I wrote this article about training Flux LoRAs with just 1 image.
https://civitai.com/articles/7618
While the technique isn't foolproof, it can produce good enough results, depending on the needs you have for the model.
Using the Flux training script on CivitAI right now has a minimum cost of 2000 buzz, and this remains the cost even if you have 1000 images, or increase the training resolution, rank and total steps, until it starts increasing at some point.
But with Flux, it may not be necessary to have all those settings so high.
I don't know if the buzz costs could be brought down at all, but that is the end goal to allow people to make smaller LoRAs, test-models, or models which can be used to generate a stronger dataset.
I'm proposing a "Flux Lite"-training configuration, which has much stricter restrictions, but allows for cheaper training. With these limitations, the training time would also go down. And perhaps there could be a "lower priority queue" that these models go into to cut costs more.
I would leave CivitAI to figure out the actual details of the limitations, but my starting point would perhaps be:
Maximum 10 images
At most 500 steps
At most 10 epochs
At most rank 16
Limit the number of preview images to 1 per epoch to save on training inference cost
Support masked training (transparent textures or black/white mask images)
If the training a model steps would be slightly simpler, I think we can lower the barrier of entry for people to just try to make a model. The more that can enjoy training models the better.
The masked training would be a great addition regardless of this feature.
