SC-FEGAN ships inference only, pins a TensorFlow 1.13 release candidate, and asks for a checkpoint directory that is not in the repository

requirements.txt pins a release candidate of TensorFlow 1.13

The dependency list is not a loose set of ranges. Every entry is pinned, and one of the pins is a release candidate:

absl-py==0.7.0
astor==0.7.1
gast==0.2.2
grpcio==1.18.0
h5py==2.9.0
Keras-Applications==1.0.7
Keras-Preprocessing==1.0.9
Markdown==3.0.1
mock==2.0.0
numpy==1.16.1
opencv-python==4.0.0.21
pbr==5.1.2
protobuf==3.6.1
PyQt5==5.12
PyQt5-sip==4.19.14
pyyaml>=4.2b1
six==1.12.0
tensorboard==1.12.2
tensorflow==1.13.0rc2

`tensorflow==1.13.0rc2` is the line that decides everything else. A release candidate is not a stable build, and TensorFlow 1.13 belongs to a series that was superseded years ago, so the interpreter it runs on is correspondingly old as well. `numpy==1.16.1`, `h5py==2.9.0` and `protobuf==3.6.1` come from the same period.

The one unpinned entry is `pyyaml>=4.2b1`, and the reason it matters is that `demo.yaml` is the only configuration surface in the project. Everything else is frozen; the file that decides which GPU the process uses is the one thing that can drift.

The README's own dependency list is much shorter, naming tensorflow, numpy, Python3, PyQt5, opencv-python and pyyaml, which is enough to install but not enough to reproduce.

Setup assumes a checkpoint directory outside the repository that nothing creates

The model is not in the repository. The first instruction is to download it from a Google Drive file id, and the next two commands are:

mv /${HOME}/SC-FEGAN.ckpt.* /${HOME}/ckpt/
python3 demo.py

Read the paths carefully. The checkpoint is assumed to be sitting directly in your home directory, and it is moved into `${HOME}/ckpt/`. That destination is not part of this repository. The top-level entries are .gitignore, LICENSE, README.md, demo.py, demo.yaml, imgs/, model.py, ops.py, requirements.txt, samples/, ui/ and utils/, and there is no ckpt directory among them. So the move either works because you happened to have that directory already, or it fails.

The repository is not archived and the last push to master is dated 2024-05-20, but there are no GitHub releases and no version marker anywhere in the tree, so there is nothing to compare that Drive file against. The archive is the only source of weights and the file is not in the tree beside them.

The development environment is stated in one line in the notes: Linux running Ubuntu 18.04.1.

Training code is listed as future work, so the repository is the demo only

The Next section of the README contains exactly one item: update training code.

That is the whole of it, and it is the most important line in the file. What is here is the inference path, not the training path. There is no data pipeline, no loss configuration, no dataset loader, and no way to retrain or fine-tune the network from what is in the tree.

The Python surface is correspondingly small. `demo.py` is the entry point the setup instructions run, `model.py` holds the network, `ops.py` the operations, with `ui/` and `utils/` as the supporting directories. Four files carry a model that the paper describes as an SN-patchGAN discriminator with a Unet-like generator using gated convolutional layers.

What the model does is narrower than the paper title suggests for a practical purpose. It does not synthesise a face from nothing. The README is explicit that only erased regions of the image are filled in by the network, which means you supply an existing face image, a sketch and colour strokes, and a mask, and the network fills inside that mask. The Example Results headings also cover edit earring and face restoration, including a variant using only sketch and colour.

The acknowledgement section credits the official DeepFillv1 inpainting code, which is where the inpainting half of this comes from.

The licence field says nothing, the README says CC BY-NC 4.0, and both restrict you to research

Three sources describe the licence and they do not line up.

The repository metadata field for licence records nothing, the NOASSERTION value that means GitHub's licence detector found nothing it could name. The README states a licence in its own section: CC 4.0 Attribution-NonCommercial International. And a LICENSE file is present at the root of the tree, which the README does not describe.

