Player not loading? Watch on YouTube
This PhotoMaker tutorial, uploaded on January 21, 2024, covers a local Windows installation and tests portraits generated from reference photos. It identifies no numbered software version; the setup instructions and interface behavior describe the configuration shown at that time.
The presenter points Windows users to the bmaltais fork after reporting trouble running the original TencentARC repository. He stresses using the dependency versions specified in its instructions, though the transcript does not give those version numbers. The demonstrated Gradio interface accepts reference images, a prompt and a style selection. Advanced controls include a negative prompt, steps, style strength, guidance scale and seed.
For this local AI workflow, the presenter says the IMG trigger must appear exactly once in the prompt, alongside a subject term such as man or woman. His test uses 1024 by 1024 output. On his RTX 4080, he reports roughly 25 to 30 seconds for two images after a slower first batch; requesting four causes a CUDA out-of-memory error in his setup.
The examples show his main reservation: he finds the generated face consistent across photographic and stylized outputs, but not identical to the reference person. These are observations from his tests. He also demonstrates selecting each image and clicking download, because this configuration has no output directory.
