Commit
•
3f92697
1
Parent(s):
0db5342
Add diffusers code loading example
Browse filesit already works out of the box with diffusers, just requires proper tagging and code examples
README.md
CHANGED
@@ -3,31 +3,65 @@ license: mit
|
|
3 |
language:
|
4 |
- en
|
5 |
library_name: diffusers
|
|
|
|
|
|
|
|
|
6 |
---
|
7 |
|
8 |
# Mann-E Turbo
|
9 |
|
10 |
This is a part of _Mann-E Community Edition_ models. This model is a basic implementation of [Mann-E](https://mann-e.com)'s original model (which is not Stable Diffusion anymore) as a LoRa for SDXL. Since the whole business of Mann-E started around SD ecosystem, we've decided to release our LoRa for SD users and people who enjoy using models locally!
|
11 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
## Considerations
|
13 |
|
14 |
- Model __IS__ capable of making accidental and intentional _NSFW_ material. So be careful while using the LoRa in presence of people who might be sensitive to this material.
|
15 |
- The model has tested with SDXL 1.0, SDXL Turbo and even SDXL Lightning models and it works perfectly.
|
16 |
-
- We've tested the model with
|
17 |
- DreamShaperXL, TurboVisioXL and RealVis were models with the best results. You may consider using them or merges/fine-tunes based on those!
|
18 |
- This LoRa is compatible with other LoRa's as well. Be careful to weight it exactly 1 to get results like what we're presenting here.
|
19 |
- DPM SDE++ Karras is the best scheduler. 6-10 steps for turbo models work perfectly, also CFG of 2.5 to 4.0 works perfectly!
|
20 |
|
21 |
-
##
|
|
|
|
|
|
|
22 |
|
23 |
-
|
24 |
-
|
25 |
-
|
|
|
26 |
|
27 |
-
|
|
|
|
|
|
|
28 |
|
29 |
-
|
|
|
30 |
|
31 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
|
33 |
-
|
|
|
|
3 |
language:
|
4 |
- en
|
5 |
library_name: diffusers
|
6 |
+
tags:
|
7 |
+
- lora
|
8 |
+
- text-to-image
|
9 |
+
base_model: "stabilityai/stable-diffusion-xl-base-1.0"
|
10 |
---
|
11 |
|
12 |
# Mann-E Turbo
|
13 |
|
14 |
This is a part of _Mann-E Community Edition_ models. This model is a basic implementation of [Mann-E](https://mann-e.com)'s original model (which is not Stable Diffusion anymore) as a LoRa for SDXL. Since the whole business of Mann-E started around SD ecosystem, we've decided to release our LoRa for SD users and people who enjoy using models locally!
|
15 |
|
16 |
+
## Sample Generations
|
17 |
+
|
18 |
+
<div style="text-align:center;">
|
19 |
+
|
20 |
+
<img src="./hf-1.png" width=512 height=512>
|
21 |
+
|
22 |
+
<img src="./hf-2.png" width=512 height=512>
|
23 |
+
|
24 |
+
<img src="./hf-3.png" width=512 height=512>
|
25 |
+
|
26 |
+
<img src="./hf-4.png" width=512 height=512>
|
27 |
+
|
28 |
+
</div>
|
29 |
+
|
30 |
## Considerations
|
31 |
|
32 |
- Model __IS__ capable of making accidental and intentional _NSFW_ material. So be careful while using the LoRa in presence of people who might be sensitive to this material.
|
33 |
- The model has tested with SDXL 1.0, SDXL Turbo and even SDXL Lightning models and it works perfectly.
|
34 |
+
- We've tested the model with A111 and `diffusers`. For ComfyUI, we may need contributions if it doesn't work.
|
35 |
- DreamShaperXL, TurboVisioXL and RealVis were models with the best results. You may consider using them or merges/fine-tunes based on those!
|
36 |
- This LoRa is compatible with other LoRa's as well. Be careful to weight it exactly 1 to get results like what we're presenting here.
|
37 |
- DPM SDE++ Karras is the best scheduler. 6-10 steps for turbo models work perfectly, also CFG of 2.5 to 4.0 works perfectly!
|
38 |
|
39 |
+
## Load with 🧨 diffusers
|
40 |
+
```py
|
41 |
+
from diffusers import DiffusionPipeline, DPMSolverSinglestepScheduler
|
42 |
+
import torch
|
43 |
|
44 |
+
pipe = DiffusionPipeline.from_pretrained(
|
45 |
+
"stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
|
46 |
+
).to("cuda")
|
47 |
+
#You can use other base models such as `Lykon/dreamshaper-xl-1-0`, `0x4f1f/TurboVisionXL-v3.2` or `SG161222/RealVisXL_V4.0`
|
48 |
|
49 |
+
pipe.load_lora_weights(
|
50 |
+
"mann-e/Mann-E_Turbo",
|
51 |
+
weight_name="manne_turbo.safetensors",
|
52 |
+
)
|
53 |
|
54 |
+
#This is equivalent to DPM++ SDE Karras, as noted in https://huggingface.co/docs/diffusers/main/en/api/schedulers/overview
|
55 |
+
pipe.scheduler = DPMSolverSinglestepScheduler.from_config(pipe.scheduler.config, use_karras_sigmas=True)
|
56 |
|
57 |
+
image = pipe(
|
58 |
+
prompt="a cat in a bustling middle eastern city",
|
59 |
+
num_inference_steps=8,
|
60 |
+
guidance_scale=4,
|
61 |
+
width=768,
|
62 |
+
height=768,
|
63 |
+
clip_skip=1
|
64 |
+
).images[0]
|
65 |
|
66 |
+
image.save("a_cat.png")
|
67 |
+
```
|