Create main.py

main.py

@@ -0,0 +1,21 @@
+import torch
+from diffusers import FluxPipeline
+
+pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16) # can replace schnell with dev
+# to run on low vram GPUs (i.e. between 4 and 32 GB VRAM)
+pipe.enable_sequential_cpu_offload()
+pipe.vae.enable_slicing()
+pipe.vae.enable_tiling()
+
+pipe.to(torch.float16) # casting here instead of in the pipeline constructor because doing so in the constructor loads all models into CPU memory at once
+
+prompt = "A cat holding a sign that says hello world"
+out = pipe(
+    prompt=prompt,
+    guidance_scale=4,
+    height=768,
+    width=1024,
+    num_inference_steps=4,
+    max_sequence_length=256,
+).images[0]
+out.save("image.png")