Image Generation

Image generation is where the integrated GPU on the Ryzen AI Halo is working well, with more than enough memory to hold the big diffusion models. A discrete card looks like the obvious choice, but the model does not fit here. Qwen-Image at full bf16 precision needs about 40 GB, more than a 32 GB RTX 5090 and far beyond a 16 GB 5060 Ti, so both fail right away. You can force a GPU into running it by offloading to system RAM or dropping to a smaller precision, but that is much slower as we've seen on the LLM page.


Running Qwen-Image at 30 steps, from the same default prompt and the same random seed, the Halo finished in about two minutes; the NVIDIA DGX Spark was twice as fast, around a minute. I think what's at play here is that CUDA is more mature, with better-optimized kernels. The output from the two boxes is visually identical, so this is purely a difference in speed.


Here is one of the images it produced at 1024×1024