I will train a lora model for stable diffusion, flux or others

In most cases, Stable Diffusion XL is the best choice as it can generate high-quality images in realistic, digital and cartoon settings. Flux has state-of-the-art performance in generating realistic images with human characters.

A minimum of 5 high-quality images are required to fine-tune a LoRA. An ideal training set is made of 30 consistent images. If you have only a few images at your disposal, I can synthetically augment the dataset. The more images, the better the LoRA quality.

It doesn't matter, as long as enough samples are provided for each aspect ratio! Training works in aspect ratio buckets, grouping together images into buckets of similar aspect ratios. To get a LoRA that works equally well in both directions, include a mixture of portrait and landscape images.

Stable Diffusion 1.5 is optimized for 512×512 images. In contrast, Stable Diffusion XL supports a wider range of aspect ratios and higher resolutions, typically between 1024×1024 and 1216×832. Flux-1-dev supports generating images in full HD and up to 2K resolution.