What this workflow does

👉 1st generates a subject; then, generates a 2x2 grid with 4 different facial expressions.

This is experimental and your mileage may vary.

I am not qualified to assert if this works, or why; any feedback would be much appreciated.

How to use this workflow

👉 The main prompts' nodes are colored green, you just have to change them to your liking.

All the image nodes are Preview nodes, so you need to save the image manually!

If you want to save individual face expressions, instead of the full grid, you can bypass the FlatLatentsIntoSingleGrid node; or just connect a new PreviewNode before it.

Tips about this workflow

👉 You may want to use ClipSkip, LoRas, Textual Embeddings, or ControNet for some expressions; this is a vanilla template, and you will have to tweak it to your needs. Without relying on additional points of control in the workflow your output will heavily depend on the quality of the used model.

Connect ab_sampler to an ancestral sampler and avoid the use of IPAdapter (or keep its influence low).

If you use ControlNet, mind that B steps in ab_sampler run over a 2x2 grid, so adjust it accordingly.

There is no guarantee that the 1st generation faces are of similar size, which can be a problem; not only due the the used mask, which is somewhat relaxed, but because a change in a smaller area results in a more pronounced change in the subject's expression. There are different strategies you can employ to solve this if you require higher consistency (e.g. ControlNet).

🎥 Video demo link (optional)

👉 [Please add here]