v33ts commited on
Commit
4e779b3
·
verified ·
1 Parent(s): f56301f

Upload models/vae/sdxl_vae.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. models/vae/sdxl_vae.md +68 -0
models/vae/sdxl_vae.md ADDED
@@ -0,0 +1,68 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ author: nucleardiffusion
3
+ baseModel: SDXL 1.0
4
+ hashes:
5
+ AutoV1: 584F63AB
6
+ AutoV2: 235745AF8D
7
+ AutoV3: 55F20A1016E7
8
+ BLAKE3: D46EB6988A26BCAF9E9CFA5E5C6264C4EE1A70F2018F33B8BC2DD7CA0681B490
9
+ CRC32: 5BB75FB4
10
+ SHA256: 235745AF8D86BF4A4C1B5B4F529868B37019A10F7C0B2E79AD0ABCA3A22BC6E1
11
+ metadata:
12
+ format: SafeTensor
13
+ modelPage: https://civitai.com/models/140686?modelVersionId=155933
14
+ preview:
15
+ - https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/4596b7dc-4fd4-494c-bf8c-36d6a0ee89bc/width=450/2369135.jpeg
16
+ website: Civitai
17
+ ---
18
+
19
+ # Trigger Words
20
+
21
+ No trigger words
22
+
23
+ # About this version
24
+
25
+ No description about this version
26
+
27
+ # FIX FP16 Errors SDXL - Lower Memory use! --- sdxl-vae-fp16-fix by madebyollin
28
+
29
+ **"As good as SDXL VAE but runs twice as fast and uses significantly less memory."** <https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/discussions/7>
30
+
31
+ "Same license on [stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
32
+
33
+ same vae license on [sdxl-vae-fp16-fix](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix)
34
+
35
+ **Troubleshoot:**
36
+
37
+ **Do not use the refiner with VAE built in**
38
+ **Try launch param: --medvram --opt-split-attention --xformers**
39
+
40
+ SDXL-VAE-FP16-Fix is the [SDXL VAE](<https://huggingface.co/stabilityai/sdxl-vae>, but modified to run in fp16 precision without generating NaNs.
41
+
42
+ Details:
43
+
44
+ SDXL-VAE generates NaNs in fp16 because the internal activation values are too big:
45
+
46
+ SDXL-VAE-FP16-Fix was created by finetuning the SDXL-VAE to:
47
+
48
+ 1. keep the final output the same, but
49
+ 2. make the internal activation values smaller, by
50
+ 3. scaling down weights and biases within the network
51
+
52
+ There are slight discrepancies between the output of SDXL-VAE-FP16-Fix and SDXL-VAE, but the decoded images should be close enough for most purposes." - [bdsqlsz](https://www.reddit.com/user/bdsqlsz/)
53
+
54
+ ---
55
+
56
+ **NOT MY WORK - REUPLOADED HERE FOR EASE OF USE**
57
+
58
+ ---
59
+
60
+ **COMMISSIONS NOW ACCEPTED!**
61
+ =============================
62
+
63
+ **I have been away saving to upgrade my pc, please help fund more work:**
64
+ -------------------------------------------------------------------------
65
+
66
+ [**https://www.patreon.com/nucleardiffusion**](https://www.patreon.com/nucleardiffusion)
67
+
68
+ <https://ko-fi.com/nucleardiffusion>