|
|
|
|
|
|
by ausbitbank
833 days ago
|
|
|
Can you share some of your process, like is a full custom sdxl finetune really required for high quality results ?
Did you experiment with IPA face adapters or other techniques ?
I hope this works out for you, its a great niche good luck :) |
|
|
It's a proper fine tune (LoRA, not "full"). I experimented with IPA faceID and a few others but got poor results. Especially when training characters that don't look like celebrities. It's really obvious when you train it on yourself vs a stranger. And in this case the character is the customer, so that matters. As an aside this whole space of "custom fine tunes" will eventually get hurt by people promoting IPA products, which perform worse, but are 100x cheaper than real fine tunes. That is, until IPA-style tricks get better. (So, 3 months from now?)
Here's what I learned after lots of experimenting: stick to koyha's defaults. I wasted a lot of time on custom pipelines to normalize training images. I also wasted a lot of time fixing Replicate's SDXL cog, which I now know is more of a proof of concept than something people are supposed to actually use.
Don't use reg images (not necessary in this case), use a "celebrity doppelgänger" tool to train an existing token instead of a rare token, and, importantly, give folks a tool for spot correction (inpainting) to e.g. fix hands. Pretty much every image has some AI artifact but you can cycle through a few corrections to fix it.
Not sure if you're building in this space but I'll add: try paying for a few of the headshot generators out there. They're a bit disappointing. I assumed they had all "cracked it" so I kept working until I had a decent process, and the result is arguably better than the current leading products (which mostly use upscaled & face-corrected SD1.5).
On the market and product – this landing page is converting paid traffic very poorly and has had zero sales, so it's impossible to predict which ideas resonate, even if the tech is interesting!