Ccadic/Anna-ZImage-Turbo-LoRA

Anna — Cinematic Virtual Actress

Anna is an original fictional female character created for cinematic storytelling, virtual production, concept art and consistent character generation.

This Character LoRA was trained for Tongyi-MAI/Z-Image-Turbo. It is designed to preserve Anna's recognizable facial identity across different camera angles, expressions, hairstyles, clothes, lighting conditions, locations and historical or science-fiction settings.

Anna has a mature Mediterranean appearance, very dark brown hair, gray-green hazel eyes and a calm, intelligent and self-assured cinematic presence. She is intended to function as a reusable virtual actress rather than as a single fixed portrait.

Anna is an original fictional character and is not intended to represent any real person.

Technical information

Item	Value
Public name	Anna — Cinematic Virtual Actress
Type	Character LoRA
Base model	Z-Image Turbo
Trigger word	ana_biosyn
Recommended initial strength	0.8–1.0 — still being evaluated
Steps	8–9
Guidance	1
Training resolution	512–768
Recommended generation resolution	768–1024
Format	Safetensors BF16
Tested interface	ComfyUI
Character type	Original fictional character

Note: the tested ComfyUI workflow uses Guidance 1.
When using the official Diffusers ZImagePipeline, use guidance_scale=0.0.

What this LoRA does

The LoRA adds Anna's identity to Z-Image Turbo. It can be used to generate:

• close-up and medium cinematic portraits;
• three-quarter and profile views;
• full-body compositions;
• different facial expressions and emotions;
• contemporary, historical and science-fiction costumes;
• natural daylight, studio lighting and dramatic cinematic lighting;
• recurring images of the same character for storyboards, films and visual narratives.

The LoRA controls the character identity. The surrounding scene, clothing, pose, expression, camera and lighting should still be described in the prompt.

Installation in ComfyUI

1. Download ANA_IDENTITY_ZIMAGE_V1.safetensors.

2. Copy it into:

   ComfyUI/models/loras/
   

3. Restart ComfyUI, or refresh its model list.

4. Load Z-Image Turbo as the base model.

5. Add a LoRA Loader node and select ANA_IDENTITY_ZIMAGE_V1.safetensors.

6. Start with a LoRA strength between 0.8 and 1.0.

7. Include the trigger ana_biosyn near the beginning of the prompt.

Recommended settings

Setting	Recommended value
Base model	Tongyi-MAI/Z-Image-Turbo
LoRA strength	0.8–1.0
Sampling steps	8–9
Guidance	1
Sampler	FlowMatch-compatible Z-Image workflow
Resolution	768–1024 pixels on the principal dimension

These are starting values rather than absolute requirements. A lower strength can give the prompt more freedom, while a higher strength can reinforce Anna's identity.

Trigger word

ana_biosyn

For best identity retention, use ana_biosyn woman close to the beginning of the prompt.

Example prompts

Natural close-up

close-up cinematic photograph of ana_biosyn woman, calm neutral expression,
direct eye contact, loose dark brown hair, pale blue cotton shirt, modern apartment,
soft window daylight, realistic skin texture, shallow depth of field,
natural cinematic color grading

Science-fiction scene

ana_biosyn woman inside a vast futuristic archive, intelligent and determined gaze,
dark green technical uniform, translucent data displays, restrained cinematic lighting,
subtle atmospheric haze, realistic skin, highly detailed science-fiction film still

Historical character

ana_biosyn woman as a nineteenth-century scientific explorer, chest-up portrait,
authentic period clothing, leather field notebook, soft natural window light,
realistic skin texture, cinematic historical drama

Emotional scene

cinematic portrait of ana_biosyn woman, restrained sadness, moist gray-green eyes,
subtle natural tears, windswept dark hair, overcast coastal landscape,
realistic emotional expression, dramatic film still

Prompting advice

• Put ana_biosyn woman near the beginning of the prompt.
• Describe the desired framing explicitly: close-up, chest-up, three-quarter view, profile or full-body.
• Describe expression, clothing, lighting and location separately.
• If the identity becomes too dominant or the pose too rigid, reduce the LoRA strength.
• If facial identity becomes too weak, move the trigger earlier and slightly increase the strength.
• For comparisons, keep the same prompt and seed while changing only the LoRA strength.

Current limitations

• Identity consistency can vary in very distant full-body shots.
• Extreme camera angles, heavy facial occlusion or highly stylized rendering may reduce resemblance.
• Complex scenes containing several similar female characters can cause identity mixing.
• Recommended strength is still being evaluated and may be refined in a later version.
• Results depend on the exact ComfyUI Z-Image Turbo workflow and its model versions.

Training summary

• Architecture: Z-Image Turbo Character LoRA
• Training steps: 2,000
• LoRA rank / alpha: 16 / 16
• Training precision: BF16
• Optimizer: AdamW 8-bit
• Noise scheduler: FlowMatch
• Text encoder training: disabled
• Training objective: balanced character identity preservation

File

ANA_IDENTITY_ZIMAGE_V1.safetensors

Only the .safetensors file is required for inference. Training caches, optimizer states and source datasets are not required.

Intended use

This LoRA is intended for lawful creative uses including cinematic previsualization, original fiction, storyboards, concept art, virtual production and experimentation with consistent fictional characters.

Users are responsible for the images they generate and for complying with applicable laws, platform rules and the license of the base model. Do not use the model to impersonate a real person or to create misleading representations presented as authentic events.

License

This repository is released under the Apache License 2.0. Use of the LoRA also remains subject to the terms applicable to the Z-Image Turbo base model.

Creator

Created and trained by Philippe Cadic.

• Hugging Face: Ccadic
• X / Twitter: @sulfuroid

© 2026 Philippe Cadic. Released under the Apache License 2.0.

Version

v1.0 — Initial public release.