Quantizing image models isn’t the same as with LLMs. For language models, you’re compressing text prediction weights—missing one token barely shows in the text. But for images, you’re compressing the UNet or DiT, plus the text encoder, and quality loss hits the pixels directly. Even a tiny difference, your eyes catch it immediately.
My experience is FP16 is still the default go-to, best quality, and you can run it comfortably with 8GB+ VRAM. FP8 cuts UNet VRAM in half, and on natively FP8-supporting cards like the 40 or 50 series, quality loss is almost invisible. It’s the sweet spot for 12-16GB cards. NF4, aka 4-bit, is the last resort for low VRAM cards—it runs, but quality drops noticeably with complex prompts.
These numbers are all tested at 1024 resolution, single batch, actual results vary with implementation and resolution. Don’t blindly chase lower precision to save VRAM—you’ll pay for it in quality.