Update README.md
Browse files
README.md
CHANGED
|
@@ -70,29 +70,53 @@ export_to_video(output, "output.mp4", fps=24, quality=9)
|
|
| 70 |
import torch
|
| 71 |
from diffusers import Kandinsky5T2VPipeline
|
| 72 |
|
| 73 |
-
# SFT
|
| 74 |
pipe_sft = Kandinsky5T2VPipeline.from_pretrained(
|
| 75 |
"ai-forever/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers",
|
| 76 |
torch_dtype=torch.bfloat16
|
| 77 |
)
|
| 78 |
|
| 79 |
-
# Distilled 16-step model (fastest)
|
| 80 |
pipe_distill = Kandinsky5T2VPipeline.from_pretrained(
|
| 81 |
"ai-forever/Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers",
|
| 82 |
torch_dtype=torch.bfloat16
|
| 83 |
)
|
| 84 |
|
| 85 |
-
# No-CFG model (balanced speed/quality)
|
| 86 |
pipe_nocfg = Kandinsky5T2VPipeline.from_pretrained(
|
| 87 |
"ai-forever/Kandinsky-5.0-T2V-Lite-nocfg-5s-Diffusers",
|
| 88 |
torch_dtype=torch.bfloat16
|
| 89 |
)
|
| 90 |
|
| 91 |
-
# Pretrain model (most diverse)
|
| 92 |
pipe_pretrain = Kandinsky5T2VPipeline.from_pretrained(
|
| 93 |
"ai-forever/Kandinsky-5.0-T2V-Lite-pretrain-5s-Diffusers",
|
| 94 |
torch_dtype=torch.bfloat16
|
| 95 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 96 |
```
|
| 97 |
|
| 98 |
## Architecture
|
|
|
|
| 70 |
import torch
|
| 71 |
from diffusers import Kandinsky5T2VPipeline
|
| 72 |
|
| 73 |
+
# 5s SFT model (highest quality)
|
| 74 |
pipe_sft = Kandinsky5T2VPipeline.from_pretrained(
|
| 75 |
"ai-forever/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers",
|
| 76 |
torch_dtype=torch.bfloat16
|
| 77 |
)
|
| 78 |
|
| 79 |
+
# 5s Distilled 16-step model (fastest)
|
| 80 |
pipe_distill = Kandinsky5T2VPipeline.from_pretrained(
|
| 81 |
"ai-forever/Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers",
|
| 82 |
torch_dtype=torch.bfloat16
|
| 83 |
)
|
| 84 |
|
| 85 |
+
# 5s No-CFG model (balanced speed/quality)
|
| 86 |
pipe_nocfg = Kandinsky5T2VPipeline.from_pretrained(
|
| 87 |
"ai-forever/Kandinsky-5.0-T2V-Lite-nocfg-5s-Diffusers",
|
| 88 |
torch_dtype=torch.bfloat16
|
| 89 |
)
|
| 90 |
|
| 91 |
+
# 5s Pretrain model (most diverse)
|
| 92 |
pipe_pretrain = Kandinsky5T2VPipeline.from_pretrained(
|
| 93 |
"ai-forever/Kandinsky-5.0-T2V-Lite-pretrain-5s-Diffusers",
|
| 94 |
torch_dtype=torch.bfloat16
|
| 95 |
)
|
| 96 |
+
|
| 97 |
+
# 10s SFT model (highest quality)
|
| 98 |
+
pipe_sft = Kandinsky5T2VPipeline.from_pretrained(
|
| 99 |
+
"ai-forever/Kandinsky-5.0-T2V-Lite-sft-10s-Diffusers",
|
| 100 |
+
torch_dtype=torch.bfloat16
|
| 101 |
+
)
|
| 102 |
+
|
| 103 |
+
# 10s Distilled 16-step model (fastest)
|
| 104 |
+
pipe_distill = Kandinsky5T2VPipeline.from_pretrained(
|
| 105 |
+
"ai-forever/Kandinsky-5.0-T2V-Lite-distilled16steps-10s-Diffusers",
|
| 106 |
+
torch_dtype=torch.bfloat16
|
| 107 |
+
)
|
| 108 |
+
|
| 109 |
+
# 10s No-CFG model (balanced speed/quality)
|
| 110 |
+
pipe_nocfg = Kandinsky5T2VPipeline.from_pretrained(
|
| 111 |
+
"ai-forever/Kandinsky-5.0-T2V-Lite-nocfg-10s-Diffusers",
|
| 112 |
+
torch_dtype=torch.bfloat16
|
| 113 |
+
)
|
| 114 |
+
|
| 115 |
+
# 10s Pretrain model (most diverse)
|
| 116 |
+
pipe_pretrain = Kandinsky5T2VPipeline.from_pretrained(
|
| 117 |
+
"ai-forever/Kandinsky-5.0-T2V-Lite-pretrain-10s-Diffusers",
|
| 118 |
+
torch_dtype=torch.bfloat16
|
| 119 |
+
)
|
| 120 |
```
|
| 121 |
|
| 122 |
## Architecture
|