switch to distilled

app.py

@@ -15,10 +15,13 @@ import mediapipe as mp
 from PIL import Image
 import cv2
 
+
 dtype = torch.bfloat16 
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
-pipeline = LTXConditionPipeline.from_pretrained("Lightricks/LTX-Video-0.9.7-dev", torch_dtype=dtype)
+#pipeline = LTXConditionPipeline.from_pretrained("Lightricks/LTX-Video-0.9.7-dev", torch_dtype=dtype)
+pipeline = LTXConditionPipeline.from_pretrained("Lightricks/LTX-Video-0.9.7-distilled", torch_dtype=torch.bfloat16)
+
 pipe_upsample = LTXLatentUpsamplePipeline.from_pretrained("Lightricks/ltxv-spatial-upscaler-0.9.7", vae=pipeline.vae, torch_dtype=dtype)
 pipeline.to(device)
 pipe_upsample.to(device)
@@ -197,12 +200,8 @@ def generate_video(
     negative_prompt="worst quality, inconsistent motion, blurry, jittery, distorted",
     height=768,
     width=1152,
-    num_inference_steps=30,
-    guidance_scale=5.0,
-    guidance_rescale=0.7,
-    decode_timestep=0.05,
-    decode_noise_scale=0.025,
-    image_cond_noise_scale=0.0,
+    num_inference_steps=7,
+    guidance_scale=1.0,
     seed=0,
     randomize_seed=False,
     progress=gr.Progress()
@@ -263,11 +262,11 @@ def generate_video(
             height=downscaled_height,
             num_frames=num_frames,
             num_inference_steps=num_inference_steps,
-            decode_timestep=decode_timestep,
-            decode_noise_scale=decode_noise_scale,
-            image_cond_noise_scale=image_cond_noise_scale,
+            decode_timestep=0.05,
+            decode_noise_scale=0.025,
+            # image_cond_noise_scale=image_cond_noise_scale,
             guidance_scale=guidance_scale,
-            guidance_rescale=guidance_rescale,
+            # guidance_rescale=guidance_rescale,
             generator=torch.Generator().manual_seed(seed),
             output_type="latent",
         ).frames
@@ -293,11 +292,11 @@ def generate_video(
             denoise_strength=0.4,
             num_inference_steps=10,
             latents=upscaled_latents,
-            decode_timestep=decode_timestep,
-            decode_noise_scale=decode_noise_scale,
-            image_cond_noise_scale=image_cond_noise_scale,
+            decode_timestep = 0.05,
             guidance_scale=guidance_scale,
-            guidance_rescale=guidance_rescale,
+            decode_noise_scale = 0.025,
+            image_cond_noise_scale=0.025,
+            #guidance_rescale=guidance_rescale,
             generator=torch.Generator(device="cuda").manual_seed(seed),
             output_type="pil",
         ).frames[0]
@@ -392,7 +391,7 @@ with gr.Blocks() as demo:
                     minimum=10,
                     maximum=50,
                     step=1,
-                    value=30
+                    value=7
                 )
                 
                 with gr.Row():
@@ -401,43 +400,10 @@ with gr.Blocks() as demo:
                         minimum=1.0,
                         maximum=15.0,
                         step=0.1,
-                        value=5.0
-                    )
-                    guidance_rescale = gr.Slider(
-                        label="Guidance Rescale",
-                        minimum=0.0,
-                        maximum=1.0,
-                        step=0.05,
-                        value=0.7,
-                        visible=False
-                    )
-                
-                with gr.Row():
-                    decode_timestep = gr.Slider(
-                        label="Decode Timestep",
-                        minimum=0.0,
-                        maximum=1.0,
-                        step=0.01,
-                        value=0.05,
-                        visible=False
-                    )
-                    decode_noise_scale = gr.Slider(
-                        label="Decode Noise Scale",
-                        minimum=0.0,
-                        maximum=0.1,
-                        step=0.005,
-                        value=0.025,
-                        visible=False
+                        value=1.0
                     )
+                    
                 
-                image_cond_noise_scale = gr.Slider(
-                    label="Image Condition Noise Scale",
-                    minimum=0.0,
-                    maximum=0.5,
-                    step=0.01,
-                    value=0.0,
-                    visible=False
-                )
                 
                 with gr.Row():
                     randomize_seed = gr.Checkbox(
@@ -482,10 +448,6 @@ with gr.Blocks() as demo:
             width,
             num_inference_steps,
             guidance_scale,
-            guidance_rescale,
-            decode_timestep,
-            decode_noise_scale,
-            image_cond_noise_scale,
             seed,
             randomize_seed
         ],