Krea 2 Realistic Skin Texture
A photorealistic skin LoRA for Krea 2. It brings back real, unretouched skin: visible pores, fine vellus hair, natural subsurface scattering, and honest imperfection, instead of the airbrushed plastic look diffusion models drift into. It is a texture and realism adapter, not a specific person.
Trained on Krea 2 RAW with the Inline Studio Trainer, from 26 images. The LoRA carries over to Krea 2 Turbo for generation.
Trigger words
inline-skin-lora, detailed skin texturePut these at the front of your prompt, then describe the subject, framing, and light.
Recommended settings
Base: Krea 2 RAW to train, generate on Krea 2 Turbo
LoRA strength: 0.6 to 1.0, default 1.0. Lower it to blend the texture in, raise toward 1.0 for full effect. The sample images here are at 1.0.
Steps: 8, guidance 1 (Krea 2 Turbo defaults)
Resolution: 1024x1024
Suggested negative prompt: airbrushed, plastic skin, waxy, smooth featureless skin, over-retouched, beauty filter, cgi, 3d render, blurry, soft focus.
Training
Base: Krea 2 RAW (bf16)
Steps: 1500, batch size 1
Epochs: about 58 (1500 steps over 26 images, no flip augmentation)
Rank / alpha: 16 / 16, full scope (attention + feed-forward)
Learning rate: 1e-4
Resolution: 1024px, caption dropout 0.05
Dataset: 26 image and caption pairs
What it does well, and where it does not
Tested across eight conditions, three in-distribution and five out of distribution. Texture holds throughout, including the cases built to break it: hard midday sun keeps pores resolved inside blown highlights, and deep skin tones keep specular sheen without turning to latex. It also generalises to mature skin and male stubble, neither of which were in the training set.
Known limits, so you can prompt around them:
Age skews older than asked. Prompt younger than you want.
Framing pulls toward head-and-shoulders.
A specific lighting character (like harsh direct flash) can soften toward studio light.
Made with Inline Studio
Trained in the Inline Studio Trainer, a free, open-source app for AI filmmaking on a node canvas: dataset, captioning, training, and the loss curve are all nodes you wire together.
Website: https://inlinestudio.art
App and source: https://github.com/inlineresearch/Inline-Studio
Model on Hugging Face: https://huggingface.co/inlineresearch/skin-lora-krea-2-raw
Training dataset (26 image and caption pairs): https://huggingface.co/datasets/inlineresearch/krea2-skin-lora
License
Base model: Krea 2. This LoRA is a derivative and is released under the Krea 2 Community License: https://www.krea.ai/krea-2-licensing
Fine-tuning and LoRAs are expressly permitted.
Commercial use is free only under 1,000,000 USD total annual company revenue (trailing twelve months). At or above that threshold you need a separate enterprise license from Krea.
Derivatives inherit the same license, and attribution is required.
Do not generate unlawful content, non-consensual imagery, or content that impersonates real people.
Not affiliated with or endorsed by Krea.
Description
FAQ
Comments (9)
A lora with the purpose of improving skin quality shouldn't need trigger words in my opinion.
it also shouldn't look like a poorly rendered SDXL gen either.
@Ziz0miz0 We defiantly fix these in next release.
this just looks like overbaked older AI models, did you use real pictures or synthetic data?
I have used real pictures to train this.
I'm getting pretty good results with this LoRA, thanks! Yes, the sample pictures might not be the best (probably sub-optimal prompt + settings), but the LoRA itself is actually pretty solid, IMO. I'm getting pretty good results with Krea 2 RAW + the Turbo LoRA (rank 64), res_2s sampler + ddim_uniform scheduler, CFG 2.0 and ~14 steps. You should really give it a try.
Trained from 26 images, are you joking? Too rough
It's a skin LoRA. You don't really need hundreds of images to capture skin detail. Did you try it? I did and I'm pretty satisfied with the results.
Yes that's correct, Dataset is available on HF, Skin images doesn't have infinite variations, so just needed the right images.