53	0.9912	42.34	0.9920	50.54	0.9954	47.32	0.9907

Table 10: RAW-Diffusion (DDPM) denotes the original RAW-Diffusion model with its DDPM-based diffusion process, while RAW-Diffusion (SpiralDiff) replaces only this diffusion process with our SpiralDiff formulation, keeping the network architecture unchanged.

Table 17

Image-to-Image Translation

Image-to-image translation maps images between visual domains or modalities while preserving relevant content, structure, and semantics.

See all 115 papers

RawGen: Learning Camera Raw Image Generation
Mar 31, 2026Dongyoung Kim, Junyong Lee, Abhijith Punnappurath +4Computational Imaging Image Generation
AI Center - Toronto, Samsung Electronics · Yonsei University
RAWIC: Bit-Depth Adaptive Lossless Raw Image Compression
Mar 30, 2026Chunhang Zheng, Tongda Xu, Mingli Xie +2Learned Image Compression
School of Electronics, Peking University, Beijing, 100871, China · Institute for AI Industry Research (AIR), Tsinghua University, Beijing, 100084, China · the University of Chinese Academy of Sciences, Beijing, 100190, China
MERIT: Multi-domain Efficient RAW Image Translation
Mar 21, 2026Wenjun Huang, Shenghao Fu, Yian Jin +10Image-to-Image Translation
University of California, Irvine · University of Pennsylvania · Purdue University Northwest +1

Mar 31, 2026J/K move · Enter open · S save

RawGen: Learning Camera Raw Image Generation

Dongyoung Kim, Junyong Lee, Abhijith Punnappurath, Mahmoud Afifi+3

AI Center - Toronto, Samsung Electronics · Yonsei University

Computational Imaging Image Generation Diffusion-Based Image Generation Inverse Problems T2I Diffusion Models

Open paper AskPDF arXiv

Cameras capture scene-referred linear raw images, which are processed by onboard image signal processors (ISPs) into display-referred 8-bit sRGB outputs. Although raw data is more faithful for low-level vision tasks, collecting large-scale raw datasets remains a major bottleneck, as existing datasets are limited and tied to specific camera hardware. Generative models offer a promising way to address this scarcity -- however, existing diffusion frameworks are designed to synthesize photo-finished sRGB images rather than physically consistent linear representations. This paper presents RawGen, to our knowledge the first diffusion-based framework enabling text-to-raw generation for arbitrary target cameras, alongside sRGB-to-raw inversion. RawGen leverages the generative priors of large-scale sRGB diffusion models to synthesize physically meaningful linear outputs, such as CIE XYZ or camera-specific raw representations, via specialized processing in latent and pixel spaces. To handle unknown and diverse ISP pipelines and photo-finishing effects in diffusion-model training data, we build a many-to-one inverse-ISP dataset where multiple sRGB renditions of the same scene generated using diverse ISP parameters are anchored to a common scene-referred target. Fine-tuning a conditional denoiser and specialized decoder on this dataset allows RawGen to obtain camera-centric linear reconstructions that effectively invert the rendering pipeline. We demonstrate RawGen's superior performance over traditional inverse-ISP methods that assume a fixed ISP. Furthermore, we show that augmenting training pipelines with RawGen's scalable, text-driven synthetic data can benefit downstream low-level vision tasks.

Abstract

Image-to-Image Translation

Explore similar work

RawGen: Learning Camera Raw Image Generation

RAWIC: Bit-Depth Adaptive Lossless Raw Image Compression

MERIT: Multi-domain Efficient RAW Image Translation