Image conditioning does not preserve identity in image-to-video generation

[mkamranr]

I am testing LTX-2.3 image-to-video generation using TI2VidTwoStagesPipeline with the --image conditioning option.

The input image is an AI-generated person (not a real person). My expectation is that the generated video should keep the same character identity from the input image while only changing the motion.

However, the generated video changes the person's:

• Face
• Hairstyle
• Clothes
• Overall appearance

The output looks like a different generated person rather than the same character from the reference image.

Example usage: --image person.png 0 0.8

I tried different image conditioning strengths and prompts, but the identity is still not preserved.

My questions:

• Is --image only intended for general visual conditioning and not identity preservation?
• Is there a recommended pipeline, LoRA, or IC-LoRA for keeping the same character identity from a single image?
• Are there recommended settings or prompting techniques for better identity consistency?

Thanks!