Linear Image Generation by Synthesizing Exposure Brackets

2026-04-24 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, extended

Summary

Researchers from S-Lab, Nanyang Technological University, Adobe NextCam, and Adobe Research have developed a novel generative framework for text-to-linear-image synthesis, addressing the limitations of current generative models that primarily produce display-referred images. Their method, which uses a DiT-based flow-matching architecture, represents a linear image as a sequence of four exposure brackets (EVs of -4, -2, 0, 2) to overcome the challenges of high dynamic range and bit depth that VAEs struggle with. The framework incorporates exposure modulation self-attention and a radiance-scale token denoising mechanism for joint radiance scale and image content prediction. Trained on a dataset of 25,000 RAW images from RAISE and Adobe FiveK, the model achieves superior visual quality and dynamic range compared to adapted state-of-the-art text-to-image and text-to-video models. This approach also enables downstream applications like linear image editing, inpainting, and ControlNet-guided conditional generation.

Key takeaway

For research scientists developing advanced image generation models, this work demonstrates a robust approach to synthesizing high-dynamic-range linear images. You should consider adopting multi-exposure bracket generation and radiance-scale token denoising to overcome VAE limitations when working with scene-referred data, enabling richer post-processing capabilities and more physically accurate outputs. This framework provides a strong foundation for future work in professional photography workflows and computational imaging.

Key insights

Generating linear images as exposure brackets overcomes VAE limitations for high dynamic range content.

Principles

• Linear images offer superior post-processing flexibility.
• Decomposing HDR into exposure brackets aids generative models.
• Jointly predicting radiance scale improves scene reconstruction.

Method

The method uses a flow-matching framework with a DiT backbone to synthesize multiple exposure brackets, fusing them into a linear image. It integrates exposure modulation self-attention and a radiance-scale token denoising mechanism.

In practice

• Use 3D-RoPE for multi-bracket positional encoding.
• Employ LoRA for efficient fine-tuning of large models.
• Apply exposure modulation only to Single-DiT components.

Topics

• Linear Image Generation
• Exposure Bracketing Synthesis
• Flow-Matching Diffusion Models
• Diffusion Transformers
• Radiance Scale Prediction

Code references

• black-forest-labs/flux

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• High-Fidelity Two-Step Image Generation via Teacher-Aligned End-to-End Distillation — Tailored distillation strategies can significantly improve the quality-efficiency trade-off for few-step image generation.
• BRDFusion: Physics Meets Generation for Urban Scene Inverse Rendering — BRDFusion unifies physical modeling and generative priors for robust, controllable inverse rendering of urban scenes, enhancing quality and consistency.
• Image Restoration via Diffusion Models with Dynamic Resolution — Dynamic resolution DMs accelerate image restoration by operating in lower-dimensional subspaces, avoiding VAE overhead.
• Generating HDR Video from SDR Video — A new framework converts SDR video to HDR using generative models and a multi-exposure prediction and merging process.
• Fast and Lightweight Novel View Synthesis with Differentiable Multiplane Image — MPI, enhanced with diffusion and point maps, offers a fast, lightweight alternative for novel view synthesis.
• AsyncPatch Diffusion: spatially-flexible image generation — AsyncPatch Diffusion enables spatially flexible image generation by assigning distinct noise levels to different regions, unifying various generative tasks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.