cs.CVMar 31, 2026
SaveRawGen: Learning Camera Raw Image Generation
Organizations: AI Center - Toronto, Samsung Electronics · Yonsei University
Abstract
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.
Figures & tables
Figure 2: Overview of the RawGen framework. During training, raw images are converted to CIE XYZ and sRGB representations to (A) fine-tune the DiT to denoise CIE XYZ latents conditioned on an sRGB image, and (B) fine-tune the VAE decoder to reconstruct CIE XYZ images. During inference (C), either an sRGB image (image-to-raw, I2R) or a text prompt (text-to-raw, T2R) is used to condition the DiT to generate a CIE XYZ latent, which is decoded to obtain a CIE XYZ image and subsequently mapped to the target camera’s raw space using its calibration metadata, which can be easily acquired from a single DNG file of target camera.
| Expert A | Expert B | Expert C | Expert D | Expert E | ||||||
| Method | PSNR | SSIM | PSNR | SSIM | PSNR | SSIM | PSNR | SSIM | PSNR | SSIM |
| CIE XYZ Net [ 3 ] | 19.60 | 0.7861 | 21.04 | 0.8431 | 19.49 | 0.7795 | 19.44 | 0.8058 | 18.64 | 0.7918 |
| InvISP [ 70 ] | 16.04 | 0.6854 | 16.04 | 0.6854 | 14.92 | 0.6323 | 14.24 | 0.6802 | 13.30 | 0.6473 |
| Raw-Diffusion [ 54 ] | 19.30 | 0.7832 | 20.66 | 0.8397 | 18.79 | 0.7705 | 18.69 | 0.7966 | 17.49 | 0.7751 |
| RawGen | 19.57 | 0.7782 | 21.21 | 0.8349 | 19.48 | 0.7741 | 18.74 | 0.7929 | 18.07 | 0.7839 |
| RawGen (Ours) | 23.20 | 0.8432 | 24.35 | 0.8581 | 23.37 | 0.8387 | 23.51 | 0.8531 | 23.89 | 0.8500 |
Table 1: Quantitative comparison of sRGB-to-XYZ reconstruction across five photo-finishing styles on the MIT-Adobe FiveK dataset [ 12 ] . We report results for ablated variant of RawGen. Best results are highlighted in yellow .
Figure 3: CIE XYZ reconstruction results. Shown is an sRGB input image rendered with different rendering preferences (Expert A–E) and the corresponding CIE XYZ reconstructions produced by Raw-Diffusion [ 54 ] , CIE XYZ Net [ 3 ] , and our RawGen. The last row shows the ground-truth CIE XYZ image.
Figure 5: Raw reconstruction and generation results. Shown are the input sRGB images (A) and corresponding ground-truth raw images (B). Our reconstructed raw results are shown in (C). Additional generated raw images of the same semantic scene are shown in (D), where InstructBLIP-2 [ 18 ] is used to generate text descriptions from image (A).
Figure 6: Rendering results of pre-trained neural ISPs on real data from the Samsung S24 main camera (S24 dataset [ 6 ] ). (A) CIE XYZ image generated by our RawGen. (B) Image mapped to the S24 main camera raw space with synthetic noise. (C) Result of PyNet [ 33 ] . (D) Result of Modular Neural ISP [ 5 ] .
| Denoising | ||||||||||
| Illuminant estimation | Neural ISP | ISO 1600 | ISO 3200 | |||||||
| Method | Mean | Median | Worst 25% | PSNR | SSIM | PSNR | SSIM | PSNR | SSIM | |
| EnlightenGAN [ 35 ] | 7.01 | 6.82 | 11.07 | 35.58 | 0.965 | 3.137 | 48.82 | 0.991 | 47.25 | 0.988 |
| UPI [ 11 ] | 6.26 | 5.89 | 10.33 | 36.43 | 0.966 | 2.907 | 49.05 | 0.990 | 47.51 | 0.988 |
| Graphics2RAW [ 61 ] | 4.21 | 3.38 | 8.57 | 38.10 | 0.974 | 2.301 | 49.37 | 0.991 | 48.16 | 0.989 |
| RawGen | 3.14 | 2.11 | 7.37 | 38.42 | 0.970 | 2.183 | 50.63 | 0.994 | 48.57 | 0.992 |
Table 3: Quantitative comparison of downstream tasks: illuminant estimation (reported as angular error statistics), neural ISP learning, and denoising at two ISO levels. Each row corresponds to a model trained using data generated by the method listed. Models trained on real data are shown in the bottom row as a reference. The best results of models trained on non-real raw data are highlighted in yellow .
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
| Parameter | Distribution | Range | Stage |
| Red WB multiplier | Uniform | White balance | |
| Blue WB multiplier | Uniform | White balance | |
| Tone-map | Normal | (clip ) | Tone mapping |
| Tone-map | Normal | (clip ) | Tone mapping |
| Contrast | Uniform | sRGB |
Table 4: Photo-finishing parameter ranges used for sRGB variant generation.
