DeFLOCNet
Deep Image Editing via Flexible Low-Level Controls 路 CVPR 2021
Paper 路 Live demo 路 GitHub 路 Training
Selected outputs from both released checkpoints. White marks the hole. These examples use the mask alone. Image credits.
DeFLOCNet fills masked regions at 256 脳 256 RGB and accepts optional sketch and color guidance. Choose the checkpoint that fits your image.
Checkpoints
| Checkpoint | Intended images | Training data | Step |
|---|---|---|---|
| HDTF Portrait | Faces and close portraits | Dedicated HDTF face crops | 10,000 |
| General Scenes | Landscapes, buildings, and everyday images | Mixed images including Places2, Unsplash, T2I, and a small HDTF component | 23,500 |
Each file is an FP32 generator-only inference bundle of about 323 MiB. See the local inference guide for input formats and commands.
Edit with sketch and color
In DeFLOCNet Studio, paint the edit region in white. Add a sketch for structure or sparse color marks inside the hole. Compare both uses the same image, mask, hints, and seed for the two checkpoints. The completion copies pixels outside the mask from the prepared image.
Both outputs use the same crop, mask, seed, and General Scenes checkpoint. The black arc and brown dots are the supplied sketch and color hints.
Architecture
Both checkpoints use sd_color_v2, a five-level residual U-Net that completes
the image in one forward pass. Separate pyramids encode the sketch and the
supplied color pixels. Sketch features gate the color state as it propagates
from coarse to fine. See the architecture notes
and the original CVPR 2021 code.
Checksums
| File | SHA-256 |
|---|---|
deflocnet-hdtf-step10000.pt |
2de135908683ca9c3a882de794efebc103e34101a6e813b3348c662007ed1bb0 |
deflocnet-general-step23500.pt |
bab2bf7eedbf6f905895e40be3731ee499d3f485ff7546071ebb90041fc17806 |

