Preface

While working on an Inpainting Workflow I figured out that the whole image out of the mask are also changed a little bit. I gradually began to suspect that the original was being changed by internal conversions between loading and saving [1].

Workflow Preview

Figure 1 shows the workflow I used for my test [2].

Figure 1: Workflow preview

What the Workflow Does

The workflow is loading an image and is saving a lossless compressed JPEG image file. Image quality is set to 100%.

Comparison of Input and Output

Test Image

The original test image used here was created using a SDXL base model and one of my LoRAs.

Input and Output Images

Figure 2 and Figure 3 shows the input and output images of the simplified workflow.

Figure 2: Image A (Input)

Figure 3: Image B (Output)

Direct Visual Comparison of Input and Output

Figure 4 is a visual comparison of image A and image B. The red parts differ between the images without indicating how much the difference is. For better visualization, a so-called shadow of the second image is added to the white unchanged and red changed parts of the image.

Figure 4: Visual comparison of image A and image B

Remarks

The commands I used for comparing images can be found in [1].

Conclusion

When using the ComfyUI Load Image you have to be aware of losing information. there is no way around it. You cannot see this, but the comparison makes it clear.

References

[1] https://www.copus.io/work/8e2416cd9f2e029ce91f85814de3fdce

[2] https://www.copus.io/work/ec283712b91b346d74155342977e51b7