Do the execution ...)
Double-check all your parameters, then click the Run button on this job (if using “Run all” for the first time, this will happen automatically). Remember if making changes from a prior trial, you should re-run the Parameters job to validate the inputs before running this one.
After a moment for initialization, you’ll see a muddy rectangle (if “initial_image” was blank), or your starter image. The VQGAN model will then iteratively refine this mud/image toward the situation you described (and/or the target_image provided).
Have patience. At the default 480x480 size, it’ll take about one second per iteration, and at the default 50 interval you’ll get about one visual update per minute. The first couple of updates will be vague nonsense, but with each successive update you’ll see the scene coming into focus. If you don’t like where it’s headed, stop the job, edit any parameters, then run the Parameters job and the Execute job to begin again.
There’s a certain point where a scene will “congeal” and further iterations make little difference. This will vary among the different models, parameters and just random chance…but very occasionally it will surprise you with a sudden shift in overall color or texture. With the imagenet_16384 model, I typically set a limit of 800 iterations, but the scene is usually established about halfway through and I might stop it earlier.
Accompanying each preview image you’ll see a “loss” value which gradually decreases with each successive image. In theory, a value under 0.5 means the discriminator finds the scene “plausible.” However, the image-to-image change tends to decelerate and good chance won’t cross that threshold. Also, I’ve only got so many hours in my life and can’t afford to wait forever, I just want to see weird distorted cat pictures, y’know?
Another good reason for limiting the size and number of iterations is that session time and system resources are limited, especially in the free mode. You might squeeze our three or four attempts in a 24 hour period, then have to wait for the following day to experiment more.
Genera un vídeo con los resultados (Generate a video with the results)
If you just want a still image (or a few along the way), you can right-click any of the in-progress execution images and “Save as…” This a PNG image file, you’ll need other software if you want to convert to JPEG.
For video, a couple more jobs must be run. If “Run all” was selected, these last steps will happen automatically once the Execute job finishes (iteration limit reached) or is cancelled in-progress.
The next job below “Execute” is “Generate a video…”
When run, this will coalesce every image iteration (not just the preview images, but every frame in-between) into a MP4 video file.
It takes a couple minutes to complete, and there’s a progress bar while it works. Once the conversion is complete, the bar turns green.
Next there’s a “View video in browser” job, but I find it’s generally not worth the effort. The preview images along the way give a pretty good impression whether the results are worth keeping. Instead, skip down to the last job on the page, “Descargar vídeo / Download video.”
Just click the “Run job” button and this will download to your computer, wherever the browser preferences deposit downloaded files.
The first time you run this in a session, you might get a confirmation dialog to allow downloaded files. That’s fine, just the browser being vigilant. It’s easy to miss while distracted with all the fun graphics though. Allow downloads and be on your way.
Keep in mind: if you run another “Execute” job, you’ll need to re-run “Generate a video” prior to downloading, otherwise you’ll download your prior video. The MP4 isn’t produced automatically each time.