Merge pull request #6 from hodanov/feature/modify_to_add_meta_datas

Modify to add seed to filename.

.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

@@ -6,5 +6,4 @@ run:
 	--height 768 \
 	--width 512 \
 	--samples 5 \
-	--steps 50 \
-	--seed 500
+	--steps 50

sd_cli.py

@@ -75,23 +75,31 @@ 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",
         )
 
-        self.pipe = diffusers.StableDiffusionPipeline.from_pretrained(
-            cache_path,
-            scheduler=scheduler,
-            vae=vae,
-            custom_pipeline="lpw_stable_diffusion",
-            torch_dtype=torch.float16,
-        ).to("cuda")
+        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,
+                vae=vae,
+                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()
@@ -213,9 +221,6 @@ def entrypoint(
     gets back a list of images and outputs images to local.
     """
 
-    if seed == -1:
-        seed = util.generate_seed()
-
     inputs: dict[str, int | str] = {
         "prompt": prompt,
         "n_prompt": n_prompt,
@@ -233,9 +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, inputs, 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

@@ -4,8 +4,6 @@ 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")
 
@@ -65,13 +63,13 @@ def count_token(p: str, n: str) -> int:
     return max_embeddings_multiples
 
 
-def save_images(directory: Path, images: list[bytes], inputs: dict[str, int | str], 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}_{inputs['seed']}_{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)