Someone did a real-world test of PhoneDiffusion, running local image generation on an iPhone, even with the SDXL model loaded. They covered a lot: SD1.5 generation, SDXL and RealVisXL img2img, upscaling locally, background removal, negative prompts, various styles, custom workflows, and even generating in airplane mode.
My takeaway is that fast models like SD1.5 LCM are already usable on phones for quick inspiration; SDXL is heavier, first model load takes a while, but still impressive that it runs on a phone at all. The standout feature is chaining multiple steps together: generate, transform, upscale, remove background—all in one go.
OP also asked what the biggest bottleneck is now: speed, model size, heat, interaction, or image quality. Personally, I think heat is the real killer—after a few images, the phone starts burning up.