Subtle Styles is a collection of loras for more versatile looks with the Big Love XL checkpoints. They also work with a few other checkpoints (see below), but have no effect with most.
These are style & detailer loras. Trained on 100-230 images. Not only the effect, but also contrast & details usually increase with a higher lora strength. Softness may increase with some loras and prompts. Colors may change too. If nudity should vanish, add additional nudity terms to your prompt.
You can use Big Love here to create 45 free images per day in highest quality with these settings. For discussing photorealistic image generation join me on Discord.
Recommended Lora Strength:
0.25-0.5: Cuteness
0.5-0.9: Craziness, Rayness1/2, Fantasy, Real Cuteness, FluxStyle, ChatGPTStyle, Candidness3
0.25-0.75: All others
The showcase images were created with Big Love XL1/XL2 and DMD2.
Credits:
Special thanks to RaymondLuxuryYacht for giving his permission to train the Rayness lora on 216 of his images.
Description/Training Set:
Cuteness: Cute women, glamour photos
Craziness: Unusual poses & composition, artistic images
Candidness: Unposed spontanous woman, amateur photos
Rayness1/2: Artistic images of women by RaymondLuxuryYacht
Subtleness: Attractive women, photorealistic images by SubtleShader
Retro: Woman from photos ranging from 1920 to 2010.
Beauty: Beautiful women, glamour photos
Fantasy: Photorealistic fantasy images, mostly SD1.5
Realness: Average looking young to mature women, amateur photos
RealCuteness: Cute woman, amateur photos
Candidness2/3: Unposed spontanous shots, amateur photos
MultipleGirls: 2 to 6+ woman, amateur to pro photos
FluxStyle: Flux generated women
ChatGPTStyle: Images generated by ChatGPT and Sora
Pros:
Cuteness: More photorealism, cuter faces, cuter poses
Craziness: Very creative, unusual poses, artistic lighting, strong contrast, surprising
Candidness: Amateur look, natural faces, retro look
Rayness1/2: Vivid colors, faces of Ray's girls, Special effects
Subtleness: Expressive poses, strong contrast increase
Retro: Old photo look, retro hair & clothes
Beauty: Gorgeous faces, Glamour style
Fantasy: Beautiful fantasy motives, creative details
Realness: Everyday faces, amateur look
RealCuteness: Cute Amateur Woman, Candid Poses
Candidness2/3: Amateur look, natural faces
MultipleGirls: More posing variety, less similar faces in one image
FluxStyle: Very creative, Flux lighting, intense colors
ChatGPTStyle: Crazy skin texture, artistic style
Cons:
Cuteness: No effect sometimes, softness at high strength
Craziness: Lower quality at very high strength, full body tendency
Candidness: Soft at high strength
Rayness1/2: Sometimes more graphical than photorealistic
Subtleness: Can get too dark with too high lora strength
Retro: Too strong yellow color cast sometimes
Beauty: ?
Fantasy: Soft skin sometimes
Realness: Lower quality at high lora strength
RealCuteness: ?
Candidness2/3: Grainy at high lora weight
MultipleGirls: Difficult to get good anatomy with multiple people (but a SDXL problem)
FluxStyle: Soft skin (Reduce strength or use img2img)
ChatGPTStyle: Hyperrealistic (not photorealistic), brown colors
Upscaling:
Reduce the lora strength during upscaling or don't use the lora at all to produce more sharpness. Not using the lora for upscaling may change faces though.
Cuteness Advice:
Sometimes it has no real effect. Especially if the woman is already cute. Works best with portraits & medium shots. Less good with full body.
To apply it (to an existing image or prompt) without changing the pose use it in img2ing with 0.75 lora weight & 0.5 denoise. Controlnet in txt2img works less good as it produces bad colors.
Craziness Advice:
Reduce its strength to do xxx scenes for a higher hit rate. As it has a tendancy to rather do full body images, you have to give terms like portrait or closeup a higher weight. Reduce the lora strength if you get too many deformations. Lower lora strength reduces the raw look, but also creativity. Use lower lora strength or remove the lora in img2img to increase image quality.
Candidness Advice:
Higher lora strength produces a more amateur look with more natural faces and softness. You need to find your sweet spot. Does not have the tendancy towards less nudity like the others.
Retro Advice:
Use high enough lora strength for a retro effect. Add 1970s photo or 1950s photo to the prompt to pin the period. Otherwise you may get only a modern photo with vintage fx.
Fantasy Advice:
If you don't use a fantasy-style prompt, it just adds embellishments and creative details. You can add fantasy, mystical atmosphere or science fiction at the beginning of the prompt to produce fantastical images.
MultipleGirls Advice:
You have to add multiple girls or 2girls to 6+girls to the prompt to create more than one woman. Otherwise it only applies its style. It cannot fix anatomical problems with multiple people as that is a basic SDXL problem. You just need to keep on generating till you get good anatomy. It helps with creating different looking women in one image, but better to also prompt different women. It definitely provides more posing variety when it comes to multiple girls
Works With These Checkpoints:
* Big Love XL (recommended)
* Big Love Pony
* Lustify Endgame/OLT
* Big Lust 1.6
* XXX-Ray 1.1
* Analog Madness XL2
* The Araminta Experiment Fv3
* Monolith6
* bigASPv2Lustify
* bigASP 2 (not recommended)
Description
Double as many training images. Higher quality. More variation.
FAQ
Comments (11)
Insane!
best sdxl lora that assure realistic body poees so far from time of start of sdxl , as nude body not so far good with cloths or clothed characters .but forget about consistency ..
, realness version is best , others have no idea what it do ! what cuteness do also ??
There are not many clothes in the training set, so it cannot be good with clothes 😆. Realness was trained on 200 average looking women. Real Cuteness was trained on cute looking women from amateur photos. Cuteness was trained on glamour photos of cute woman.
@SubtleShader wait , question ... is your subtle loras is a failure with models other than your model biglove ?
@amazingbeauty As I wrote they also work with Big Love Pony2, Lustify Endgame, Big Lust 1.6, The Araminta Experiment Fv3, Monolith, bigASP 2 and likely earlier versions of them. Likely with other models that are based on Lustify & bigASP too. I rather prefer them to work with few models and be more effective with them than work with a lot of models and be less effective.
@SubtleShader yes yes i see , make i works effective with minimum lose is better , good work
This is awesome. Too many other lora give 'same-face syndrome' this does not. We need more lora like this with good variety
Big Love XL1 also produces different faces. Even 4 different faces in one image.
Yeah...These enhance Monolith6 a great deal. Thanks for the love on the list! I will see how I can return the love. Great work here!
Thanks. No problem. Saw that you released a new checkpoint. Also based on Lustify or bigASP?
@SubtleShader Yes, it is pretty much a V2, but I added so much in terms of merging and training I figured a new name was better than a new version. BigLove is insanely good btw, I still consider it the best sdxl checkpoint for the style.
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.