AnyBokeh: Physics-Guided Any-to-Any Bokeh Editing with Optical Fingerprint Transfer
Paper β’ 2606.31959 β’ Published β’ 5
How to use itsmag11/AnyBokeh with Diffusers:
pip install -U diffusers transformers accelerate
import torch
from diffusers import DiffusionPipeline
from diffusers.utils import load_image
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("itsmag11/AnyBokeh", dtype=torch.bfloat16, device_map="cuda")
prompt = "Turn this cat into a dog"
input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png")
image = pipe(image=input_image, prompt=prompt).images[0]Paper | Project Page | Code
AnyBokeh transforms an image from an arbitrary source optical state to a desired target focus and aperture setting. It consists of two LoRAs on top of FLUX.1-Fill-dev:
stage1-pytorch_lora_weights.safetensors: estimates the circle-of-confusion (CoC) map and the disparity map of the input image.stage2-pytorch_lora_weights.safetensors: re-renders the image with the target focus point and aperture.stage1-prompt_embeds.pt, stage1-pooled_prompt_embeds.pt, stage2-prompt_embeds.pt, and stage2-pooled_prompt_embeds.pt are the precomputed text embeddings, so the text encoders are not needed at inference.
git clone https://github.com/itsmag11/AnyBokeh.git
cd AnyBokeh
python inference_full.py --image_path examples/127_f5.0.JPG --focus_x 904 --focus_y 613 --source_aperture 5.0 --target_aperture 2.8 --output_path outputs/127_f5.0_to_f2.8.jpg
The weights in this repo are downloaded automatically. See the GitHub repo for more details.
@inproceedings{hou2026anybokeh,
title = {AnyBokeh: Physics-Guided Any-to-Any Bokeh Editing with Optical Fingerprint Transfer},
author = {Hou, Xinyu and Li, Xiaoming and Yue, Zongsheng and Loy, Chen Change},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
year = {2026}
}
Base model
black-forest-labs/FLUX.1-Fill-dev