ComfyUI INT8 vs FP8: comparisons and workflow tutorial

Nerdy Rodent compares INT8 ConvRot image models with FP8, MXFP8 and BF16 variants in ComfyUI, using a 3090. In the examples, the presenter reports faster generation and points to cleaner lettering and logos. These are results from selected prompts, rather than a guarantee that INT8 will outperform every alternative.

The local AI workflow combines Krea-2 Turbo, Qwen Image Edit 2511, Z-Image-Turbo and Boogu Image Edit with two selectable refiner stages. A pose-based Qwen generation passes through Z-Image and Krea-2 for a more photographic result. Other examples cover comic art and masked editing with Boogu's four-step turbo model. Upscaling options include SeedVR2 INT8 for a 4x result of 5472 by 3072, plus Ultimate Upscale for tiled refinement.

The tutorial explains how model labels, text comparisons and Boolean logic route conditioning through Lenient Switch nodes. Different models need different text encoders; an Any Switch approach also requires managing which companion groups are enabled. In the final example, a refiner changes a face. Reusing the original edit conditioning in the second pass restores its similarity in the presenter's comparison. The workflow also accepts FP8 and other formats supported by ComfyUI.