Try inline pipeline to reduce memory usage.

app.py

@@ -62,7 +62,7 @@ class DropOutput(Callback):
 #     is_mega=True,
 #     is_reusable=True
 # )
-pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", custom_pipeline="stable_diffusion_mega", torch_dtype=torch.float32, use_auth_token=AUTH_TOKEN)
+# pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", custom_pipeline="stable_diffusion_mega", torch_dtype=torch.float32, use_auth_token=AUTH_TOKEN)
 
 # pipeline.to("cuda")
 
@@ -73,6 +73,9 @@ def gen_image(prompt):
     # Hugging Space faces seems to run out of memory if grads are not disabled
     # torch.set_grad_enabled(False)
     print(f'RUNNING gen_image with prompt: {prompt}')
+    pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5",
+                                                 custom_pipeline="stable_diffusion_mega", torch_dtype=torch.float32,
+                                                 use_auth_token=AUTH_TOKEN)
     images = pipeline.text2img(prompt, width=512, height=512).images
     # images = model.generate_images(
     #     text=prompt,