Automatic Image Captioning (Identity-LoRA Training)

Subject: adult woman, trigger word: axad woman
Base model: Z-Image / SD3.5 / Flux (Qwen/T5 encoders)
Tool: OneTrainer, masked training
Method: Hybrid Natural Language (Precise categories embedded in English sentences)

General Rules

1. Syntax (Hybrid Natural Language)
Modern encoders understand context from sentence structures. Do not use comma-separated tags alone. Write short, clear English sentences.
The trigger word "axad woman" must be placed in the very first sentence.

2. Consistency
Use the same standardized terms throughout the dataset. Do not use synonyms for the same poses or expressions (e.g., always "looking aside", never "gazing away").

3. Factuality and Noise Minimization
Write only what the image shows. Do not guess or infer missing information.
No subjective words or quality tags (beautiful, sexy, masterpiece, 8k, photorealistic).
Do not name the person (Amy Adams, actress, celebrity). The subject is always: axad woman.

4. Prohibition of Static Features
Never describe the person's static anatomical features. These are learned from the pixels of the source images.
Prohibited: eye color, skin tone, freckles, nose shape, age.
Hair COLOR is mentioned only if it deviates from the norm. Hair STRUCTURE (wavy, curly, bun) is always written.

5. Separation of Variable Features
Image cropping, pose, gaze, expression, hairstyle, clothing, visible accessories, lighting, and location must be noted in the text. Everything variable must be separated from the identity.


Standardized Terms

Image Cropping (Mandatory in the first sentence)
tight close-up, close-up, headshot, waist-up, half body, three-quarter, full body

Gaze (Critical: zoom in on the iris)
looking at the camera, looking aside (iris turned to the side regardless of face direction), looking past the camera (iris looking past the camera), looking up, looking down, looking over her shoulder

Expression (Mandatory)
neutral expression, serious expression, slight smile, smiling with teeth (mouth closed), wide smile showing teeth, laughing with her mouth open, mouth open, talking (mouth open without laughing), mouth slightly open

Hair (Style, no color)
wavy hair, straight hair, curly hair, long hair, shoulder-length
hair in a low bun, hair in a ponytail, hair in an updo, hair in rollers, hair tucked behind the ears, wet hair

Poses and Angles
rear view, side view, three-quarter view
standing, sitting in a chair, bending forward at the waist, arms crossed, one hand on her chest, walking

Clothing and Accessories
Short identifiable description (e.g., short grey dress, black top with a white v-neck).
Visible prominent jewelry must be noted (drop earrings, thin necklace, hoop earrings) so they are not baked into the anatomy.
Visible bare skin if necessary (cleavage, bare shoulders, bare legs).


Examples of Structure

Good (Hybrid Natural Language):
A full body photograph of axad woman bending forward at the waist with one hand on her chest. She is looking aside with her mouth open, talking. She has curly hair and is wearing a short grey dress with heels and bare legs, in indoor light.

Good (Hybrid Natural Language):
A close-up photograph of axad woman looking at the camera with a serious expression and her mouth slightly open. She has long wavy hair and is wearing a green jacket over a collared shirt, with a brown leather strap over her shoulder, in indoor light.

Good (Hybrid Natural Language):
A waist-up photograph of axad woman standing with her arms crossed. She is looking past the camera with a neutral expression. She has her hair in rollers and is wearing a sheer black robe over dark lingerie, illuminated by window light.

Bad (Outdated CLIP tag list, not for Z-Image):
axad woman, full body, bending forward, looking aside, mouth open, talking, curly hair, short grey dress, heels, indoor light.

Bad (Static features and subjective quality tags):
A masterpiece 8k portrait of axad woman. She has auburn hair, hazel eyes, and a detailed natural face. She is looking at the camera with a beautiful smile.