Fix bug in Z image Control Training (#403)

scripts/z_image_fun/train_control.py

@@ -86,7 +86,6 @@ from videox_fun.models import (AutoencoderKL, AutoProcessor, AutoTokenizer,
 from videox_fun.pipeline import Flux2Pipeline
 from videox_fun.utils.discrete_sampler import DiscreteSampling
 from videox_fun.utils.utils import get_image_to_video_latent, save_videos_grid
-from videox_fun.utils.utils_yolo import ObjectInstanceDetector
 
 if is_wandb_available():
     import wandb
@@ -842,7 +841,6 @@ def main():
         low_cpu_mem_usage=True,
         transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']),
     ).to(weight_dtype)
-    yolo_instance = ObjectInstanceDetector(device=accelerator.device)
 
     # Freeze vae and text_encoder and set transformer3d to trainable
     vae.requires_grad_(False)

scripts/z_image_fun/train_control_2.0.sh

@@ -30,5 +30,6 @@ accelerate launch --mixed_precision="bf16" scripts/z_image_fun/train_control.py
   --max_grad_norm=0.05 \
   --enable_bucket \
   --uniform_sampling \
+  --add_inpaint_info \
   --transformer_path="models/Personalized_Model/Z-Image-Turbo-Fun-Controlnet-Union-2.0.safetensors" \
   --trainable_modules "control"