I was amazed by the simplicity, and effectiveness of IC-LoRA, a method to use Flux for virtual try-on, comic book generation with very good character consistency and more.

It's really simple, and that's apparent in the code - just train a LoRA on 10-100 images which are a combination of input and ouput (i.e. a piece of apparel combined with it worn). The LoRA can be used in further workflows - using inpainting or other methods.  

It got me thinking that this might be useful in generating virtual staging pictures. The current wokflows rely on controlnets and segmentation. While these can work really well, there are a lot of design choices that need to be made. But - on a deeper level - these very convoluted ComfyUI workflows seem to be missing the point of AI. Why add a whole lot of human logic, when we know how to train machines to make choices for us?

Anyway, I collected a data set of 3D renders, with and without furniture, and trained a LoRA. Flux LoRAs need a lot of VRAM. RunPod now has L40s available with 48 GB of RAM. I used a SPOT instance to save 50% on the cost, and restarted twice - spending a few dollars on training.

The safetensors is avaible on my Huggingface account.