add seed and latents stuff

api/serve.py

@@ -12,18 +12,19 @@ from flask import Flask, make_response, request, send_file
 from stringcase import spinalcase
 from io import BytesIO
 from os import environ, path, makedirs
+import numpy as np
 
 # defaults
 default_prompt = "a photo of an astronaut eating a hamburger"
+default_cfg = 8
+default_steps = 20
 default_height = 512
 default_width = 512
-default_steps = 20
-default_cfg = 8
 
+max_cfg = 30
+max_steps = 150
 max_height = 512
 max_width = 512
-max_steps = 150
-max_cfg = 30
 
 # paths
 model_path = environ.get('ONNX_WEB_MODEL_PATH', "../../stable_diffusion_onnx")
@@ -50,6 +51,14 @@ def get_from_map(args, key, values, default):
   else:
     return values[default]
 
+def get_latents_from_seed(seed: int, width: int, height: int) -> np.ndarray:
+    # 1 is batch size
+  latents_shape = (1, 4, height // 8, width // 8)
+    # Gotta use numpy instead of torch, because torch's randn() doesn't support DML
+  rng = np.random.default_rng(seed)
+  image_latents = rng.standard_normal(latents_shape).astype(np.float32)
+  return image_latents
+
 # setup
 if not path.exists(model_path):
   raise RuntimeError('model path must exist')
@@ -68,14 +77,20 @@ def hello():
 def txt2img():
   user = request.remote_addr
 
-  cfg = get_and_clamp(request.args, 'cfg', default_cfg, max_cfg)
-  height = get_and_clamp(request.args, 'height', default_height, max_height)
   prompt = request.args.get('prompt', default_prompt)
-  steps = get_and_clamp(request.args, 'steps', default_steps, max_steps)
   scheduler = get_from_map(request.args, 'scheduler', scheduler_list, 'euler-a')
+  cfg = get_and_clamp(request.args, 'cfg', default_cfg, max_cfg, 0)
+  steps = get_and_clamp(request.args, 'steps', default_steps, max_steps)
+  height = get_and_clamp(request.args, 'height', default_height, max_height)
   width = get_and_clamp(request.args, 'width', default_width, max_width)
 
-  print("txt2img from %s: %s/%s, %sx%s, %s" % (user, cfg, steps, width, height, prompt))
+  seed = int(request.args.get('seed', -1))
+  if seed == -1:
+    seed = np.random.randint(np.iinfo(np.int32).max)
+
+  latents = get_latents_from_seed(seed, width, height)
+
+  print("txt2img from %s: %s/%s, %sx%s, %s, %s" % (user, cfg, steps, width, height, seed, prompt))
 
   pipe = OnnxStableDiffusionPipeline.from_pretrained(
     model_path,
@@ -88,7 +103,8 @@ def txt2img():
     height,
     width,
     num_inference_steps=steps,
-    guidance_scale=cfg
+    guidance_scale=cfg,
+    latents=latents
   ).images[0]
 
   output = '%s/txt2img_%s.png' % (output_path, spinalcase(prompt[0:64]))