Posts

I think I’m qualified to comment on this subject as I’ve made over 200 MLP character LoRAs for SDXL-based models (mostly Illustrious, though I started with PDXL 6 and have recently shifted to NoobAI).

One thing I have found is that a small, varied and well-tagged dataset is far better than a large one with just passable tagging and not that much variety. I generally use around 30 distinct images of the character (a minimum of around a dozen pics and a maximum of around 50) when training my LoRAs.

After I gather the images for the dataset, I run them all through an autotagger and then touch up its results (adding a tag for the character, removing tags that don’t apply, adding missing ones).

I think I have the capacity to train LoRAs for Flux.2, but I have no clue how to tag images for one (at the moment), so I haven’t done anything of that sort yet.

I’ve been gaining experience with PonyDiffusion and enjoying the easy knowledge it has of MLP characters, but run into some problems with fringe characters it has misconceptions of (characters are the wrong gender, have 90% correct features but random horns or other features, presented in biases acitons or poses, etc). I sometimes find pre-existing LORA’s that help with this, but sometiems the LORA’s themselves cause similar hiccups or interact with existing projects in a negative way.

I’ve recently started dabbling with FLUX2, and while it seems much more useful in creative direction and revision, it really seems to have no concepts of the pony world. Which gets me considering getting into LORA creation. There must be some potential here, seeing how impressive the PonyDiffusion project’s results are for the fandom as a whole and how little the base models seem to know.

I’m under the impression that making a LORA is simply a matter of organizing samples, and relativly trivial to run if you are willing to use a cloud serivce?

I’ve seen some others on here referencing their own LORA creation, and wondering what your experiences have been like. Is this an unrealistic thing to consider doing solo for specific characters, or do any of you find it useful? Is it relatively simple on a base model like SDXL/PXL, and unrealistic on a pony-naive model like Flux?

@Cypher
Well… that’s sucks. I liked these LoRAs… I can’t gen locally so it’s quite a blow for me.

Civit, Channel’s gone. Who cares about the 800 followers, I’m a nobody anyway

@Cypher
On Civit? Or here on Tantabus?

Here on Tantabus, you’ll need to have an administrator do it. On Civitai, you’ll need to go to the bottom of the account settings page (account settings is the gear at the bottom of the top-right dropdown menu)

How do I delete my account?

Chapter 11 added.

@Thoryn
Glad I’m not the only one. My uploads have slowed dramatically.

@Equum Unfortunately, those are limitations of CLIP, which is what a lot of older models use, including PonyV6, Illustrious, and basically almost every model that competently renders ponies. One of the major limitations of CLIP is that it doesn’t really comprehend spatial relationships, which means it really only can do spatial relationships when the specific spatial relationship you’re prompting exists a lot in the training data. It’s not capable of dynamically putting together a new spatial relationship without a LoRA to assist.

Newer models have a much better understanding of spatial relationships (and most of them are also natural-language capable) which makes them much more useful for prompts like what you’re describing. Unfortunately, most of them suck at general pony knowledge; some of them might kinda vaguely get major characters like Twilight or Celestia right, but many of them can’t even comprehend the CMC. They’re also generally heavier models that require better hardware just to get tolerable generation times.

@Equum
No, sadly haven’t had the time to play around with image generation that much the past year because of IRL things.

@Thoryn
Did you end up making progress on this? Settle on any specific workflows?
I’m basically reliving this experience you’re having as I try to get up to speed on all this image generation stuff after stumbling into it in 2026. Lots of the same issues here.

Step 10 is super useful. Use that a lot.

Thanks for the feedback. I also added chapter 10.

Interesting, thanks.

Thank you for this!

@🐴
It’s already the year of Linux gaming, thanks to Valve. Maybe laptop manufacturers will finally open up their eyes and see the world from where I stand.

@Background Pony #E6E5
Year of the GNU+Linux laptop, baby!

Maybe they would if they weren’t chained to Windows.

@Background Pony #5272
I’ve been a Mac addict ever since I first discovered the 2008-era touchpads on the MacBook Pros. There’s been no other laptop that’s been worth a d~~n ever since. PC manufacturers don’t bother with multi-touch gestures.

@🐴
I am very happy with mine. It is certainly doing the job.

Who else enjoys their Macs and the unified memory of the M-series Apple silicon chips?

Any macOS-specific gripes or tips to share?

I’m just a regular horse-fucker. It was fated since my birth.

@Cypher
I would give them a try but Civitai has blocked my country. Is there another site these Loras are backed up on?

1.Mage-Meadowbrook https://civitai.com/models/2543185/my-little-ponymlp-g4-mage-meadowbrook 2.Somnambula https://civitai.com/models/2365757/my-little-ponymlp-g4-somnambula Nobody’s going to use it anyway

Well, I guess any web based generator has it’s own limitations.

Thank goodness even an old 10x series 8gb card allows decent quality image generation.