up

run_bug_4297.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-from diffusers import DiffusionPipeline
+from diffusers import DiffusionPipeline, AutoencoderKL
 import torch
 torch.backends.cudnn.deterministic = False
 torch.backends.cuda.matmul.allow_tf32 = False
@@ -7,7 +7,8 @@ torch.backends.cudnn.allow_tf32 = False
 torch.backends.cudnn.benchmark = True
 torch.backends.cuda.enable_flash_sdp(False)
 
-base_pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-0.9", torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
+vae = AutoencoderKL.from_pretrained("stabilityai/sdxl-vae", torch_dtype=torch.float16)
+base_pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", vae=vae, torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
 base_pipe.to("cuda") # OR, pipe.enable_sequential_cpu_offload() OR,
 #pipe.enable_model_cpu_offload()
 
@@ -17,12 +18,11 @@ base_pipe.to("cuda") # OR, pipe.enable_sequential_cpu_offload() OR,
 # Reproducibility.
 torch_seed = 4202420420
 refiner_seed = 698008569
-refiner_strength = 0.50
 prompt = "happy child flying a kite on a sunny day"
 negative_prompt = ''
 # Batch size.
 batch_size = 2
-do_latent = False
+do_latent = True
 prompt = [ prompt ] * batch_size
 negative_prompt = [ negative_prompt ] * batch_size
 # We're going to schedule 20 steps, and complete 50% of them using either model.
@@ -30,9 +30,9 @@ total_num_steps = 20
 # We need multiple Generators.
 generator = [ torch.Generator(device="cuda").manual_seed(torch_seed) ] * batch_size
 
-pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-refiner-0.9", vae=base_pipe.vae, text_encoder_2=base_pipe.text_encoder_2, torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
+pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-refiner-1.0", vae=base_pipe.vae, text_encoder_2=base_pipe.text_encoder_2, torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
 # Using channels last layout.
-pipe.unet.to(memory_format=torch.channels_last)
+# pipe.unet.to(memory_format=torch.channels_last)
 pipe.to("cuda") # OR, pipe.enable_sequential_cpu_offload() OR, 
 
 # Generate the base image.

run_bug_4297_new.py

@@ -0,0 +1,49 @@
+#!/usr/bin/env python3
+from diffusers import DiffusionPipeline
+import torch
+torch.backends.cudnn.deterministic = False
+torch.backends.cuda.matmul.allow_tf32 = False
+torch.backends.cudnn.allow_tf32 = False
+torch.backends.cudnn.benchmark = True
+torch.backends.cuda.enable_flash_sdp(False)
+
+# vae = AutoEncoderKL.from_pretrained("stabilityai/sdxl-vae", torch_dtype=torch.float16)
+# base_pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", vae=vae, torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
+base_pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", vae, torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
+base_pipe.to("cuda") # OR, pipe.enable_sequential_cpu_offload() OR,
+
+# Reproducibility.
+torch_seed = 4202420420
+refiner_seed = 698008569
+refiner_strength = 0.50
+prompt = "happy child flying a kite on a sunny day"
+negative_prompt = ''
+# Batch size.
+batch_size = 2
+do_latent = True
+
+# We're going to schedule 20 steps, and complete 50% of them using either model.
+total_num_steps = 20
+# We need multiple Generators.
+generator = torch.Generator(device="cuda").manual_seed(torch_seed)
+
+pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-refiner-1.0", vae=base_pipe.vae, text_encoder_2=base_pipe.text_encoder_2, torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
+# Using channels last layout.
+pipe.unet.to(memory_format=torch.channels_last)
+pipe.to("cuda") # OR, pipe.enable_sequential_cpu_offload() OR, 
+
+# Generate the base image.
+pre_image = base_pipe(prompt=prompt, generator=generator,
+        num_inference_steps=total_num_steps, negative_prompt=negative_prompt, num_images_per_prompt=batch_size, output_type="latent" if do_latent else "pil").images
+
+# Generate a range from 0.1 to 0.9, with 0.1 increments.
+test_strengths = [0.5]
+for refiner_strength in test_strengths:
+    # Generate a new set of random states for each image.
+    generator_two = torch.Generator(device="cuda").manual_seed(refiner_seed)
+    # Put through the refiner now.
+    images = pipe(prompt=prompt, image=pre_image, aesthetic_score=10, negative_aesthetic_score=2.4, generator=generator_two,
+                num_inference_steps=total_num_steps, num_images_per_prompt=batch_size, strength=refiner_strength, negative_prompt=negative_prompt).images # denoising_start
+    for idx in range(0, len(images)):
+        print(f'Image: {idx}')
+        images[idx].save(f'/home/patrick/images/refiner_bug/test-{refiner_strength}-{idx}--{batch_size}--{do_latent}.png', format='PNG')