malcolmrey

/

zimage

like 119

for those that i have datasets you can just drop the names anywhere, for those i do not have datasets (no model on the browser of any kind) - just drop it on my coffee page :)

Tswift!

kat dennings

Kat is already there but I'll add Taylor:)

can you do valkyrae, cinna por favor :)

adriaarjona, ivanabaquero, leeleesobieski, lindseystirling, marisharay,

+1 for Leelee Sobieski ! Also, I notice you have a couple of adult movie stars here, so to add to those, Odette Delacroix would be great!

Jenna Jameson,(vintage look), Disha Patani, Veronika Zemanova, Kriti Sanon

I would really like "Taeyeon" :) That would be great!

Alessandra Ambrosio, Elsa Hosk, Sara Jean Underwood

Sophie Turner for Z-image.

Madelaine Petsch, Florence Pugh

Dani Jensen, Rebecca Linares,

Any chance for a Lora of Jean Louisa Kelly? From back in the 90s when she srtill had her babyfat and her face was smoother.

A couple of things I noticed:

1. I get much better results from your Z-Image Loras than the Wan ones so good to see that.

2. Their look can be very different from when they were younger and when they are older (especially with all the surgery these days). Maybe you could do a couple of different versions for age range? Celebs who got popular in their 20's but back in the 1980s sometimes your loras have more of a look of when they are in their 30s and 40s. Or if doing just one lora maybe restrict your dataset more for when they initially got really popular? For example your Courteney Cox looks more like her 30's and 40+ version of her than she did in her 20s and very, very different. Same with Elisabeth Shue, Jennifer Connelly, and many others.

Thanks for all your efforts and your willingness to share!

Really happy with your models so far. Thanks for going to the trouble. I created a dataset for Erica Rose Campbell (crswa/ERC). Perhaps you could include it in your next release? Thx

Danke für die viele Arbeit

LG

AnniTheDuck, Honeypuu and Reved

Also the rest of OfflineTV, namely AriaSaki, Tinakitten, Poki, Jodi etc

Showing my age, but would love to see:

Barbara Eden (from I Dream of Jeannie fame)
Raquel Welch
Dawn Wells (from Gilligan's Island)

I realize the issue with older celebrities is finding decent, color (or colorized) photos of them, though, so it's perfectly understandable that this request would not get anywhere, but I figured it's worth trying!

Thanks for all of your outstanding work!

Catalina Sandino Moreno and Hannah Cheramy for z-image would be much appreciated

Thanks for all your work, very surprising some of the people that have been trained like I didn't expect that.

Two people I think are worthy of being there and I'm surprised aren't yet that I think are worthy of being trained are Alexandra Shipp,Sarah Wright,Lucy Punch, Rosalyn Sanchez, AJ Odudu.

Anyway.. still to have fun with all the models I've downloaded so far, legend!

Sarah Chalke, Gates Mcfadden,America Ferrera, Nichole Sakura,Jennifer Love Hewitt for Zimage

Hi Guys!

Thanks for all the suggestions but I really need your help with image sets if you want to see those new people trained. My time is limited and I can't just hunt the imageset on my own anymore. But I can definitely spend some time cutting them into dataset and of course train them :)

Cheers!

Thanks for all the suggestions but I really need your help with image sets if you want to see those new people trained. My time is limited and I can't just hunt the imageset on my own anymore. But I can definitely spend some time cutting them into dataset and of course train them :)

@malcolmrey Is there a guide you can share that describes how many and what kind of images you are looking for in a training dataset? I think that would be helpful for everyone who is requesting loras.

From looking at some of your reddit posts, I see that you are mostly interested in close-up shots of the face, with some medium shots, and a few full body shots. I saw you mentioned training with under 50 images, but not sure what the minimum should be.

Also, how does age of the person in the image factor into the training? For example, someone like Marisa Tomey has been an actress for many years, so there are photos of her at different ages. I find that some of the loras I've tried skew towards a person's current look rather than younger, even when specifically prompted. I wonder if it's because younger photos were not a part of the dataset or if there is a training limitation.

Hey hey, I guess I need to make a guide somewhere :-)

I believe you're mixing datasets and imagesets. In datasets I do focus on headshots mainly with some bodyshots sprinkled here and there.

For imagesets I just accept good images of that person and the higher the resolution the better. From those I can cut faces and bodies (nowadays I try a bit more body shots).

I create a dataset that has around 22-25 images. This is why i request around 30-50 images because I can then pick those that are best for dataset (based on my experience). It is easier if you prepare a bit more and the error margin can be higher.

Also, how does age of the person in the image factor into the training? For example, someone like Marisa Tomey has been an actress for many years, so there are photos of her at different ages. I find that some of the loras I've tried skew towards a person's current look rather than younger, even when specifically prompted. I wonder if it's because younger photos were not a part of the dataset or if there is a training limitation.

It is best to train on certain era/age of the subject. Mixing various ages is actually detrimental as it they tries to average it our (even if you provide captioning). It is better to train two separate loras - for younger and current age.

The reason why some of the loras skew towards current look is that it is much easier to grab images from recent years. It is very difficult to find 30+ images of someone from 1980s, you're pretty much limited to screengrabbing from movies (which is certainly an option, but not a quick one).

This is why some models have v2 or v3 or more, because people were providing additional imagesets from specific period or even a movie or a particular look :)

@malcolmrey Thank you for the quick guide! Does it matter if the person is looking directly into the camera or not, or if their head is turned so the face isn't symmetrical?

This is why some models have v2 or v3 or more, because people were providing additional imagesets from specific period or even a movie or a particular look :)

Ah, I thought the version numbers meant that they were improvements over the previous version, not a specific look. The way the browser works also makes it difficult to tell what the difference is, as right now the sample images aren't captioned by version.
I wonder if keeping the look lora separate from the base lora would be more informative, something like marisatomey_young_v1. Then the example images would also match better, illustrating the difference more clearly.