A pre-merge experimental test aimed at delicately refining and calibrating complex anatomy in the Z-image base model. This is not a global restructuring of the latent space, but a careful, targeted tuning of problem areas and textures.
📊 About the dataset and structure
The model was trained on a balanced dataset of 1,000 images:
500 images targeted macro: the aesthetics of female and male hands, grips, cuticle texture, nail plates, and premium manicures.
500 images general physiological anatomy: correct elbow bends, knee joints in complex sitting/kneeling postures, natural skin folds, and muscle dynamics.
🛠 What this refinement provides
Focus on the hand: The main emphasis is on eliminating floating finger geometry. The model better understands the physics of holding objects (pens, glasses, and bars).
Minor anatomical adjustments: Subtle adjustments have been made throughout the anatomical mesh. The basic Z-image model already holds proportions well, and this test only makes joint curves (knees and elbows) smoother and cleaner, removing defects at the joints.
Branding effect: Due to the high training density at high ranks, the model has assigned a token to the concept of quality and can automatically generate a trigger word in the form of engraving or embossing on metal and glossy surfaces.
⚙️ Usage recommendations
Trigger word: ZIATt512
LoRA Weight: 0.60 to maintain prompt flexibility in wide shots and up to 1.0 for extreme macro close-ups of the hands.
CFG: For close-ups of hands and objects, it is recommended to keep it around 3.5 for ideal plasticity and softness of the skin.