num_inference_steps for animtatediff

dkackman · Mar 23, 2024 · 08f34de · 08f34de
1 parent d8ed7f7
commit 08f34de
Showing 2 changed files with 4 additions and 3 deletions.
diff --git a/swarm/test.py b/swarm/test.py
@@ -184,7 +184,7 @@
     "prompt": "A dancing marmot, 4k, high resolution",
     "negative_prompt": "bad quality, worse quality, low resolution",
     "workflow": "txt2vid",
-    "num_inference_steps": 8,
+    "num_inference_steps": 6,
     "guidance_scale": 2.0,
     "outputs": ["primary"],
     "num_frames": 32,
@@ -196,9 +196,9 @@
         "motion_adapter":
         {
             "model_name": "ByteDance/AnimateDiff-Lightning",
-            "checkpoint_file": "animatediff_lightning_8step_diffusers.safetensors",
+            "num_inference_steps": 4,
+            "checkpoint_file": "animatediff_lightning_4step_diffusers.safetensors",
         },
-
         "scheduler_args": {
             "scheduler_type": "EulerDiscreteScheduler",
             "beta_schedule": "linear",

diff --git a/swarm/video/tx2vid.py b/swarm/video/tx2vid.py
@@ -29,6 +29,7 @@ def txt2vid_diffusion_callback(device_identifier, model_name, **kwargs):
         if "checkpoint_file" in motion_adapter_args:
             motion_adapter = MotionAdapter()
             motion_adapter.load_state_dict(load_file(hf_hub_download(motion_adapter_args["model_name"], motion_adapter_args["checkpoint_file"])))
+            kwargs["num_inference_steps"] = motion_adapter_args.pop("num_inference_steps", 4)
         else:
             motion_adapter = MotionAdapter.from_pretrained(motion_adapter_args["model_name"], torch_dtype=torch_dtype)