Welcome to LanPaint Discussions!

[scraed]

👋 Welcome!

We’re using Discussions as a place to connect with other members of our community. We hope that you:

• Ask questions you’re wondering about.
• Share ideas and parameter tweaking tips.
• Engage with other community members.
• Welcome others and are open-minded. Remember that this is a community we
  build together 💪.

To get started, comment below with an introduction of yourself and tell us about what you do with this community.

[VantomPayne]

Hi! Heavy user of inpainting features and always looking for new solutions! First testing session with LanPaint provide really great results, hopefully the speed can be further improved as well as support for new samplers!

There's also the inherent issue in the inpainting process where vae encoding/decoding changing the non-masked area of the original image as well, sometimes resulting in noticeable deviation from the original image. I wonder if LanPaint can be used with Inpaint Crop&Stitch node so it doesn't affect the non-inpainted areas?

[scraed]

Hey, thank you so much for your kind words! We’ll gradually try more samplers. For faster generation, you can reduce the LanPaint num steps—though it might affect quality. Also, I think it can work with Inpaint Crop&Stitch by simply replacing the KSampler with the LanPaint K Sampler.

[NonaBuhtig]

I just wanted to thank you for your work. LanPaint is really useful and works perfectly with all the different models I use. 👍👌😁

[scraed]

Thanks for your kind words!

[joeyshelton]

I am super impressed by the results using Flux1.dev. Really awesome work and thank you for making it so easy to use! I'll be trying this next with the newly-released Qwen-Image. Should this be a pretty simple drop-in replacement starting from one of the sample workflows?

[scraed]

There are still some bug with Qwen-Image working with some comfy build in "set latent noise mask" inpainting workflow. Since LanPaint reuses nodes and codes from "set latent noise mask", it is not working with Qwen Image at current stage. we have to wait for comfy to fix these bugs.

[joeyshelton]

Do you mean the error in the Repeat Latent Noise Mask node that looks something like this:

Number of dimensions of repeat dims can not be smaller than number of dimensions of tensor

Or if bypassing the repeat latent noise mask step, an error in the LanPaint KSampler node that looks something like this:

ValueError: Input and output must have the same number of spatial dimensions, but got input with spatial dimensions of [1431, 1920] and output size of torch.Size([1, 178, 240]). Please provide input tensor in (N, C, d1, d2, ...,dK) format and output size in (o1, o2, ...,oK) format.

If so, I was able to resolve this issue by bypassing the Repeat Latent Batch node and making a tiny code change in LanPaint's nodes.py. It seems like the Qwen encoder output is a different shape than the output from the FLUX models. The flux model VAE outputs a 4D tensor like "[B, C, H, W]" whereas the qwen image VAE outputs a tensor like [B, C, T, H, W] with this extra time/frame dimension inserted. This is mentioned in their technical report here:
https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-Image/Qwen_Image.pdf#subsection.4.3

To enable the model to distinguish between multiple images, we extend MSRoPE
by introducing an additional frame dimension, in addition to the height and width used to locate image
patches within a single image (see the right part of Figure 14)

In nodes.py, I made this small change in reshape_mask() to always grab the last two dimensions (instead of all dimensions starting from index 2) and handle the 5D shape if it sees it:

def reshape_mask(input_mask, output_shape):
    dims = len(output_shape) - 2

    scale_mode = "nearest-exact"
    mask = torch.nn.functional.interpolate(input_mask, size=output_shape[-2:], mode=scale_mode)
    if len(output_shape) == 5:
        mask = mask.unsqueeze(2)
    if mask.shape[1] < output_shape[1]:
        mask = mask.repeat((1, output_shape[1]) + (1,) * dims)[:,:output_shape[1]]
    mask = repeat_to_batch_size(mask, output_shape[0])
    return mask

After I made this small change I was able to get it to work! If a user needs Repeat Latent Batch, they may need to wait for a similar fix there as you mentioned.

EDIT: I see you posted a bug for this and the Comfy devs just patched Repeat Latent Batch to fix this! I can test if the change in LanPaint reshape_mask is also needed

[scraed]

Hi, that is brilliant! If it is working please raise a pull request. I will also test the fix
EDIT: might also need to post a bug for reshape_mask, I think the set latent noise workflow also have same bug
EDIT2: I have fixed more bugs about batch size. now LanPaint should support repeated batch with the latest version of ComfyUI.

[joeyshelton]

Thank you for the quick fix and for posting the example workflow. I can confirm I see great results after the fix and the results are much better with LanPaint than with a standard Qwen-Iimage inpainting workflow!