Welcome to the L2M repository! This is the official implementation of our ICCV'25 paper titled "Learning Dense Feature Matching via Lifting Single 2D Image to 3D Space".
Accepted to ICCV 2025 Conference
fromromatchimportl2mpp_modelimporttorchimportcv2device=torch.device("cuda") iftorch.cuda.is_available() elsetorch.device("cpu")
l2mpp=l2mpp_model(device=device, version="v2") # , version="v1"# Matchwarp, certainty=l2mpp.match("assets/sacre_coeur_A.jpg", "assets/sacre_coeur_B.jpg", device=device)
matches, certainty=l2mpp.sample(warp, certainty)
# Convert to pixel coordinates (RoMa produces matches in [-1,1]x[-1,1])kptsA, kptsB=roma_model.to_pixel_coordinates(matches, H_A, W_A, H_B, W_B)
# Find a fundamental matrix (or anything else of interest)F, mask=cv2.findFundamentalMat(
kptsA.cpu().numpy(), kptsB.cpu().numpy(), ransacReprojThreshold=0.2, method=cv2.USAC_MAGSAC, confidence=0.999999, maxIters=10000
)Pretrained checkpoints for L2M++ are publicly available on Hugging Face:
👉 Hugging Face:
https://huggingface.co/datasets/Liangyingping/L2Mpp-checkpoints
The model will automatically download the required weights when running inference.
Lift to Match (L2M) is a two-stage framework for dense feature matching that lifts 2D images into 3D space to enhance feature generalization and robustness. Unlike traditional methods that depend on multi-view image pairs, L2M is trained on large-scale, diverse single-view image collections.
To enable training from single-view images, we simulate diverse multi-view observations and their corresponding dense correspondence labels in a fully automatic manner.
We lift a single-view image to a coarse 3D structure and then render novel views from different camera poses. These synthesized multi-view images are used to supervise the feature encoder with dense matching consistency.
Run the following to generate novel-view images with ground-truth dense correspondences:
python get_data.py \
--output_path [PATH-to-SAVE] \
--data_path [PATH-to-IMAGES] \
--disp_path [PATH-to-MONO-DEPTH]
This code provides an example on novel view generation with dense matching ground truth.
The disp_path should contain grayscale disparity maps predicted by Depth Anything V2 or another monocular depth estimator.
Below are examples of synthesized novel views with ground-truth dense correspondences, generated in Stage 2.1:
These demonstrate both the geometric diversity and high-quality pixel-level correspondence labels used for supervision.
For novel-view inpainting, we also provide a better inpainting model fine-tuned from Stable-Diffusion-2.0-Inpainting:
from diffusers import StableDiffusionInpaintPipeline
import torch
from diffusers.utils import load_image, make_image_grid
import PIL
model_path = "Liangyingping/Lift3Dreamer"
pipe = StableDiffusionInpaintPipeline.from_pretrained(
model_path, torch_dtype=torch.float16
)
pipe.to("cuda")
init_image = load_image("assets/debug_masked_image.png")
mask_image = load_image("assets/debug_mask.png")
W, H = init_image.size
prompt = "a photo of a person"
image = pipe(
prompt=prompt,
image=init_image,
mask_image=mask_image,
h=512, w=512
).images[0].resize((W, H))
print(image.size, init_image.size)
image2save = make_image_grid([init_image, mask_image, image], rows=1, cols=3)
image2save.save("image2save_ours.png")
If you use this inpainting model, please cite our paper:
@article{LIANG2026,
title = {Lift3Dreamer: Boosting Text-Driven Novel View Synthesis via Lifted 3D Inpainting Model from Single Images.},
journal = {Fundamental Research},
year = {2026},
issn = {2667-3258},
author = {Yingping Liang and Ying Fu and Jiaming Liu and Debing Zhang},
}
To improve feature robustness under varying lighting conditions, we apply a physics-inspired relighting pipeline to the synthesized 3D scenes.
Run the following to generate relit image pairs for training the decoder:
python relight.py
All outputs will be saved under the configured output directory, including original view, novel views, and their camera metrics with dense depth.
If desired, you can run sky_seg.py to mask out sky regions, which are typically textureless and not useful for matching. This can help reduce noise and focus training on geometrically meaningful regions.
python sky_seg.py
We build upon recent advances in ROMA, GIM, MINIMA, and FiT3D.
@inproceedings{Liang2025L2M,
author = {Yingping Liang and Yutao Hu and Wenqi Shao and Ying Fu},
title = {Learning Dense Feature Matching via Lifting Single 2D Image to 3D Space},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
year = {2025},
pages = {6621--6631}
}