ControlLight Turns Photo Brightening into a Smooth Dimmer Switch
ControlLight is a new open-source model that lets you brighten low-light photos using a simple slider, adjusting the enhancement strength from subtle to full correction. The system builds on the FLUX.2 Klein 9B base model and applies a LoRA adapter to handle illumination without changing the scene’s original layout, colors, or fine textures. Because it was trained on a continuous range of darkness levels, you can smoothly dial in the exact look you want instead of settling for a single preset.
A team of researchers led by Yufeng Yang developed this tool and released both the training code and a new dataset called Light100K alongside the model. Their paper highlights that older low-light enhancers often fail on real-world photos because they learn from limited examples with one fixed target brightness. ControlLight solves that by using flow matching and a special loss function that keeps results consistent no matter where you set the strength slider.
Adjust enhancement strength like a dimmer switch
Key features that set ControlLight apart
- Continuous enhancement control via a single alpha parameter.
- Preserves original scene structure and fine details.
- Built as a lightweight LoRA for FLUX.2 Klein 9B.
- Works in ComfyUI for local diffusion workflows.
- Trained on Light100K, a continuous illumination dataset.
- Consistent output across all enhancement strength values.
Photographers and video editors who often fix poorly lit footage can use ControlLight to recover shadow areas while keeping skin textures and object edges natural. Privacy-minded professionals benefit from running the model entirely on their own GPU, avoiding cloud uploads. Small creative studios can adopt it to batch-process dark images with predictable, adjustable results that maintain visual consistency across a series of shots.
Developer notes and what’s next
The model weights and code are released under an Apache 2.0 license, but users must also follow the upstream license terms from FLUX.2 Klein 9B and any other third-party components. Compared to existing low-light tools, ControlLight showed state-of-the-art performance in the researchers’ tests while adding the flexible strength control that most approaches lack. The team also made the Light100K dataset public to encourage further research into continuous illumination editing.
"The model is designed to enhance low-light images while preserving the original scene structure, visual content, and fine-grained details." — Source: Reddit
Project Links
