Hugging Face's logo

Bedovyy
/
Anima-FP8

Text-to-Image
Diffusion Single File
comfyui
Model card Files
xet
 Community
5
Bedovyy commited on
Commit
9625f2a
·
verified ·
1 Parent(s): 433323c

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +29 -22
README.md CHANGED
@@ -9,33 +9,36 @@ base_model_relation: quantized
9
 
10
  # FP8 Quantized model of ANIMA
11
 
12
- There are two models - FP8 and FP8Mixed.
13
 
14
- FP8Mixed quantized fewer layers, so expects better quality.
 
 
 
15
 
16
 
17
  ## Generation speed
18
 
19
  Tested on
20
- - RTX5090 (400W), ComfyUI with torch2.10.0+cu130
21
  - Generates 832x1216, 30steps, cfg 4.0, er sde, simple
22
 
23
- | Dtype (Quant) | Sage Attn | it/s | Time (s) | Relative Speedup (%) |
24
- | ------------- | --------- | ---- | -------- | -------------------- |
25
- | bf16 | X | 4.33 | 7.07 | 0% |
26
- | bf16 | O | 4.66 | 6.59 | +7.6% |
27
- | fp8mixed | X | 4.79 | 6.42 | +10.6% |
28
- | fp8mixed | O | 5.25 | 5.87 | +21.3% |
29
- | fp8 | X | 4.94 | 6.23 | +14.1% |
30
- | fp8 | O | 5.40 | 5.72 | +24.7% |
31
 
32
 
33
  ## Sample
34
 
35
- | bf16 | fp8mixed | fp8 |
36
- |------|----------|-----|
37
- |<img src="https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/4OzIFBhl_FipyILF3ZXyb.png" width="208">|<img src="https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/GiVW4VHm1AnMB73q1mv1w.png" width="208">|<img src="https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/Bk0o1slEWvdbO9NynBWGr.png" width="208">|
38
- |<img src="https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/_u2hrgEulSMOPiNux9bYI.png" width="208">|<img src="https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/dWEUFK3lFLK8_wsiRDMzJ.png" width="208">|<img src="https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/b769101lzVGLjPIGJApJp.png" width="208">|
 
 
39
 
40
  ## Quantized layers
41
 
@@ -43,24 +46,28 @@ Tested on
43
  ```json
44
  {
45
  "format": "comfy_quant",
46
- "block_names": [""],
47
  "rules": [
48
- { "policy": "keep", "match": ["blocks.0", "blocks.27"] },
49
- { "policy": "float8_e4m3fn", "match": ["q_proj", "k_proj", "v_proj", "o_proj", "output_proj", "mlp"] },
50
  { "policy": "nvfp4", "match": [] }
51
  ]
52
  }
53
  ```
54
 
55
- ### fp8mixed
56
  ```json
57
  {
58
  "format": "comfy_quant",
59
  "block_names": ["net.blocks."],
60
  "rules": [
61
- { "policy": "keep", "match": ["blocks.0.", "blocks.1.", "blocks.27.", "adaln_modulation", "v_proj"] },
62
- { "policy": "float8_e4m3fn", "match": ["q_proj", "k_proj", "output_proj", "mlp"] },
63
- { "policy": "nvfp4", "match": [] }
 
 
 
 
64
  ]
65
  }
66
  ```
 
9
 
10
  # FP8 Quantized model of ANIMA
11
 
12
+ ## !! I hava changed models recently. Please redownload if hash is different. !!
13
 
14
+ There are two models - FP8 and NVFP4Mixed.
15
+
16
+ - FP8 : (***recommend***) maximize generation speed while preserving quality as much as possible.
17
+ - NVFP4Mixed : (***marginal quality***) Mixture of FP8 and NVFP4.
18
 
19
 
20
  ## Generation speed
21
 
22
  Tested on
23
+ - RTX5090 (400W), ComfyUI with `--fast`option, torch2.10.0+cu130
24
  - Generates 832x1216, 30steps, cfg 4.0, er sde, simple
25
 
26
+ | quant | none | sage+torch.compile |
27
+ |------------|----------------------|-----------------------|
28
+ | bf16 | 7.13s/4.21it/s | 5.16s/5.81it/s (+38%) |
29
+ | fp8 | 6.66s/4.50it/s (+11%)| **4.52s/6.64it/s (+58%)** |
30
+ | nvfp4mix | 6.37s/4.71it/s (+12%)| 4.99s/6.01it/s (+43%) |
31
+
 
 
32
 
33
 
34
  ## Sample
35
 
36
+ | quant | sample |
37
+ |------------|----------------------|
38
+ | bf16 | ![anima-preview-bf16](https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/MUYIxQjZZxX5wGGwPHS3G.webp) |
39
+ | fp8 | ![anima-preview-fp8](https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/b09IufBT31yDxg_BRZR3w.webp) |
40
+ | nvfp4mixed | ![anima-preview-nvfp4](https://cdn-uploads.huggingface.co/production/uploads/63fbf6951b4b1bd4e706fed1/FMcirkGeDMAdbzkniOARF.webp) |
41
+
42
 
43
  ## Quantized layers
44
 
 
46
  ```json
47
  {
48
  "format": "comfy_quant",
49
+ "block_names": ["net.blocks."],
50
  "rules": [
51
+ { "policy": "keep", "match": ["blocks.0", "blocks.1."] },
52
+ { "policy": "float8_e4m3fn", "match": ["q_proj", "k_proj", "v_proj", "o_proj", "output_proj", ".mlp"] },
53
  { "policy": "nvfp4", "match": [] }
54
  ]
55
  }
56
  ```
57
 
58
+ ### nvfp4mixed
59
  ```json
60
  {
61
  "format": "comfy_quant",
62
  "block_names": ["net.blocks."],
63
  "rules": [
64
+ { "policy": "keep", "match": ["blocks.0."] },
65
+ { "policy": "float8_e4m3fn", "match": [
66
+ "blocks.1.k_proj", "blocks.1.q_proj", "blocks.1.output_proj",
67
+ "blocks.27.k_proj", "blocks.27.q_proj", "blocks.27.output_proj",
68
+ "v_proj", "adaln_modulation", ".mlp"
69
+ ] },
70
+ { "policy": "nvfp4", "match": ["k_proj", "q_proj", "output_proj"] }
71
  ]
72
  }
73
  ```