Add model card

README.md

@@ -0,0 +1,167 @@
+---
+base_model: Comfy-Org/Krea-2
+tags:
+  - text-to-image
+  - lora
+  - krea
+  - krea-2
+  - photorealistic
+  - skin-texture
+  - portrait
+  - realism
+  - inline-studio
+license: other
+license_name: krea-2
+license_link: https://huggingface.co/Comfy-Org/Krea-2
+instance_prompt: inline-skin-lora
+pipeline_tag: text-to-image
+widget:
+  - text: >-
+      inline-skin-lora, detailed skin texture, extreme macro beauty close-up,
+      fair skin densely covered in freckles, visible pores and vellus peach fuzz,
+      editorial beauty photography, unretouched
+    output:
+      url: samples/01-beauty-macro_1_seed90100.png
+  - text: >-
+      inline-skin-lora, detailed skin texture, portrait in soft window light,
+      olive skin, dewy highlights, natural micro-texture, no beauty filter
+    output:
+      url: samples/02-window-light_1_seed90200.png
+  - text: >-
+      inline-skin-lora, detailed skin texture, elderly face, deep creases and
+      crepe texture, age spots, harsh midday sun, documentary photograph
+    output:
+      url: samples/07-mature-skin_1_seed90700.png
+---
+
+# Inline Skin LoRA - Krea 2 (RAW)
+
+A photorealistic **skin texture** LoRA for [Krea 2](https://huggingface.co/Comfy-Org/Krea-2). It pushes the base toward 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 the **RAW** base with [Inline Studio](https://github.com/inlineresearch/Inline-Studio). The LoRA carries over to Krea 2 Turbo at generation time if you prefer the 8-step path.
+
+## Samples by condition
+
+Eight conditions, four seeds each, on the RAW base. Every sheet prints the prompt and settings under each render, so you can read the recipe straight off the image. The first three are in-distribution; the rest push the adapter out of what it trained on.
+
+### 01. Beauty macro, extreme close-up
+
+In-distribution control: the framing and light the adapter was trained on.
+
+![Beauty macro](samples/sets/01-beauty-macro.png)
+
+### 02. Soft window light, bare face
+
+Gentle directional daylight, no makeup. Does texture survive low contrast?
+
+![Soft window light](samples/sets/02-window-light.png)
+
+### 03. Hard midday sun, high contrast
+
+Specular sweat and blown highlights. Does the adapter hold under harsh light? Pores stay resolved inside the blown areas instead of clipping to flat white.
+
+![Hard midday sun](samples/sets/03-hard-sun.png)
+
+### 04. Direct on-camera flash
+
+Worst case for skin: flat harsh flash flattens texture and exaggerates shine.
+
+![Direct on-camera flash](samples/sets/04-direct-flash.png)
+
+### 05. Environmental candid, half body
+
+Pulled back to half-body. Does skin detail survive when the face is small in frame?
+
+![Environmental candid](samples/sets/05-environmental.png)
+
+### 06. Deep skin tones, studio
+
+Tone generalisation. The training set skewed fair, so this is the honest test. Deep tones keep specular sheen without turning to latex.
+
+![Deep skin tones](samples/sets/06-deep-tones-studio.png)
+
+### 07. Mature skin, out of distribution
+
+Stress test: the dataset was all young faces. Renders deep creases, crepe texture, and age spots rather than forcing youth.
+
+![Mature skin](samples/sets/07-mature-skin.png)
+
+### 08. Male stubble, out of distribution
+
+Stress test: coarse hair and heavy texture the training set never contained. Individual follicles and razor detail hold up.
+
+![Male stubble](samples/sets/08-male-stubble.png)
+
+## Trigger
+
+Put this at the front of your prompt:
+
+```
+inline-skin-lora, detailed skin texture
+```
+
+`inline-skin-lora` is the trigger word; `detailed skin texture` is the phrase every training caption led with. Then describe the subject, framing, and light as usual.
