Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Free memory before doing tiled decode.

comfyanonymous2865f913f754
diff --git a/comfy/sd.py b/comfy/sd.py
index 7e76f6fa1..95fc6d271 100644
--- a/comfy/sd.py
+++ b/comfy/sd.py
@@ -362,7 +362,8 @@ class VAE:
return pixel_samples
def decode_tiled(self, samples, tile_x=None, tile_y=None, overlap=None):
- model_management.load_model_gpu(self.patcher)
+ memory_used = self.memory_used_decode(samples.shape, self.vae_dtype) #TODO: calculate mem required for tile
+ model_management.load_models_gpu([self.patcher], memory_required=memory_used)
dims = samples.ndim - 2
args = {}
if tile_x is not None: