Merge pull request #5 from hodanov/feature/modify_to_use_seed

Modify to use seed.

.env.example

@@ -1,3 +1,4 @@
 HUGGING_FACE_TOKEN=""
 MODEL_REPO_ID="stabilityai/stable-diffusion-2-1"
 MODEL_NAME="stable-diffusion-2-1"
+USE_VAE="false"

Makefile

@@ -2,8 +2,8 @@ run:
 	modal run sd_cli.py \
 	--prompt "A woman with bob hair" \
 	--n-prompt "" \
+	--upscaler "RealESRGAN_x4plus_anime_6B" \
 	--height 768 \
 	--width 512 \
 	--samples 5 \
-	--steps 50 \
-	--upscaler "RealESRGAN_x4plus_anime_6B"
+	--steps 50

sd_cli.py

@@ -75,16 +75,16 @@ class StableDiffusion:
 
         torch.backends.cuda.matmul.allow_tf32 = True
 
-        vae = diffusers.AutoencoderKL.from_pretrained(
-            cache_path,
-            subfolder="vae",
-        )
-
         scheduler = diffusers.EulerAncestralDiscreteScheduler.from_pretrained(
             cache_path,
             subfolder="scheduler",
         )
 
+        if os.environ["USE_VAE"] == "true":
+            vae = diffusers.AutoencoderKL.from_pretrained(
+                cache_path,
+                subfolder="vae",
+            )
             self.pipe = diffusers.StableDiffusionPipeline.from_pretrained(
                 cache_path,
                 scheduler=scheduler,
@@ -92,6 +92,14 @@ class StableDiffusion:
                 custom_pipeline="lpw_stable_diffusion",
                 torch_dtype=torch.float16,
             ).to("cuda")
+        else:
+            self.pipe = diffusers.StableDiffusionPipeline.from_pretrained(
+                cache_path,
+                scheduler=scheduler,
+                custom_pipeline="lpw_stable_diffusion",
+                torch_dtype=torch.float16,
+            ).to("cuda")
+
         self.pipe.enable_xformers_memory_efficient_attention()
 
     @method()
@@ -101,6 +109,7 @@ class StableDiffusion:
         """
         import torch
 
+        generator = torch.Generator("cuda").manual_seed(inputs["seed"])
         with torch.inference_mode():
             with torch.autocast("cuda"):
                 base_images = self.pipe(
@@ -111,6 +120,7 @@ class StableDiffusion:
                     num_inference_steps=inputs["steps"],
                     guidance_scale=7.5,
                     max_embeddings_multiples=inputs["max_embeddings_multiples"],
+                    generator=generator,
                 ).images
 
         if inputs["upscaler"] != "":
@@ -197,12 +207,13 @@ class StableDiffusion:
 def entrypoint(
     prompt: str,
     n_prompt: str,
+    upscaler: str,
     height: int = 512,
     width: int = 512,
     samples: int = 5,
     batch_size: int = 1,
     steps: int = 20,
-    upscaler: str = "",
+    seed: int = -1,
 ):
     """
     This function is the entrypoint for the Runway CLI.
@@ -219,7 +230,7 @@ def entrypoint(
         "batch_size": batch_size,
         "steps": steps,
         "upscaler": upscaler,
-        # seed=-1
+        "seed": seed,
     }
 
     inputs["max_embeddings_multiples"] = util.count_token(p=prompt, n=n_prompt)
@@ -227,9 +238,11 @@ def entrypoint(
 
     sd = StableDiffusion()
     for i in range(samples):
+        if seed == -1:
+            inputs["seed"] = util.generate_seed()
         start_time = time.time()
         images = sd.run_inference.call(inputs)
-        util.save_images(directory, images, i)
+        util.save_images(directory, images, int(inputs["seed"]), i)
         total_time = time.time() - start_time
         print(f"Sample {i} took {total_time:.3f}s ({(total_time)/len(images):.3f}s / image).")
 

util.py

@@ -1,14 +1,23 @@
 """ Utility functions for the script. """
+import random
 import time
 from datetime import date
 from pathlib import Path
 
-from PIL import Image
-
 OUTPUT_DIRECTORY = "outputs"
 DATE_TODAY = date.today().strftime("%Y-%m-%d")
 
 
+def generate_seed() -> int:
+    """
+    Generate a random seed.
+    """
+    seed = random.randint(0, 4294967295)
+    print(f"Generate a random seed: {seed}")
+
+    return seed
+
+
 def make_directory() -> Path:
     """
     Make a directory for saving outputs.
@@ -16,7 +25,7 @@ def make_directory() -> Path:
     directory = Path(f"{OUTPUT_DIRECTORY}/{DATE_TODAY}")
     if not directory.exists():
         directory.mkdir(exist_ok=True, parents=True)
-        print(f"Make directory: {directory}")
+        print(f"Make a directory: {directory}")
 
     return directory
 
@@ -54,13 +63,13 @@ def count_token(p: str, n: str) -> int:
     return max_embeddings_multiples
 
 
-def save_images(directory: Path, images: list[bytes], i: int):
+def save_images(directory: Path, images: list[bytes], seed: int, i: int):
     """
     Save images to a file.
     """
     for j, image_bytes in enumerate(images):
         formatted_time = time.strftime("%Y%m%d%H%M%S", time.localtime(time.time()))
-        output_path = directory / f"{formatted_time}_{i}_{j}.png"
+        output_path = directory / f"{formatted_time}_{seed}_{i}_{j}.png"
         print(f"Saving it to {output_path}")
         with open(output_path, "wb") as file:
             file.write(image_bytes)