Hugging Face's logo

AX1Y2JP
/
FLUX.2-dev-INT8-ConvRot

Diffusion Single File
comfyui
image-generation
image-editing
flux
Model card Files
xet
 Community
1
AX1Y2JP commited on
Commit
b0bb53f
·
verified ·
1 Parent(s): 1b779b5

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +3 -1
README.md CHANGED
@@ -13,7 +13,9 @@ tags:
13
  ---
14
 
15
  A FLUX.2-dev model quantized to INT8 with ConvRot using a conservative quantization policy.
16
- In T2I on my potato machine, int8-convrot-aggressive is 20 seconds faster than int8-convrot. Both models output exactly the same structure/composition as bf16.
 
 
17
  T2I Example:
18
  <img src="ComparisonA.webp" width="1000">
19
  <img src="ComparisonB.webp" width="1000">
 
13
  ---
14
 
15
  A FLUX.2-dev model quantized to INT8 with ConvRot using a conservative quantization policy.
16
+
17
+ On my potato machine, in T2I without using distilled LoRA, int8-convrot-aggressive is 20 seconds faster than int8-convrot. Neither model produced images that deviated from bf16.
18
+
19
  T2I Example:
20
  <img src="ComparisonA.webp" width="1000">
21
  <img src="ComparisonB.webp" width="1000">