Update app_fast.py

app_fast.py

@@ -2,6 +2,7 @@ import spaces
 import torch
 from diffusers import AutoencoderKLWan, WanPipeline, WanImageToVideoPipeline, UniPCMultistepScheduler
 from diffusers.utils import export_to_video
+from diffusers.hooks import apply_first_block_cache, FirstBlockCacheConfig
 import gradio as gr
 import tempfile
 from huggingface_hub import hf_hub_download
@@ -11,6 +12,7 @@ import random
 
 MODEL_ID = "FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers"
 vae = AutoencoderKLWan.from_pretrained(MODEL_ID, subfolder="vae", torch_dtype=torch.float32)
+#vae = AutoencoderKLWan.from_pretrained("Kijai/WanVideo_comfy", filename="Wan2_2_VAE_bf16.safetensors", torch_dtype=torch.bfloat16)
 
 # Initialize pipelines
 text_to_video_pipe = WanPipeline.from_pretrained(MODEL_ID, vae=vae, torch_dtype=torch.bfloat16)
@@ -20,7 +22,7 @@ for pipe in [text_to_video_pipe, image_to_video_pipe]:
     pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=5.0)
     pipe.to("cuda")
 
-
+apply_first_block_cache(pipe.transformer, FirstBlockCacheConfig(threshold=0.2))
 
 
 ##Lora testing
@@ -42,7 +44,7 @@ for pipe in [text_to_video_pipe, image_to_video_pipe]:
 #LORA_FILENAME = "HSToric_color_Wan22_5b_LoRA.safetensors"
 
 LORA_REPO_ID = "AlekseyCalvin/Phenakistiscopes_Wan22_5B_T2V_LoRA"
-LORA_FILENAME = "Phenakistiscopes_MergedVers4_3_2_wan5b_wEMAsigmaRel020.safetensors"
+LORA_FILENAME = "Phenakistiscopes_V3_Wan5b_PowerEMA_SigmaRel013.safetensors"
 causvid_path = hf_hub_download(repo_id=LORA_REPO_ID, filename=LORA_FILENAME)
 pipe.load_lora_weights(causvid_path, adapter_name="wan_lora")
 pipe.set_adapters(["wan_lora"], adapter_weights=[1.05])
@@ -61,7 +63,7 @@ MIN_FRAMES_MODEL = 17
 MAX_FRAMES_MODEL = 193
 
 default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
-default_negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards, watermark, text, signature"
+default_negative_prompt = "dull, overexposed, flashing, stuttering, static, blurred, vapid, banal, static, overall gray, worst, low, JPEG compression residue, incomplete, extra,  error, missing, vanishing, lapse, broken, wrong, deformed, disfigured, misshapen, fused fingers, still, messy, watermark"
 
 def _calculate_new_dimensions_wan(pil_image, mod_val, calculation_max_area, min_slider_h, max_slider_h, min_slider_w, max_slider_w, default_h, default_w):
     orig_w, orig_h = pil_image.size
@@ -99,13 +101,13 @@ def get_duration(input_image, prompt, height, width,
                    seed, randomize_seed, 
                    progress):
     if steps > 5 and duration_seconds > 5:
-        return 30
+        return 50
     elif steps > 5 or duration_seconds > 5:
-        return 25
+        return 45
     else:
-        return 20
+        return 40
 
-@spaces.GPU(duration=get_duration)
+@spaces.GPU
 def generate_video(input_image, prompt, height, width, negative_prompt=default_negative_prompt, duration_seconds=2, guidance_scale=0, steps=4, seed=44, randomize_seed=False, progress=gr.Progress(track_tqdm=True)):
     target_h = max(MOD_VALUE, (int(height) // MOD_VALUE) * MOD_VALUE)
     target_w = max(MOD_VALUE, (int(width) // MOD_VALUE) * MOD_VALUE)