Dataset was 41 images, most over 1024 pixels in at least one dimension.

This is the first LoRA I've released after Civitai limited the batch size on their online trainer. It means only two images are being compared at once, not four. This does appear to degrade quality.

I do think it's a lot better than a 'very vague resemblance' though, otherwise I wouldn't have released it.

@beardfine100274 Thank you for the kind words and I'm glad this LoRA's working well for you. Unfortunately I'm really struggling to get new LoRAs to be as good as those I've made before. I've done a lot of experimentation to make up for the reduced batch size and none of it works well enough. The batch size restriction seems to be a cost-cutting exercise on Civitai's part, it likely requires a lot more vRAM to compare four 1024x1024 images at once than just two.

I've been able to get good likenesses again by training for a lot longer, but it then over-trains and pulls out necklaces, jewellery etc, even the photographers' styles. Finding the sweet spot will take a lot more experimentation than I can reasonably afford - it used to take 3,000-3,500 Buzz for a LoRA, my experiments have taken tens of thousands just to end up with worse results!

I've taken the decision to put training on-hold for now, until either Civitai walk the change back or I can find another way to train well.

I haven't given up for good, but for the time being I'm going to concentrate on learning Comfy/Hunyuan and creating myself some nice videos!

Thanks, but I've paused making LoRAs for the time being. This LoRA should have reasonable support for NSFW though (within the usual limitations of Flux).

I may be wrong though - this one had to be made with different parameters to my usual settings, so it may be missing something.