This workflow was designed to get the most out of the new InstantID-6M model published earlier this year. It is a workflow which improves upon the abilities of other ID models in the past, such as PuLID, ReActor, InstantID and FaceIDplusv2. It excels at identity preservation of any characters, whilst ensuring expressions and occluding objects in front of the character are preserved in the final image. This includes food, hands, extreme expressions like anger, disgust etc. I have tested out various sdxl and pony models and found that as long as the first KSampler uses the recommended model settings this workflow will produce good results. Anyhow, I have provided full instructions within the workflow and details on what should not be changed, as well as provided some bypassed nodes in case something was changed due to testing or error, and you want to find an easy way back to the default parameters. I have begun testing on illustrious/noobAI models, and if I find settings that work well with this workflow I will post an update in the future.

High accuracy identity transfer with object and expression preservation

Just wanted to give credit to Super-shuhe on huggingface for releasing the 6M version of InstantID - see his repos here <a target="_blank" rel="ugc" href="https://huggingface.co/Super-shuhe">Super-shuhe (Wang Shuhe)</a> To keep expression, either change the InstantID "start at" setting to 0.99 while using the latent from ReActor, or keep it at 0.95 and use the latent from the 1st stage. I have implemented a switch for this.

Files with SHA256: 5d1c98e5b82b2e770801455f7814f27ef9e0c6dab8039d98b65b3168ab42aac0