Create LiteVisionPipeline.py
Browse files- LiteVisionPipeline.py +44 -0
LiteVisionPipeline.py
ADDED
|
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from diffusers import StableDiffusionPipeline, LCMScheduler
|
| 3 |
+
|
| 4 |
+
|
| 5 |
+
class LiteVisionPipeline(StableDiffusionPipeline):
|
| 6 |
+
"""
|
| 7 |
+
LiteVisionPipeline v1
|
| 8 |
+
- Custom pipeline for LiteVision models
|
| 9 |
+
- LCM scheduler preset
|
| 10 |
+
- Low VRAM & fast inference defaults
|
| 11 |
+
- Compatible with DiffusionPipeline ecosystem
|
| 12 |
+
"""
|
| 13 |
+
|
| 14 |
+
def __init__(self, *args, **kwargs):
|
| 15 |
+
super().__init__(*args, **kwargs)
|
| 16 |
+
|
| 17 |
+
# ---- Scheduler preset (LCM) ----
|
| 18 |
+
self.scheduler = LCMScheduler.from_config(self.scheduler.config)
|
| 19 |
+
|
| 20 |
+
# ---- Memory optimizations ----
|
| 21 |
+
self.enable_attention_slicing()
|
| 22 |
+
self.enable_vae_slicing()
|
| 23 |
+
|
| 24 |
+
@torch.inference_mode()
|
| 25 |
+
def __call__(
|
| 26 |
+
self,
|
| 27 |
+
prompt,
|
| 28 |
+
negative_prompt=None,
|
| 29 |
+
num_inference_steps: int = 6,
|
| 30 |
+
guidance_scale: float = 1.5,
|
| 31 |
+
**kwargs,
|
| 32 |
+
):
|
| 33 |
+
"""
|
| 34 |
+
LiteVision default generation call
|
| 35 |
+
- Optimized for LCM-style inference
|
| 36 |
+
"""
|
| 37 |
+
|
| 38 |
+
return super().__call__(
|
| 39 |
+
prompt=prompt,
|
| 40 |
+
negative_prompt=negative_prompt,
|
| 41 |
+
num_inference_steps=num_inference_steps,
|
| 42 |
+
guidance_scale=guidance_scale,
|
| 43 |
+
**kwargs,
|
| 44 |
+
)
|