The young version of Dua Lipa

I’m finally ready to release V2!

I trained it with a resolution of 1024 for 1500 steps with about thirty images, for a total cost of 2678 buzz. I didn’t use any descriptions, was that a mistake? I haven’t really found an answer—some say it's pointless, while others say it’s necessary. Sometimes I feel like it’s all just luck ^^".

This model is much closer to the young Dua Lipa I was aiming for, though it’s still not perfect. I’d love to make a V3, but I don’t have enough buzz anymore, so don’t expect it anytime soon.

I’ve also tried all the combinations of simpler and scheduler, but again, I haven’t found a clear winner. For me, it still feels like it’s just luck.

How do you make the most of LoRA?

As long as the framing doesn’t exceed a mid-shot, the image quality will be great with any image size. However, if you want to generate a full-body shot, the head might turn out poorly. To solve this, I recommend generating images at 2.5 megapixels for such framing. For the optimal resolution for each image ratio, I suggest installing the Flux Calculator: https://github.com/gseth/ControlAltAI-Nodes

Txt2Img

I no longer use CLIP at all in my workflow, I simply have a text input and guidance that I keep at 3.5. I’ve run several tests, sometimes going up to 8, but again, depending on the prompt, it all feels like luck to me.

The first technique is simply to type "a young woman" and you’ll get a very high-quality portrait. That’s usually enough if your goal is to have a portrait of a young Dua Lipa.

To go further, I use JoyCaption: https://huggingface.co/spaces/fancyfeast/joy-caption-pre-alpha Drag in an image, and it’ll generate a detailed description. From that description, I ask ChatGPT to add the following list of words:

with lighting that accentuates her facial features and skin texture, symmetrical, well-proportioned face, detailed eyes with realistic light reflections, long dark brown hair, oval-shaped head, full lips

It’s a technique that works well. However, Flux is terrible for generating images in a highly futuristic space setting or more complex environments that are out of the ordinary. To push image generation further, I use img2img.

Img2Img

Now, let me tell you, I’m completely in love with this technique. When it comes to breaking free from reality and getting a much more creative result, LoRA exceeds my expectations.

Personally, I start by looking for an image on Pinterest. Today, 90% of Pinterest is probably composed of images generated with MidJourney. I use the JoyCaption technique with ChatGPT, as I explained earlier, to create the prompt that will go along with the image. I set the denoise to 0.6 (this is very important). If you leave it at 1, the img2img will be canceled, and between 0.6 and 1, the image drifts too far from the source image. Set the number of steps to 30; I personally use Euler + SGM_uniform.

Trigger Word

The only thing I can tell you is that if you use "young girl" instead of "young woman," you’ll get a child, which I didn’t have with the first version. You can also try "a 22-year-old girl" for a different result from "a young woman."

You can easily change her hair as well, but I like to stick with "long dark hair" or "brown hair."

That’s all I have to say about this V2. I hope you’ll know how to use it and achieve the result you were hoping for.

Feel free to experiment with img2img and show me your results—it’ll help me gain buzz for a V3.

In my example images, I’m showing you my favorites but not necessarily the ones that most resemble Dua Lipa’s appearance.

See you soon!