Update README.md

README.md

@@ -26,7 +26,12 @@ tags:
 > - [Comfy-Org/comfy-model-tools](https://github.com/Comfy-Org/comfy-model-tools) (`quant_int8_convrot.py`) 
 > - [Starnodes2024/comfyui-starnodes-modelconverter](https://github.com/Starnodes2024/comfyui-starnodes-modelconverter)
 
-> **Note on my quantized/ INT8 conversions:** 🪶 = Custom quantized / Int8 ConvRot converted version, SD1.5 and Pony models are **intentionally excluded** from this process. SD1.5 is already lightweight, making speed gains negligible. For Pony, besides already being fast, its dual-CLIPs are tightly integrated with the UNet—splitting them would cause severe storage bloat (e.g., requiring 60x2 CLIP-L/G files for 60 checkpoints).
+> **Note on my quantized/ INT8 conversions:**
+> 🪶 = Custom quantized / Int8 ConvRot converted version,
+>
+> Based on testing, INT8 quantized versions of SD1.5 and SDXL architectures (including Pony, Illustrious, etc.) are actually slower than the original models due to VRAM bandwidth bottlenecks and the overhead of runtime dequantization. Since SD1.5 is already lightweight and SDXL architectures are already fast, the speed gains from quantization are negligible.
+>
+> Furthermore, for SDXL variants, the dual-CLIPs are tightly integrated with the UNet—splitting them would cause severe storage bloat (e.g., requiring 60 CLIP-L+ 60 CLIP_G files for 60 checkpoints). However, I still provide these converted versions specifically for low-VRAM environments or testing, allowing users with limited hardware to run models that would otherwise OOM."
 
 
 <details open>