Skip to content

LaMa (Large Mask Inpainting)

Feed-forward inpainting model using Fast Fourier Convolution (FFC) for image-wide receptive field from the first layer. Excels at filling large masks with coherent textures. Resolution-robust: trained at 256x256, infers up to ~2000x2000.

Paper: WACV 2022. Authors: Samsung AI Moscow / AIRI. arXiv:2109.07161.

Architecture

U-Net-like fully-convolutional network: 

Input (4ch: masked_image + mask) → Encoder → 9-18 FFC Residual Blocks → Decoder → Output (3ch)

Fast Fourier Convolution (FFC)

The key innovation. Each FFC block splits channels into two parallel branches:

Feature map
    ├─ Local branch (standard convolutions) → high-frequency local details
    └─ Global branch:
         → channel-wise Real FFT (spatial → frequency domain)
         → 1×1 conv in frequency domain
         → inverse Real FFT (frequency → spatial domain)
         → image-wide receptive field from layer 1

Why this matters for inpainting: standard CNNs need many layers to propagate information across large masked gaps. FFC's spectral branch covers the entire spatial extent in a single layer — information from opposite sides of the mask is immediately available.

Variants

Variant Params Size Training Data
Big-LaMa ~51M ~410 MB Places365-Challenge
Standard LaMa ~27M smaller Places365-Standard
CelebA variant ~27M smaller CelebA-HQ (faces)

Training

  • Resolution: 256×256
  • Masks: on-the-fly random polygonal chains + rectangles (deliberately large)
  • Loss: adversarial (non-saturating) + feature matching + High Receptive Field perceptual loss (ResNet50-dilated)
  • Discriminator: patch-level, "fake" labels only for mask-intersecting areas

Performance

  • VRAM: 2-4 GB at 512×512 — runs on consumer GPUs
  • Speed: ~2s GPU (HD), ~25s CPU, 26-45ms mobile (Qualcomm NPU)
  • Resolution generalization: trains 256px → infers to ~2000px without retraining (FFC property)
  • 20% slower but 3-4x fewer params than competing baselines

Relation to Diffusion-Based Inpainting

LaMa is a feed-forward model (single pass, deterministic). Diffusion-based inpainting (FLUX Kontext, Step1X Edit) is iterative (20-50 steps) but better at semantic content (faces, complex objects). LaMa excels at textures and patterns.

Practical combination: use LaMa for fast background/texture inpainting, diffusion for semantic regions.

License

Apache 2.0 — fully commercial.

  • GitHub: github.com/advimman/lama
  • Widely adopted: IOPaint, cleanup.pictures, ComfyUI nodes