PREEN
A preset regression model based on an example editor network. Given a content image and a style reference image, it predicts 29-dimensional smartphone editor preset parameters, enabling interpretable and re-editable color style transfer.
Timeline (last 90 days)
PREEN (Preset Regression via an Exemplar-based Editor Network) proposes a color style transfer method: given a content image and a style reference image, it predicts a set of smartphone editor preset parameters (29-dimensional slider values) rather than outputting pixels.
PREEN transforms style transfer into a regression problem over the editor parameter space, and its output is a set of preset parameters that can be manually adjusted and re-edited, offering interpretability.
The training supervision signal of PREEN is generated by introducing a style transfer teacher model: the teacher model's target appearance is inverted into the editor parameter space to generate pseudo-presets.
The specific training pipeline of PREEN consists of two steps: first pretraining a CLIP-conditioned regressor, then fine-tuning with LoRA under an image-level CIELAB loss.
In PREEN, CLIP is used as the condition encoder for the style reference; compared with directly using the reference image's RGB pixels as the condition, CLIP features are more robust to semantic content and can mitigate interference from specific objects in the reference image on style judgment.
On the PST50 benchmark, PREEN leads baseline methods on pixel-level and perceptual fidelity metrics while maintaining structural fidelity.
Because its output is a compact 29-dimensional preset vector, PREEN achieves sub-second latency for on-device inference, making it suitable for real-time smartphone application scenarios.
The low-dimensional parameter subspace of PREEN provides an inference-time strength knob: in a single forward pass, users can adjust along a continuous trajectory from "default" to "vivid", and every intermediate state is a valid editor preset.
A two-axis human study shows that raters prefer the vivid end when matching the reference image, but prefer the default end when viewing it as a standalone photo.
The limitation of PREEN is that it depends on a fixed 29-parameter editor, and if the editor parameter system changes, the model needs to be re-adapted; in addition, the quality of the pseudo-presets is influenced by the teacher model, and the teacher model's bias may propagate into the final presets.
All Facts (10)
PREEN (Preset Regression via an Exemplar-based Editor Network) proposes a color style transfer method: given a content image and a style reference image, it predicts a set of smartphone editor preset parameters (29-dimensional slider values) rather than outputting pixels.
50%UnverifiedPREEN transforms style transfer into a regression problem over the editor parameter space, and its output is a set of preset parameters that can be manually adjusted and re-edited, offering interpretability.
50%UnverifiedThe training supervision signal of PREEN is generated by introducing a style transfer teacher model: the teacher model's target appearance is inverted into the editor parameter space to generate pseudo-presets.
50%UnverifiedThe specific training pipeline of PREEN consists of two steps: first pretraining a CLIP-conditioned regressor, then fine-tuning with LoRA under an image-level CIELAB loss.
50%UnverifiedIn PREEN, CLIP is used as the condition encoder for the style reference; compared with directly using the reference image's RGB pixels as the condition, CLIP features are more robust to semantic content and can mitigate interference from specific objects in the reference image on style judgment.
50%UnverifiedOn the PST50 benchmark, PREEN leads baseline methods on pixel-level and perceptual fidelity metrics while maintaining structural fidelity.
50%UnverifiedBecause its output is a compact 29-dimensional preset vector, PREEN achieves sub-second latency for on-device inference, making it suitable for real-time smartphone application scenarios.
50%UnverifiedThe low-dimensional parameter subspace of PREEN provides an inference-time strength knob: in a single forward pass, users can adjust along a continuous trajectory from "default" to "vivid", and every intermediate state is a valid editor preset.
50%UnverifiedA two-axis human study shows that raters prefer the vivid end when matching the reference image, but prefer the default end when viewing it as a standalone photo.
50%UnverifiedThe limitation of PREEN is that it depends on a fixed 29-parameter editor, and if the editor parameter system changes, the model needs to be re-adapted; in addition, the quality of the pseudo-presets is influenced by the teacher model, and the teacher model's bias may propagate into the final presets.
50%