PLEASE post your images directly to my lora page or add a tip for I can generate some buzz back to continue making loras. Thank you &lt;3 &lt;3 -- Diamond, a personal lora I've finally decided to let out in the world. She is 100% artificial, made from a small dataset of AI images (of fake people) I've been using to get better faces out of Flux, originally with face swap at first then later on with this lora. Why the name Diamond you ask? Because she brings a sparkle to my eye and was hard as nails for me to get trained, because I'm still learning. A lot. She's trained to be 28-32 years old or "mature" if you like. Though she will certainly do young adult and can sorta do a elderly adult, though her resemblances I trained into it start to fade off a bit. She usually had a defined button nose, brown eyes and either dirty blonde or light brunette hair and full lips. She had a good amount of skin texture, blemishes, so on so on. KREA 2 version: Trigger, tho not needed, is D14M0ND. If you want her as I intended her to be, prompt her as a 30 year old/early-30s and sometimes depending on the model you can also use middle aged, but that might push her into much older area. The two most common trained phrases in the dataset to achieve that was "woman in her early-30s" and "photography of a caucasian woman, she's in her very early 30s" weight 1.0 is suggested but Ive also had her works as low as 0.4, just depends on the model and if you have other loras going on. ALL OF MY EXAMPLES HAVE BEEN UP-SCALED USING SEEDvr2 -- this often hides minor artifacts. These artifacts ARE NOT caused by the lora in almost all cases but caused by the used checkpoint model, mainly seen with int8 because people still don't make them correctly. NO FEE BUT I WOULD LOVE SOME TIPS TO MAKE UP FOR IT! Ernie-image: Best practices is to add the trigger at/near the beginning of the prompt: Trigger: D14M0ND, photography of a light-skinned woman, with long hair, she's in her early 30s Additional if you want even more accuracy: blank expression/neutral expression. I'm still feeling out how Ernie handles characters. So don't expect this to be perfect 100% of the time but I have been getting pretty good consistency, using the trigger provided. I think I need a larger dataset (this one is only 50 images). Expect a eventual update/replacement for this eventually probably. I want to see how she works with other people at the helm first. Works good with other loras, keep her weight above other loras if you want to maintain her facial qualities. All images posted at upscale using SeedVR2 upscaler using 35 steps @ cfg 4 using whatever sampler/sche that is displayed on their metadata (I tested a few but mainly stuck with uni/sgm_uni. A few were made with a turbo lora also which worked well. Onsite folks should stick with the default, tho I do think boosting the steps will increase quality but will cost you more. Weight: 1.0 is suggested if you want to maintain accuracy ZiT VERSION: Trigger: D1AM0ND Also suggest only using: Woman/Mature woman. Personally think Mature woman looks best, putting her at a attractive late 20s early 30s something of a woman. Middle-aged also worked but changed face lightly more for reasons. Weight: 0.8-1.0 is best. In my testing she works okay with other loras and checkpoints. Keep her weight significantly higher than than others and she keeps her awesomeness. ALL CARD IMAGES WERE MADE WITH Z-Image Turbo official model using Seed3/sgm_uniform. The eyes, and only the eyes, were also ran though Facedetailer to give them a more professional make-up feel to them. Something I do on all my up-scaled images. Alt settings I like to use with her: Seed2/sgm_uniform - Res_multistep/sgm_uniform - Euler_a/Simple - deis/simple - Heun/Simple -- 9-12 steps FLUX VERSION: It will do full body shot with enough weight but does better at portrait type images. She wears more or less what you tell her to wear and will go to location to tell her to be at. Poses usually work but you might have to fiddle with it a bit with side views or complex poses. I mainly focused on the face, eyes, hair and lips with the dataset I used. She does have a small problem playing with other loras if their weights are set higher than hers, or if she isn't in the prompt before they are. So keep that in mind. Because she seems to be easy to be "over-written" by them if she isn't the star of the <s>show</s> prompt. USE THE TRIGGER WORD: d1am0nd Helpful additional prompts: #*-year-old mature woman, looking at viewer #* = adult age, 28-32 range is best imo. Optional quality boosting prompts: award-winning, professional, highly detailed, photorealistic If you use hi-res fix or run your images through a second sampler, keep the denoise really low, otherwise you're gonna end up with probably too much skin texture. SDXL Version: It follows the resemblance as the flux version pretty well. Make sure you are using the triggers: D14M0NDXL, mature woman or woman Additional optional tags to help with realism: Photorealistic, Hyperrealistic, film photography, soft lighting, heavy shadow And because its SDXL I highly suggest always using the Negatives: child, cartoon, bad quality If you have a problem making full body shot, make sure to include the prompt "full body shot" as well as describing the whole body, eg: pants, socks, boots or describing them interacting with something on the floor All examples were made using <a target="_blank" rel="ugc" href="https://civarchive.com/models/239561">WildcardX-XL </a>+ hi-res fix using 4x_NMKD-Siax_200k and a 0.26 denoise at a 1.25 upscale, use whatever sampler you wish but I suggest Restart/Exponential or DPM++ 3M SDE/Exponential

Diamond, a fake person

Late but here she is! Krea 2 version. Trigger, tho not needed, is D14M0ND. If you want her as I intended her to be, prompt her as a 30 year old/early-30s and sometimes depending on the model you can also use middle aged, but that might push her into much older area. The two most common trained phrases in the dataset to achieve that was "woman in her early-30s" and "photography of a caucasian woman, she's in her very early 30s" weight 1.0 is suggested but Ive also had her works as low as 0.4, just depends on the model and if you have other loras going on. ALL OF MY EXAMPLES HAVE BEEN UP-SCALED USING SEEDvr2 -- this often hides minor artifacts. These artifacts ARE NOT caused by the lora in almost all cases but caused by the used checkpoint model, mainly seen with int8 because people still don't make them correctly. I also used <a rel="ugc" href="https://civitai.red/models/2268008/realistic-snapshot-z-image-turbo-krea-2?modelVersionId=3084537">Realistic Snapshot</a> pretty frequently because it just makes the images more interesting. Totally not needed but you'll see it often in the meta. NO FEE BUT I WOULD LOVE SOME TIPS TO MAKE UP FOR IT!

Files with SHA256: 5c3a5d6f80b5bf6e253cc9e8c80e0bbb594176cf5cd8f7f6dec268a1816190c4