Train_controlnet_sdxl.py and tensorboard log images

Describe the bug

When logging validation images on tensorboard, one tile is always created, even if there are several validation projects and images. I suggest making changes to the code:

for tracker in accelerator.trackers:
        if tracker.name == "tensorboard":
            i=0
            for log in image_logs:
                images = log["images"]
                validation_prompt = log["validation_prompt"]
                validation_image = log["validation_image"]

                formatted_images = [np.asarray(validation_image)]

                for image in images:
                    formatted_images.append(np.asarray(image))

                formatted_images = np.stack(formatted_images)

                tag = f"{validation_prompt}_{i}"
                tracker.writer.add_images(tag, formatted_images, step, dataformats="NHWC")
                i=i+1
            tracker.writer.flush()
        elif tracker.name == "wandb":

Reproduction

Start Train_controlnet_sdxl.py with params:

accelerate
launch
train_controlnet_sdxl_orig.py
--pretrained_model_name_or_path="model/stable-diffusion-xl-base-1.0"
--pretrained_vae_model_name_or_path="model/sdxl-vae-fp16-fix"
--output_dir="output"
--train_data_dir="fillsmall2"
--cache_dir="cachedir"
--mixed_precision="fp16"
--resolution=512
--learning_rate=1e-5
--max_train_steps=10000
--train_batch_size=1
--gradient_accumulation_steps=1
--max_grad_norm=1.0
--seed=42
--checkpointing_steps=1000
--enable_xformers_memory_efficient_attention
--use_8bit_adam
--report_to="tensorboard"
--lr_scheduler="constant_with_warmup"
--lr_warmup_steps=100
--logging_dir="logs"
--validation_image
"validation_images/1.png"
"validation_images/2.png"
"validation_images/3.png"
--validation_prompt
"red circle with blue background"
"green circle with brown background"
"blue circle with white background"
--validation_steps=10
--num_validation_images=2

Open http://localhost:6006/ (Tensorboard)

Logs

System Info

Diffusers version: 0.34.0.dev0
Torch version: 2.7.0+cu126
XFormers version: 0.0.30
Python version: 3.10.11
Windows 10

Who can help?

No response

Beginne in Train_controlnet_sdxl.py oder am im Reproduktionsbeispiel verwendeten Einstiegspunkt train_controlnet_sdxl_orig.py, und zwar beim im Issue gezeigten tensorboard-Tracker-Zweig. Führe den bereitgestellten accelerate launch-Befehl aus und untersuche die TensorBoard-Bildausgabe. Als erledigt gilt die Aufgabe, wenn separate Validierungs-Prompt- und Bildgruppen unterschiedliche protokollierte Kacheln erzeugen statt einer einzigen Kachel.