Update README.md
Browse files
README.md
CHANGED
|
@@ -21,14 +21,14 @@ tags:
|
|
| 21 |
- image-editing
|
| 22 |
---
|
| 23 |
|
| 24 |
-
> **Format Notice:** Except for SAM3, all models in this repository are provided in **INT8** format. Some of these INT8 versions were converted by me using the following tools:
|
| 25 |
-
> - [silveroxides/convert_to_quant](https://github.com/silveroxides/convert_to_quant)
|
| 26 |
-
> - [Comfy-Org/comfy-model-tools](https://github.com/Comfy-Org/comfy-model-tools) (`quant_int8_convrot.py`)
|
| 27 |
-
> - [Starnodes2024/comfyui-starnodes-modelconverter](https://github.com/Starnodes2024/comfyui-starnodes-modelconverter)
|
| 28 |
-
|
| 29 |
> **Note on my quantized/ INT8 conversions:**
|
| 30 |
>
|
| 31 |
> 🪶 = Custom quantized / Int8 ConvRot converted version,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
>
|
| 33 |
> 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.In my simple benchmarks, only the Illustrious (IL) INT8 version showed a positive speed optimization, while all other INT8 versions were slower.
|
| 34 |
>
|
|
|
|
| 21 |
- image-editing
|
| 22 |
---
|
| 23 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
> **Note on my quantized/ INT8 conversions:**
|
| 25 |
>
|
| 26 |
> 🪶 = Custom quantized / Int8 ConvRot converted version,
|
| 27 |
+
>
|
| 28 |
+
> Some of these INT8 versions were converted by me using the following tools:
|
| 29 |
+
> - [silveroxides/convert_to_quant](https://github.com/silveroxides/convert_to_quant)
|
| 30 |
+
> - [Comfy-Org/comfy-model-tools](https://github.com/Comfy-Org/comfy-model-tools) (`quant_int8_convrot.py`)
|
| 31 |
+
> - [Starnodes2024/comfyui-starnodes-modelconverter](https://github.com/Starnodes2024/comfyui-starnodes-modelconverter)
|
| 32 |
>
|
| 33 |
> 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.In my simple benchmarks, only the Illustrious (IL) INT8 version showed a positive speed optimization, while all other INT8 versions were slower.
|
| 34 |
>
|