This ai-toolkit fork has some nice QoL improvements, training folks are gonna love it.

Sharing a fork of ai-toolkit I’ve been using lately. The author did a ton of VRAM optimization, picked all those low-hanging fruits the main branch missed. More usefully, they moved a bunch of options that used to be only in config files to the UI, like Prodigy optimizer, DoRA support, and those old-school slider LoRA training that doesn’t need a dataset.

A few things in daily use really hit the spot: each row in the training queue table now shows the first image from the dataset, so you know at a glance what that task is training on. You can also drag and drop to reorder the queue. Best of all, you can change sampling generation settings without stopping training—used to be you had to interrupt just to tweak a sampler, which was annoying.

Also, if you have multiple machines with GPUs on your local network, sampling generation can be offloaded to another ComfyUI instance running in parallel, so it doesn’t eat into your training machine’s resources. The catch is that machine needs direct filesystem access to the training machine.

My question: if a single machine has enough RAM, would swapping models in and out for sampling slow down the main training?

The queue showing the first image is such a thoughtful little detail.

You don’t have to stop training to change the sampler—that alone makes me want to switch.

Finally, they moved those options from the config file to the UI.

Multi-machine distributed sampling is great for people with GPU farms, but I can’t even use it with my single card.

Swapping models in and out locally will definitely eat up some bandwidth, slows things down a bit.

DoRA being added to the UI saves a ton of time messing with configs.

Got a link?