Nov 18 12:57 · AI content
Jane is a fictional adult character created for photorealistic image generation. This LoRA focuses on her facial identity, with a training dataset that includes different camera angles, framing, clothing, and lighting conditions.
Her appearance combines green-hazel eyes, lightly freckled skin, and long, wavy blonde hair with darker roots and pale strands framing her face. Her visual character is grounded in natural portrait photography: visible skin texture, understated expressions, relaxed poses, and everyday clothing.
The goal behind Jane is to develop a recognizable character who can be placed in different photographic situations. The dataset explores that idea through close-up portraits, seated compositions, full-length photographs, indoor scenes, and outdoor settings.
Jane’s training images include a neutral frontal portrait, three-quarter views from both sides, a side profile, softly lit studio photographs, a portrait with more directional side lighting, and casual scenes near a café window, in a park, and beside a city building. Clothing varies across the images, including simple T-shirts, knitwear, a cardigan, a chambray shirt, and casual jackets.
This variation was included to give the training process examples of Jane beyond a single outfit, background, or composition. Preserving her identity across new generations remains something to evaluate through practical testing; the diversity of the dataset describes the training approach rather than a guarantee of identical results in every scene.
Workflow
1. Establishing the character
The process started with an anchor portrait defining Jane’s appearance. This image served as the identity reference for the additional dataset images.
The main features to preserve were her facial proportions, eye shape and spacing, brow shape, nose, lips, jawline, and chin. Her lightly freckled skin and distinctive blonde hair with darker roots were also part of the visual reference.
Keeping these characteristics consistent was the central objective when expanding the dataset.
2. Building the dataset
Additional images were generated using the anchor portrait as the identity reference, with instructions to preserve Jane’s appearance while changing the photographic situation.
The final training set contains 12 captioned images covering:
- Frontal and near-frontal close-ups.
- Left and right three-quarter portraits.
- A left-facing side profile.
- Seated portraits with different body orientations.
- Full-length studio and street photographs.
- Outdoor portraiture under overcast light.
- A café scene with window illumination.
- A studio portrait with directional side lighting.
The variations include changes in expression, hair arrangement, clothing, background, and lighting. Some images show loose waves, while others use a loose ponytail or hair swept toward one shoulder.
This creates a compact dataset with several views of the same character and a range of everyday photographic contexts.
3. Captioning
Each image received an individual English caption beginning with the character trigger:
janev4
The rest of each caption describes the visible scene, including framing, pose, clothing, background, and lighting.
For example:
janev4. Frontal close-up photograph of an adult woman with loose waves, a neutral expression and a white crew-neck T-shirt against a light gray backdrop.
The captioning approach associates the shared trigger with the character while describing the elements that change between images. Clothing and setting are explicitly named so that the captions distinguish those scene details from the recurring identity.
4. Training
Jane was trained as a LoRA for Krea 2 RAW using Musubi Tuner.
The training configuration uses:
- Dataset: 12 captioned images.
- Configured training resolution: 740 × 740, with aspect-ratio bucketing enabled.
- Training steps: 120.
- LoRA rank / alpha: 8 / 8.
- Optimizer: AdamW.
- Learning rate: 0.0001.
- Learning-rate schedule: Constant.
- Batch size: 1.
A checkpoint was saved at step 60, followed by the step-120 checkpoint and the final LoRA file. These provide intermediate and final versions that can be compared during evaluation.
5. Using Jane
Load the LoRA in a workflow compatible with Krea 2 RAW and include janev4 in the positive prompt. Then describe the photograph you want to create: framing, pose, clothing, environment, expression, and lighting.
A simple prompt structure is:
janev4. [framing], [pose and expression], [clothing], [environment], [lighting].
For example, a starting prompt to test is:
janev4. Chest-up photograph of an adult woman wearing a navy knit sweater, sitting beside a cafe window, looking slightly away from the camera, soft natural window light, realistic skin texture.
This is an illustrative prompt, rather than a documented generation result. LoRA strength and generation settings should be adjusted through testing in the workflow being used.
6. Evaluating the result
The next stage is to compare the saved checkpoints across the same prompts and generation settings. Useful checks include frontal likeness, three-quarter views, profile shape, facial detail at wider framing, and how the character responds to changes in clothing and lighting.
Close-up portraits help reveal differences in facial structure, while full-length compositions test how recognizable Jane remains when her face occupies a smaller part of the image. Keeping the prompt and seed fixed when comparing checkpoints makes those differences easier to assess.
Jane is the starting point for exploring this character across a broader range of scenes, while keeping the training process and its practical limitations visible.
lora
ai
text2img
krea2