Figure 7: Variations of raw images generated from the CIE XYZ image shown in (A). (B) shows raw images mapped to the Samsung Galaxy S24 main camera raw space under different correlated color temperatures (CCTs). (C) shows raw images mapped with a fixed CCT of 7000 K and different noise strength factor values . All images are gamma-corrected for visualization throughout this supplementary material.
Figure 8: Additional CIE XYZ reconstruction results. Shown is an sRGB input image rendered with different rendering preferences (Expert A–E) and the corresponding CIE XYZ reconstructions produced by Raw-Diffusion [ 54 ] , CIE XYZ Net [ 3 ] , and our RawGen. The last row shows the ground-truth CIE XYZ image.
Figure 9: Additional CIE XYZ reconstruction results. Shown is an sRGB input image rendered with different rendering preferences (Expert A–E) and the corresponding CIE XYZ reconstructions produced by Raw-Diffusion [ 54 ] , CIE XYZ Net [ 3 ] , and our RawGen. The last row shows the ground-truth CIE XYZ image.
Figure 10: Additional CIE XYZ reconstruction results. Shown is an sRGB input image rendered with different rendering preferences (Expert A–E) and the corresponding CIE XYZ reconstructions produced by Raw-Diffusion [ 54 ] , CIE XYZ Net [ 3 ] , and our RawGen. The last row shows the ground-truth CIE XYZ image.
Figure 11: Additional CIE XYZ reconstruction results. Shown is an sRGB input image rendered with different rendering preferences (Expert A–E) and the corresponding CIE XYZ reconstructions produced by Raw-Diffusion [ 54 ] , CIE XYZ Net [ 3 ] , and our RawGen. The last row shows the ground-truth CIE XYZ image.
Figure 12: Additional CIE XYZ reconstruction results. Shown is an sRGB input image rendered with different rendering preferences (Expert A–E) and the corresponding CIE XYZ reconstructions produced by Raw-Diffusion [ 54 ] , CIE XYZ Net [ 3 ] , and our RawGen. The last row shows the ground-truth CIE XYZ image.
Figure 13: Additional raw image reconstruction and generation results. Shown are the input sRGB images (A) and corresponding ground-truth raw images (B). Our reconstructed raw results are shown in (C). Additional generated raw images of the same semantic scene are shown in (D), where InstructBLIP-2 [ 18 ] is used to generate text descriptions from image (A).
Figure 14: Additional results of our RawGen, generated from the text prompts shown at the bottom of each image. For each prompt, we show generated raw images for three target cameras.
Figure 15: Examples of generated raw images for the Canon 1Ds MkIII DSLR camera. The images were generated using the following text prompts: 1) Cracked ceramic mug with visible hairline fractures, natural lighting , 2) convenience store at 2am, viewed through a glass facade , 3) commuters stepping onto an escalator at the exact moment of boarding , 4) mountain valley just before sunrise with rising ground fog , 5) peeling paint on a weathered metal door , and 6) city sidewalk just as a shallow puddle finishes forming .
Figure 16: Examples of generated raw images for the Olympus EPL6 DSLR camera. The images were generated using the following text prompts: 1) ocean waves striking a seawall at the exact moment of impact , 2) shadow band phenomenon in the final seconds before totality , 3) wind-driven sand crossing a desert road at sunset , 4) coastal fishing village harbor at dawn before activity begins , 5) suburban parking lot after freezing rain , and 6) pedestrian mid-step with both feet briefly off the ground .
Figure 17: Examples of generated raw images for the Nikon D5200 DSLR camera. The images were generated using the following text prompts: 1) Close-up of human skin with visible pores and fine hairs , 2) narrow alley after rainfall , 3) roadside gas station in midday heat shimmer , 4) tree pollen drifting in golden backlight , 5) coastal cliff facing the ocean in strong onshore winds , and 6) subway platform during non-peak hours .
Figure 18: Examples of generated raw images for the Samsung NX2000 DSLR camera. The images were generated using the following text prompts: 1) Thin ice forming across a pond at the transition point , 2) person adjusting glasses at the precise midpoint of the motion , 3) visible breath at the exact moment of exhale in cold air , 4) lightning illuminating low cloud cover without direct bolt visibility , 5) coastal fog rolling over a highway overpass , and 6) newspaper print under fluorescent lighting with slight glare .
Figure 19: Rendering results of pre-trained neural ISPs on real data from the Samsung S24 main camera (S24 dataset [ 6 ] ). (A) CIE XYZ images generated by our RawGen. (B) Images mapped to the S24 main camera raw space with synthetic noise. (C) Result of Lite ISP [ 75 ] . (D) Result of Invertible ISP [ 70 ] . (E) Result of Modular Neural ISP [ 5 ] .
More in this topic
Computational Imaging
Computational imaging combines optical and sensor design with algorithms to recover or infer scene information from indirect, compressed, or noisy measurements.
See all 123 papers