Update app.py

app.py

@@ -34,7 +34,8 @@ dtype = torch.float16
 pipe = AnimateDiffPipeline.from_pretrained(bases[base_loaded], torch_dtype=dtype).to(device)
 pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing", beta_schedule="linear")
 
-step = 2
+# unfortunately 2 steps isn't good enough for AiTube, we need 4 steps
+step = 4
 repo = "ByteDance/AnimateDiff-Lightning"
 ckpt = f"animatediff_lightning_{step}step_diffusers.safetensors"
 pipe.unet.load_state_dict(load_file(hf_hub_download(repo, ckpt), device=device), strict=False)
@@ -44,10 +45,12 @@ step_loaded = step
 helper = DeepCacheSDHelper(pipe=pipe)
 helper.set_params(
     # cache_interval means the frequency of feature caching, specified as the number of steps between each cache operation.
+    # with AnimateDiff this seems to have large effects, so we cannot use large values,
+    # even with cache_interval=3 I notice a big degradation in quality
     cache_interval=2,
 
     # cache_branch_id identifies which branch of the network (ordered from the shallowest to the deepest layer) is responsible for executing the caching processes.
-    # Note Julian: I have tried cache_branch_id=0 but quality was very "smoothed out"
+    # Note Julian: I should create my own benchmarks for this
     cache_branch_id=0,
 
     # Opting for a lower cache_branch_id or a larger cache_interval can lead to faster inference speed at the expense of reduced image quality