‘Virtual Poses’ is a project aimed at making a fashion photoshoot 100% virtual using the latest technology.
Overall, this project aims to refine and master the control on these aspects;
Unique poses,
Intuitive camera angles,
The ability to change clothing,
Models’ faces and creating consistent faces,
Flexibility in composition by using depth math,
The mood and style of the image by using IP adapters and LoRas ,
The full creative power of AI.
Unique poses,
Intuitive camera angles,
The ability to change clothing,
Models’ faces and creating consistent faces,
Flexibility in composition by using depth math,
The mood and style of the image by using IP adapters and LoRas ,
The full creative power of AI.
I built a VR photo studio where visitors can pose in a virtual room. I capture their poses using Move-AI with my phone and transfer the rigging data to a Metahuman in the virtual environment. Then, I use ControlNets and API adapters to track the virtual poses. Based on these poses, I create my final editorial images using a Flux-based AI image generation model.
My first editorial, with my first virtual model Nana.
Virtual Model Nana mimics the pose created by a visitor in the virtual studio.
Detailed Methodology
1-SIMPLE IMAGE GENERATION IN DIFFERENT MODEL TYPES SD XL MODELS FLUX MODELS ( I use gguf models which are the only ones my computer ) supports.
2-IPADAPTERS I use IPadapter for the face swap, cloth swap and style transfer.
3-CONTROLNETS I use Controlnets to create openposes and depthmap
4-INPAINTING I use this method to give a stylistic look to my models by making area-based changes, such as adjusting the hairstyle, eyes, or accessories.
5-FOCUS METHOD Focus is a new method I want to use to actually mimic the real camera in/out focuses.
6-EXPERIMENTAL METHODS- NOISE TRAINING Using noises is a 3D technique converted to AI workflows to increase the variations in images.
7-LORAS I use Loras to have consistent character and style.