So there is a file, a stated licence, and an undetected metadata field. Without opening the LICENSE file you cannot tell whether it matches the CC BY-NC 4.0 the README names, and the metadata suggests the detector did not recognise it as a standard licence either way.

The non-commercial part is the term that matters. The README repeats it twice, once as the software being for educational and academic research purpose only, and again in the notes, where the provided model and sample code are described as being under a non-commercial creative commons licence.

Read together, that is a research code release rather than an unrestricted one: the weights and the code both carry the restriction, so a commercial product is out of scope on the repository's own terms regardless of what the CC label alone implies.

Save Img writes one fixed filename, so only the last result survives

The demo is a PyQt5 window with nine buttons, and the editing model is mask-limited throughout.

`Open Image` loads the face image. `Mask` draws the region to be filled on the left viewer. `Sketches` draws lines. `Color` draws coloured lines, and the first time you press it you have to pick a colour from the palette. `Palette` changes the current colour, after which you press `Color` again to apply it. `Undo` steps back. `Complete` runs the network and shows the result on the right viewer. `Arrange` arranges the editing works.

`Save Img` is where the file gets interesting: the result is saved as name.jpg, a fixed name rather than a chosen one. Nothing in the file describes incrementing it or prompting for a path, so saving a second edit overwrites the first.

The recommended workflow is ordered, and the order is not obvious from the button order. You draw the sketch first, referring to the original image. Then you draw the mask over the sketched region. Then you press `Arrange`. Only then do you draw colour inside the masked region, and finally press `Complete`.

The mask is what keeps the edit local. Only erased regions are filled in by the network, so the strokes outside the mask are not model input at all.

One GPU or the CPU, chosen by commenting out a number in demo.yaml

Compute selection is a text edit rather than a flag or an environment variable. The README says to select the number of GPUs by editing `demo.yaml`, and that multi-GPUs are not supported:

GPU_NUM: 1 (the number you want to use)
#GPU_NUM: (if you want to use only CPU, erase the number)

The mechanism is the second line. Forcing CPU-only means commenting out the number rather than setting it to zero, so the state of the file is decided by whether a `#` is present. There is no third option, no mixed precision setting, and no memory cap.

This is the only configuration file in the repository, and it is the one dependency that is not version-pinned. Everything in requirements.txt is frozen to an exact version, including the TensorFlow release candidate, while the file that decides whether the process uses one GPU or none stays open.

For a demo that loads a checkpoint and paints a GUI window, that is a reasonable surface. For anything you would run repeatedly, it is the piece to move into code.

Four Python files, five sample images, and a citation for a paper you should read

The tree is small enough to read in full: .gitignore, LICENSE, README.md, demo.py, demo.yaml, imgs/, model.py, ops.py, requirements.txt, samples/, ui/ and utils/.

`imgs/` and `samples/` are separate directories. The samples visible in the listing are five PNGs, numbered 00001, 00011, 00018, 00221 and 00225, which are the inputs the demo expects to find. Whether the GUI opens them by browse or by index is not stated.

The README closes with a BibTeX entry for the ICCV 2019 paper, by Youngjoo Jo and Jongyoul Park, published in October 2019, with an arXiv link at 1902.06838. The citation is the better half of the documentation: the README is 349 words and describes how to run a demo, while the paper describes what the network does.

The acknowledgements thank researchers at ETRI by name and credit the official DeepFillv1 code, which places this work in the lineage of learned inpainting rather than in generative face synthesis from scratch.

Editorial conclusion

SC-FEGAN is a readable reference implementation if you want to understand one 2019 face editing network, and it is not a tool to build on. The interpreter it needs is from the TensorFlow 1 era, the checkpoint lives on a Google Drive link rather than in the repository, there is no training code to adapt, and the licence restricts use to educational and academic research. Check those four things before anything else, because the README is candid about all of them and the licence text in particular is narrower than the word free might suggest.

Frequently asked questions

Official sources

1. Official README
2. Project repository
3. README
4. Releases