+
+## Recommended settings
+
+| | |
+| --- | --- |
+| Base | Krea 2 **RAW** (`krea2_raw_bf16`); also works on Krea 2 Turbo |
+| Strength | 1.0 |
+| Resolution | 1024x1024 |
+| Steps / guidance | 28 steps, CFG 4.5, euler / simple (RAW defaults) |
+| Negative prompt | see below |
+
+Negative prompt used for every sample here:
+
+```
+airbrushed, plastic skin, waxy, plastic doll, smooth featureless skin, cgi, 3d render,
+digital painting, illustration, cartoon, anime, over-retouched, beauty filter, instagram
+filter, oversaturated, blurry, soft focus, low resolution, jpeg artifacts, watermark, text,
+deformed, extra fingers, mutated hands
+```
+
+## Using it
+
+**Inline Studio** (easiest): drop `inline-skin-lora-krea-2-raw.safetensors` into `models/loras/`, add a LoRA loader node on the canvas, and wire it into a Krea 2 node. See the [repo](https://github.com/inlineresearch/Inline-Studio).
+
+**ComfyUI / diffusers**: load it as a standard LoRA on top of Krea 2. It is a PEFT adapter (rank 16) over the Krea 2 transformer; the keys are the diffusers module names, no prefix rewriting needed.
+
+## What it does well, and where it does not
+
+From a 32-render sweep (8 conditions x 4 seeds) on the RAW base:
+
+- **Texture holds everywhere.** No plastic or waxy failure in any of the 32, including hard midday sun (pores stay resolved inside blown highlights) and deep skin tones (specular sheen without turning to latex). All four seeds hold per condition, so it is not seed luck.
+- **It generalises past the training set.** Elderly faces (~83) and male stubble were never in the 26 training images, and both render convincingly. Not overfit to young female faces.
+
+Known limitations, worth knowing before you prompt:
+
+- **Age skews older.** Prompts for 22 to 29 tend to render 35 to 45. Prompt younger than you want.
+- **Framing pulls toward head-and-shoulders.** A "face filling the frame" instruction often comes back wider.
+- **Lighting-style adherence is weak.** It nails the skin, but a specific lighting character (e.g. harsh on-camera flash) can soften toward studio light.
+- **Dark studio backdrops can matte the subject** inside a border on Krea 2 RAW. Spell the backdrop out as full-bleed if you hit it.
+
+## Training
+
+| | |
+| --- | --- |
+| Base | Krea 2 RAW (bf16) |
+| Method | LoRA (PEFT), rank 16, alpha 16, full scope (attention + feed-forward) |
+| Steps | 1500, batch 1, lr 1e-4, resolution 1024 |
+| Caption dropout | 0.05, no flip augmentation |
+| Dataset | 26 image + caption pairs (see the dataset link below) |
+| Tool | [Inline Studio](https://github.com/inlineresearch/Inline-Studio) Trainer |
+
+**Dataset:** [inlineresearch/krea2-skin-lora](https://huggingface.co/datasets/inlineresearch/krea2-skin-lora)
+
+## About Inline Studio
+
+This LoRA was trained with **Inline Studio**, a free, open-source app for AI filmmaking on a node canvas. Its Trainer tab turns the whole pipeline into nodes you wire together: load a dataset, caption it, run the training, watch the loss curve. The finished LoRA drops straight into the app to generate with. Krea 2 generation and LoRA training both run locally on your own GPU, one file per model, no external server.
+
+- Website: https://inlinestudio.art
+- GitHub: https://github.com/inlineresearch/Inline-Studio
+- Dataset for this LoRA: https://huggingface.co/datasets/inlineresearch/krea2-skin-lora
+
+## License and use
+
+The base model is Krea 2; this adapter inherits its terms. Check the [base model license](https://huggingface.co/Comfy-Org/Krea-2) and set this repo's license to match before relying on it commercially. This LoRA generates realistic human skin and faces: use it responsibly, and do not use it to impersonate real people or produce non-consensual imagery.