Text-to-Image
Diffusion Single File
comfyui
Bedovyy commited on
Commit
b3b5635
·
verified ·
1 Parent(s): fa07246

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +38 -3
README.md CHANGED
@@ -9,7 +9,11 @@ base_model_relation: quantized
9
 
10
  # FP8 Quantized model of ANIMA
11
 
 
 
12
  ## Quantized layers
 
 
13
  ```json
14
  {
15
  "format": "comfy_quant",
@@ -22,9 +26,40 @@ base_model_relation: quantized
22
  }
23
  ```
24
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
25
  ## Sample
26
 
27
- | bf16 | fp8 |
28
- |-----|------|
29
- ![ComfyUI_temp_onhsj_00007_](https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/4OzIFBhl_FipyILF3ZXyb.png)|![ComfyUI_temp_onhsj_00006_](https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/Bk0o1slEWvdbO9NynBWGr.png)|
 
 
30
 
 
9
 
10
  # FP8 Quantized model of ANIMA
11
 
12
+ There's two model. FP8 and FP8Mixed. FP8Mixed quantized smaller layer, so expect better quality.
13
+
14
  ## Quantized layers
15
+
16
+ ### fp8
17
  ```json
18
  {
19
  "format": "comfy_quant",
 
26
  }
27
  ```
28
 
29
+ ### fp8mixed
30
+ ```json
31
+ {
32
+ "format": "comfy_quant",
33
+ "block_names": ["net.blocks."],
34
+ "rules": [
35
+ { "policy": "keep", "match": ["blocks.0.", "blocks.1.", "blocks.27.", "adaln_modulation", "v_proj"] },
36
+ { "policy": "float8_e4m3fn", "match": ["q_proj", "k_proj", "output_proj", "mlp"] },
37
+ { "policy": "nvfp4", "match": [] }
38
+ ]
39
+ }
40
+ ```
41
+
42
+ ## Generation speed
43
+
44
+ Tested on
45
+ - RTX5090 (400W), ComfyUI with torch2.10.0+cu130
46
+ - Generates 832x1216, 30steps, cfg 4.0, er sed, simple
47
+
48
+ | Dtype (Quant) | Sage Attn | it/s | Time (s) | Relative Speedup (%) |
49
+ | ------------- | --------- | ---- | -------- | -------------------- |
50
+ | bf16 | X | 4.33 | 7.07 | 0% |
51
+ | bf16 | O | 4.66 | 6.59 | +7.6% |
52
+ | fp8mixed | X | 4.79 | 6.42 | +10.6% |
53
+ | fp8mixed | O | 5.25 | 5.87 | +21.3% |
54
+ | fp8 | X | 4.94 | 6.23 | +14.1% |
55
+ | fp8 | O | 5.40 | 5.72 | +24.7% |
56
+
57
+
58
  ## Sample
59
 
60
+ | bf16 | fp8mixed | fp8 |
61
+ |-----|------|------|
62
+ |![ComfyUI_temp_onhsj_00007_](https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/4OzIFBhl_FipyILF3ZXyb.png)|![ComfyUI_temp_gqpfh_00024_](https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/GiVW4VHm1AnMB73q1mv1w.png)|![ComfyUI_temp_onhsj_00006_](https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/Bk0o1slEWvdbO9NynBWGr.png)|
63
+ |![anima_bf16_1](https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/_u2hrgEulSMOPiNux9bYI.png)|![anima_fp8mixed_1](https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/dWEUFK3lFLK8_wsiRDMzJ.png)|![anima_fp8_1](https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/b769101lzVGLjPIGJApJp.png)|
64
+
65