Resolution-robust Large Mask Inpainting with Fourier Convolutions

By Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park and 1 more. 2021

1,391 citations recorded by the source corpus.

Modern image inpainting systems, despite the significant progress, often struggle with large missing areas, complex geometric structures, and high-resolution images. We find that one of the main reasons for that is the lack of an effective receptive field in both the inpainting network and the loss function. To alleviate this issue, we propose a new method called large mask inpainting (LaMa). LaMa is based on i) a new inpainting network architecture that uses fast Fourier convolutions (FFCs), which have the image-wide receptive field; ii) a high receptive field perceptual loss; iii) large training masks, which unlocks the potential of the first two components. Our inpainting network improves the state-of-the-art across a range of datasets and achieves excellent performance even in challenging scenarios, e.g. completion of periodic structures. Our model generalizes surprisingly well to resolutions that are higher than those seen at train time, and achieves this at lower parameter&time costs than the competitive baselines. The code is available at https://github.com/saic-mdal/lama.

Subfields: Computer Vision and Pattern Recognition, Generative Models

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Generative Models

Citations in the corpus

Fine-Grained Egocentric Hand-Object Segmentation: Dataset, Model, and Applications (2022, methodology): “The second way is to remove the existing handobject using an image inpainting model [65] and the labeled segmentation masks, as shown in the top right of Fig.”
Make-An-Audio: Text-To-Audio Generation with Prompt-Enhanced Diffusion Models (2023, methodology): “For audio-inpainting, we compare different masking designs, including the irregular (thick, medium, and thin) strategy from visual world (Suvorov et al., 2022), as well as the frame-based (with varying p ) strategy commonly used in speech (Baevski et al., 2020; Hsu et al., 2021).”
RePaint: Inpainting using Denoising Diffusion Probabilistic Models (2022, methodology): “We use the ImageNet validation set with 32 images over the LaMa [39] Wide mask setting.”
Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks (2024, methodology): “Furthermore, it can be integrated with models like LaMa [56] and PaintByExample [68] to achieve precise image erasure and customized image editing.”
SPHINX: The Joint Mixing of Weights, Tasks, and Visual Embeddings for Multi-modal Large Language Models (2023, methodology): “This setting integrates our SPHINX , SAM, LaMa, and Stable Diffusion to achieve language-driven image inpainting and editing.”
Gaussian Grouping: Segment and Edit Anything in 3D Scenes (2023, methodology): “After deleting the Gaussians belong to the bear, we detect the image hole region with no Gaussians projection covering and then use LAMA [41] to produce single view inpainting result to guide the learning for newly added inpainting Gaussians.”
Fast High-Resolution Image Synthesis with Latent Adversarial Diffusion Distillation (2024, positive): “Our baselines are LaMa [59] and SD1.”
Improving Diffusion Models for Inverse Problems using Manifold Constraints (2022, methodology): “Score-SDE [41] 38.68 RePAINT [32] 247.6 DDRM [25] 2.117 LaMa [43] 0.629 AOT-GAN [52] 0.082 ICT [49] 144.6 DSI [35] 36.64 IAGAN [20] 518.47 Ours 81.59 Ablation studies We perform three ablation studies: 1) As both MCG term and the projection term contain information about the measurement y , we…”
FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space (2025, neutral): “Generative inpainting systems such as LaMa [54], Latent Diffusion inpainting [42], RePaint [33], Stable Diffusion Inpainting variants 2 , and FLUX.1 Fill 3 make such context-aware edits instantaneous; see also Palette [45] and Paint-by-Example [57].”
Paint by Example: Exemplar-based Image Editing with Diffusion Models (2022, methodology): “Considering it can only fuse a foreground image to background, we first use an unconditional image inpainting model LAMA [61] to inpaint the whole mask region, then extract the foreground of reference image through an additional semantic mask, and finally harmonize it into the source image.”
MAT: Mask-Aware Transformer for Large Hole Image Inpainting (2022, methodology): “It is noted that we only use a subset Places365-Standard (1.8M images) to train our model, much fewer than CoModGAN [75] (8M images) and Big LaMa [51] (4.5M images).”
DiffIR: Efficient Diffusion Model for Image Restoration (2023, methodology): “We train our DiffIR for inpainting using the same loss functions of LaMa [12], which further introduce multiple * Corresponding Author perceptual losses and adversarial loss to the basic L 1 loss.”
VACE: All-in-One Video Creation and Editing (2025, neutral): “Augmentation of the masks [62] allows for unconditional inpainting.”
Inpaint Anything: Segment Anything Meets Image Inpainting (2023, methodology): “Finally, a state-of-the-art inpainting model, such as LaMa [13], is used to ﬁll the hole created by the removed object using the mask.”
Imagen Editor and EditBench: Advancing and Evaluating Text-Guided Image Inpainting (2022, positive): “In terms of text-image alignment, Imagen Editor trained with object-masking is preferred in 68% of comparisons with its counterpart conﬁguration trained with random masking (a commonly adopted method [27, 35, 41]).”
V*: Guided Visual Search as a Core Mechanism in Multimodal LLMs (2023, methodology): “Then we use the image inpainting model LaMa [45] to erase all the mentioned objects in the corresponding images and re-evaluate the InstructBLIP model with the modified images and only keep the questions that can not be correctly answered after this modification.”