StableDiffusion3Img2ImgPipeline.__call__() is missing width and height parameters

[chie2727]

Describe the bug

The docstring for the StableDiffusion3Img2ImgPipeline.__call__() function includes width and height parameters, but the function itself does not include these parameters.
Is this a typo or is width and height supposed to be handled by the function?

Source file:
diffusers/src/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_img2img.py

Reproduction

import torch
from diffusers import StableDiffusion3Img2ImgPipeline

pipe = StableDiffusion3Img2ImgPipeline.from_pretrained(
    "stabilityai/stable-diffusion-3.5-medium",
    torch_dtype=torch_dtype,
    cache_dir=torch.float16,
    token=hf_token,
)

image = pipe(
    prompt="Resize the input image",
    image=input_image
    width=1024,
    height=512,
    strength=1.0
).images[0]

Logs

TypeError: StableDiffusion3Img2ImgPipeline.__call__() got an unexpected keyword argument 'width'

System Info

• 🤗 Diffusers version: 0.31.0
• Platform: Linux-6.1.79-99.167.amzn2023.x86_64-x86_64-with-glibc2.34
• Running on Google Colab?: No
• Python version: 3.11.6
• PyTorch version (GPU?): 2.5.1+cu118 (True)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Huggingface_hub version: 0.26.2
• Transformers version: 4.46.2
• Accelerate version: 1.1.1
• PEFT version: not installed
• Bitsandbytes version: not installed
• Safetensors version: 0.4.5
• xFormers version: not installed
• Accelerator: NVIDIA A10G, 23028 MiB
• Using GPU in script?: yes
• Using distributed or parallel set-up in script?: distributed

Who can help?

@yiyixuxu @sayakpaul

[ghunkins]

As far as I understand it, height and width are inferred from the input image. Docstring addition appears to be a copy-paste error.

Adding any required resizing prior to sending the image to the pipeline should yield what you're looking for!

[chie2727]

@ghunkins
So it's copy-paste error in the docstring then - thank you for clarifying!

I was hoping to be able to specify an output image size that differs from the input image size, but I'll do a bit more research into how to achieve this.

[sayakpaul]

@ghunkins thanks for helping out. @chie2727 feel free to close the issue if you think if it's resolved.

[ukaprch]

Interestingly, FLUX has no such limitation in their FluxImg2ImgPipeline.

[asomoza]

Hi, what would be the use case of using a different width and height. That only will result in a distorted image if they don't match the source image, why people would want that?

If the others had it (img2img) and there's a genuine use case maybe we can add it.

[ukaprch]

To your point, resizing the image for the best sizes that FLUX supports would be the main reason taking into account aspect ratios. Based on a post I saw on Reddit which made sense, these are the best sizes to use for FLUX (divisible by 64 H X W): aspect_ratio == '1:1 1024 x 1024': aspect_ratio == '1:1 1408 x 1408': aspect_ratio == '3:2 1728 x 1152': aspect_ratio == '4:3 1664 x 1216': aspect_ratio == '16:9 1920 x 1088': aspect_ratio == '21:9 2176 x 960': aspect_ratio == '2:3 1152 x 1728': aspect_ratio == '3:4 1216 x 1664': aspect_ratio == '9:16 1088 x 1920':

…

On Fri, Nov 15, 2024 at 12:12 PM Álvaro Somoza ***@***.***> wrote: Hi, what would be the use case of using a different width and height. That only will result in a distorted image if they don't match the source image, why people would want that? If the others had it (img2img) and there's a genuine use case maybe we can add it. — Reply to this email directly, view it on GitHub <#9933 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AZTE5IBQUS3SZ3FZW6GMX232AYTRDAVCNFSM6AAAAABR2HUWS6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDINZZGQ3DKNBVGA> . You are receiving this because you commented.Message ID: ***@***.***>

[ghunkins]

@chie2727 Here is some documentation from PIL as to various image resizing techniques given a specific desired size. Best of luck!

https://pillow.readthedocs.io/en/stable/reference/ImageOps.html#resize-relative-to-a-given-size

from PIL import ImageOps

required_size = (1024, 512)
resized_input_image = ImageOps.fit(input_image, required_size)

image = pipe(
    prompt="Resize the input image",
    image=resized_input_image,
    strength=0.5,
).images[0]

[asomoza]

To your point, resizing the image for the best sizes that FLUX supports would be the main reason taking into account aspect ratios. Based on a post I saw on Reddit which made sense, these are the best sizes to use for FLUX ...

@ukaprch but this is for SD3 and not Flux, also I agree that there are some resolutions that works best for the models but that only make sense using the txt2img pipelines, with img2img as @ghunkins pointed out, people should resize the source image before feeding it to the pipeline, otherwise the generated image will be distorted.

[ukaprch]

I agree with you, the sizes I showed are for Flux.