New Feature?!: Some on-site trainer enhancements.

The training pages have a notice which mentions that "We are addressing some issues with recently deployed training interface enhancements.", but doesn't say what and this isn't given an announcement (beyond the start of year overview mentioning enhancements are coming).

Import image from generator or profile as training data

Starting off very poorly: Why is this thing? This is a bad idea. Using synthetic data for training only works when a user knows exactly what they're doing, and even then produces strange results (the only ones I know have synthetic data and seen work well at all are some poses/situations with distinctly comedic and unrealistic tones that make it easy to ignore any issues). Having an option that implicitly encourages doing something very unwise doesn't seem like a good idea.

Re-Use a Data Set

Now for some good stuff. This is handy for training an alternate base model version (without doing multi-training because you wanted to make sure it worked before committing) or a v2 without forcing you to upload files the site already has. Doesn't appear to work right now though (or is showing data from way older than 30 days).

Zoom in

YES! Finally! This corrects one of my oldest gripes with the on-site trainer: That it's impossible to see the image you're tagging in full short of looking for the original in your data set, copy/pasting into an image editor, or (only applicable to images you uploaded piecemeal, see below) open in new tab. Now you can actually see what you're working on! Should have been added long ago.

Images from compressed file (e.g., zip) now work when opened in new tab

This isn't even really needed now that zoom in is a thing, but it's nice to finally see that bug quashed even if it was only really a problem because we didn't have the above.

Epoch pictures prompts no longer labeled as optional

Edit: Realized this one on a second pass. The description for the fields to put prompts for epoch pics has been rewritten. It's effectively the same, but is far better because it doesn't push a user to a feature that is (was?) totally and completely broken. You should still

What do we still need?

This old article of mine is largely still applicable past the first two (and maybe the randomized order, which may have been fixed some time ago but I'm not sure because I've gotten good enough at tagging I can do it in a single session now). Also the "Acknowledgement" you must agree to before training, which post-dates the article, really needs its wording tweaked: It completely breaks down with real people being incidentally in a small chunk of training data for a LoRA that isn't meant to reproduce them (at the most extreme case, its applying to a random guy deep in the background of photo of a real world location who is absent from the rest of the data)

Edit: I just noticed collections is an option on search now. Is that new? I don't think it was there before, and the total inability to find collections has been something I've openly complained about for a while I think I'd have noticed.