22

app.py

@@ -6,7 +6,7 @@ from videox_fun.pipeline import ZImageControlPipeline
 from videox_fun.models import ZImageControlTransformer2DModel
 from transformers import AutoTokenizer, Qwen3ForCausalLM
 from diffusers import AutoencoderKL
-from utils.image_utils import get_image_latent, scale_image
+from utils.image_utils import get_image_latent, rescale_image
 from utils.prompt_utils import polish_prompt
 # from controlnet_aux import HEDdetector, MLSDdetector, OpenposeDetector, CannyDetector, MidasDetector
 from controlnet_aux.processor import Processor
@@ -86,7 +86,7 @@ def inference(
     prompt,
     input_image,
     image_scale=1.0,
-    control_mode='Canny'
+    control_mode='Canny',
     control_context_scale = 0.75,
     seed=42,
     randomize_seed=True,
@@ -114,7 +114,9 @@ def inference(
     else:
         processor = Processor('canny')
 
-    control_image, width, height = scale_image(input_image, image_scale, 8)
+
+
+    control_image, width, height = rescale_image(input_image, image_scale, 8)
     control_image = control_image.resize((512, 512))
 
     print("DEBUG: processor running")

utils/image_utils.py

@@ -2,12 +2,18 @@ import torch
 from PIL import Image
 import numpy as np
 
-def scale_image(img, scale, nearest=32):
+def rescale_image(img, scale, nearest=32, max_size=1280):
     w, h = img.size
     new_w = int(w * scale)
     new_h = int(h * scale)
 
-    # Adjust to nearest multiple of 32
+    if new_w > max_size or new_h > max_size:
+        # Calculate new size keeping aspect ratio
+        scale = min(max_size / new_w, max_size / new_h)
+        new_w = int(new_w * scale)
+        new_w = int(new_w * scale)
+
+    # Adjust to nearest multiple
     new_w = (new_w // nearest) * nearest
     new_h = (new_h // nearest) * nearest