Dua Lipa is an English/Albanian singer who needs little introduction, the winner of numerous Grammy, MTV, Brit Awards, etc., and one of the greatest stars in the Pop music industry nowadays.
130-step TI trained on a dataset of 15 images with these settings.
Appreciate my work? My TIs are free, but you can always buy me a coffee. :)
Curious about my work process? I have summarized it here.
Building a good prompt with my TIs
You're obviously free to experiment, but bear in mind that my TIs are trained with a more or less fixed phrasing, that normally starts with:
"photo of EMBEDDING_NAME, a woman"
So I recommend always starting your prompt like that and then building the rest of the prompt from there. For instance, "photo of extremely sexy (2ual1pa:0.99), a woman as a sexy student, (closeup portrait), bob haircut, ({white|yellow|pink|blue} tight long sleeve turtleneck top), at a cantina sitting bar (masterpiece:1.5) (photorealistic:1.1) (bokeh) (best quality) (detailed skin texture pores hairs:1.1) (intricate) (8k) (HDR) (wallpaper) (cinematic lighting) (sharp focus)"
Description
130-step TI trained on a dataset of 15 images with these settings.
FAQ
Comments (5)
Hi, if I may ask, why do you use a high gradient accumulation? Is it to reduce noise or prevent overfitness?
Basically, because so far the results have been better than with my previous settings (gradient 2, everything else remained more or less the same). I think the results are better partly because I have more control over the "sweet zone" of my TIs, which usually starts around 350 epochs (about step 120) and begins deteriorating around 500 epochs (more or less after step 170 or so). Overall, it takes me less time to train and I usually obtain better results, so it's been clearly a step forward for me.
@JernauGurgeh Thanks for the quick answer. Do you explain somewhere how you exercise that control, i.e. what tweaking you do during training?
@RetroGlamourAI No tweaking at all, really. I know where to expect good results out of experience... been testing steps for weeks now, and I know that between 120 and 180 I can expect the best results. I'll occasionally test some previous and later steps, but 99.99% of the times there will be no need to, as I'll find the exact step where the TI is closer to the subject among those I mentioned.
@JernauGurgeh Thank you very much for the explanation.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.