commit

app.py

@@ -25,7 +25,7 @@ from lxml.html import fromstring
 from diffusers.utils import export_to_video, load_image
 from huggingface_hub import hf_hub_download
 from safetensors.torch import load_file, save_file
-from diffusers import FluxPipeline, CogVideoXImageToVideoPipeline
+from diffusers import FluxPipeline, HunyuanVideoPipeline, HunyuanVideoTransformer3DModel
 from PIL import Image, ImageDraw, ImageFont
 
 # logging
@@ -43,10 +43,8 @@ root.addHandler(handler)
 
 if torch.cuda.is_available():
     device = "cuda"
-    dtype = torch.bfloat16
 else:
     device = "cpu"
-    dtype = torch.bfloat16
 
 base = "black-forest-labs/FLUX.1-schnell"
 
@@ -57,13 +55,13 @@ additional_image = None
 # precision data
 
 seq=512
-fps=12
+fps=15
 width=768
-height=768
+height=512
 image_steps=8
 video_steps=50
 img_accu=9
-vid_accu=6
+vid_accu=7
 
 # ui data
 
@@ -126,10 +124,13 @@ function custom(){
 
 # torch pipes
 
-image_pipe = FluxPipeline.from_pretrained(base, torch_dtype=dtype).to(device)
-video_pipe = CogVideoXImageToVideoPipeline.from_pretrained(
-    "THUDM/CogVideoX-5b-I2V",
-    torch_dtype=dtype
+image_pipe = FluxPipeline.from_pretrained(base, torch_dtype=torch.float32).to(device)
+
+transformer = HunyuanVideoTransformer3DModel.from_pretrained(
+    "hunyuanvideo-community/HunyuanVideo", subfolder="transformer", torch_dtype=torch.bfloat16
+)
+video_pipe = HunyuanVideoPipeline.from_pretrained(
+    "hunyuanvideo-community/HunyuanVideo", transformer=transformer, torch_dtype=torch.bfloat16
 ).to(device)
 
 video_pipe.vae.enable_tiling()
@@ -246,12 +247,12 @@ def handle_generate(*_inp):
     if len(inp[2]) >= 2:
         inp[2] = "," + inp[2].strip(",").strip(" ")
 
-    inp[2] = f"textual content,unrealistic content,divined creatures,unrealistic creatures,creatures out of this world,demon,angel,cgi quality,anime quality,cartoon quality,drawing quality,cropped photo,cropped content,worst quality,low quality,duplicating elements,weird,non-standard human body,non-standard object structure,blur,wrong body anatomy,too big, too small,text,written content{inp[2]}"
+    inp[2] = f"textual,labeled,divined,distorted,discontinuous,ugly,blurry,low resolution,motionless,static,wrong body anatomy{inp[2]}"
 
     if len(inp[1]) >= 2:
         inp[1] = "," + inp[1].strip(",").strip(" ")
     
-    inp[1] = f'looks real,feels real,similar to real photographs,dark natural colors,looks beautiful and pretty,look genuine and authentic,reasonable logic,natural,masterpiece,highly detailed{inp[1]}'
+    inp[1] = f'realistic,natural,genuine,reasonable,highly detailed{inp[1]}'
 
     print(f"""