Prompt different scenes doesn't work. (Always get the white background as output)

[baddth]

Hi there,

We used inference_IMAGdressing.py with scene-related prompts, but the output was always a plain white background. We also couldn't change the gender of the result, unlike the example in the README.

Can you help us understand what we might be doing incorrectly?

We use the weight from IMAGDressing-v1_512.pt.

inference_IMAGdressing.py

    prompt = 'a man 'in the park'
    prompt = prompt + ', best quality, high quality'
    null_prompt = ''
    negative_prompt = 'bare, naked, nude, undressed, monochrome, lowres, bad anatomy, worst quality, low quality'

    print("BADDTH: ", prompt)

    clothes_img = Image.open(args.cloth_path).convert("RGB")
    clothes_img = resize_img(clothes_img)
    vae_clothes = img_transform(clothes_img).unsqueeze(0)
    ref_clip_image = clip_image_processor(images=clothes_img, return_tensors="pt").pixel_values

    output = pipe(
        ref_image=vae_clothes,
        prompt=prompt,
        ref_clip_image=ref_clip_image,
        null_prompt=null_prompt,
        negative_prompt=negative_prompt,
        width=512,
        height=640,
        num_images_per_prompt=num_samples,
        guidance_scale=7.5,
        image_scale=1.0,
        generator=generator,
        num_inference_steps=50,
    ).images

The output: (Cannot change scene and gender)

The example on README that we tried to get:

Please keep doing the great work on the IMAGDressing! Your team rocks!

[muzishen]

The scale needs to be adjusted, similar to the webui.

[MrWhiteSmall]

It's true, I set the prompt guidance scale to max, the prompt about 'background' can work.

[TanHiep-To]

Hi, @muzishen and @MrWhiteSmall. Could you provide an example of how to configure these parameters? I tried the method mentioned above, but it didn't work.

output = pipe(
        ref_image=vae_clothes,
        prompt=prompt,
        ref_clip_image=ref_clip_image,
        null_prompt=null_prompt,
        negative_prompt=negative_prompt,
        width=512,
        height=640,
        num_images_per_prompt=num_samples,
        guidance_scale=7.5,
        image_scale=1.0,
        generator=generator,
        num_inference_steps=50,
    ).images