Getting incorrect inpainting results

[jeffufu]

Hello! I just got this running. It's very cool! :)

I understand that #origin_image, #inpaint_prompt, and #base_prompt need to target the correct layer first.

Right now:

#origin_image: I set this to the main image I am inpainting on.

#inpaint_prompt: it's a text layer, right now I just write 'ancient temple, chinese' like in your video

#base_prompt: I write a short description of the image. (is it possible to get the image from Photoshop, and use Florence2 to write base prompt?)

Question 1:

Right now I make a selection in Photoshop in a blank layer, and when I hit generate, I can tell it changed my selected area but it doesn't match my prompt. It just looks like I did inpainting with very low denoise. I tried adjusting the denoise to be higher but seem to have trouble getting a unique generation. What am I doing wrong?

Question 2:

Do I have to manually select the layers (origin, inpaint prompt, base prompt) in ComfyUI to generate correctly each time?

謝謝

[zombieyang]

Q0: Yes you can change the workflow as you like, no matter Florence2 or something. For the widget panel, All I use is three kinds of node: GetDocument, GetLayerByID, Primitive

Q1: You can add some preview node in the webview to debug. maybe the mask is not getting right?

Q2: no. I think select in widget panel is ok.

have you try other workflows and are they works well? for example the simple SDXL and Flux.

[jeffufu]

Q0: Yes you can change the workflow as you like, no matter Florence2 or something. For the widget panel, All I use is three kinds of node: GetDocument, GetLayerByID, Primitive

Q1: You can add some preview node in the webview to debug. maybe the mask is not getting right?

Q2: no. I think select in widget panel is ok.

have you try other workflows and are they works well? for example the simple SDXL and Flux.

Flux - SDPPP Inpaint.json

Hi! I managed to get it working, but it wasn't with the provided workflow.

I am not a programmer but I knew something wasn't working with the sampler/denoise (for me), so I took an inpainting with mask workflow I had, and took parts from your SDPPP nodes (getting the image, base prompt etc) and combined it together. I shared the workflow. I added Florence2 so I don't have to type the base prompt.

I am a professional designer and think there could be some simple changes to make this much easier to use. Right now it's a little hard to type a layer, and then go back to Comfy to make sure I have the right layer for prompt selected. If it's like this, it's easier for me to put the image directly into Comfy, and bring it back to Photoshop.

I know you are doing this on your own time, so I am not expecting you to do what I suggest. I think it's a cool tool and wanted to share my thoughts as someone who uses Photoshop and SD for work :)

[zombieyang]

Yes, sometimes the layer selection would not take effect, I will test it more and fix in 1.5.1. And I'm considering to add a button to quickly set the layer to current selection layer.

[jeffufu]

Yes, sometimes the layer selection would not take effect, I will test it more and fix in 1.5.1. And I'm considering to add a button to quickly set the layer to current selection layer.

Oh! that's a really great idea. It would save me a lot of time if I could do that. Right now I am switching to ComfyUI (easier than opening the small plugin window) to adjust my prompt and select the layer.

If I could do that + type directly into the Photoshop plugin for my prompt, it would be so great :D Maybe it's custom Positive Prompt node so it can be displayed in the Photoshop plugin window?

Also, for me I have many workflows saved:

It would be great if I could maybe mark which ones I use often so I don't have to search for the ones specific to SDPPP.

Thanks for the great work! I'm planning on to use your nodes with https://github.com/erosDiffusion/ComfyUI-enricos-nodes/tree/master and use the image I quickly edit in Photoshop to send into this workflow.

[jeffufu]

Yes, sometimes the layer selection would not take effect, I will test it more and fix in 1.5.1. And I'm considering to add a button to quickly set the layer to current selection layer.

Also, noticed that when I use get image, it takes the entire image of the layer/group I select, even if there are parts not visible:

Left is what I started with in Photoshop, right is what it looks like when I pull the image. This makes it a little hard to use, because sometimes I move images around and don't want to lose the pixels. Not a big problem, but sharing. You can ignore it if you want!

[zombieyang]

Yes, sometimes the layer selection would not take effect, I will test it more and fix in 1.5.1. And I'm considering to add a button to quickly set the layer to current selection layer.

Oh! that's a really great idea. It would save me a lot of time if I could do that. Right now I am switching to ComfyUI (easier than opening the small plugin window) to adjust my prompt and select the layer.

If I could do that + type directly into the Photoshop plugin for my prompt, it would be so great :D Maybe it's custom Positive Prompt node so it can be displayed in the Photoshop plugin window?

Also, for me I have many workflows saved: It would be great if I could maybe mark which ones I use often so I don't have to search for the ones specific to SDPPP.

Thanks for the great work! I'm planning on to use your nodes with https://github.com/erosDiffusion/ComfyUI-enricos-nodes/tree/master and use the image I quickly edit in Photoshop to send into this workflow.

ComfyUI allows you to bookmark a workflow. I'll show them in high priority next version.

I saw your reddit post and openAI workflow, great work!

[zombieyang]

Yes, sometimes the layer selection would not take effect, I will test it more and fix in 1.5.1. And I'm considering to add a button to quickly set the layer to current selection layer.

Also, noticed that when I use get image, it takes the entire image of the layer/group I select, even if there are parts not visible:

Left is what I started with in Photoshop, right is what it looks like when I pull the image. This makes it a little hard to use, because sometimes I move images around and don't want to lose the pixels. Not a big problem, but sharing. You can ignore it if you want!

Do you mean part of the image is outside the canvas?

[jeffufu]

Yes, sometimes the layer selection would not take effect, I will test it more and fix in 1.5.1. And I'm considering to add a button to quickly set the layer to current selection layer.

Oh! that's a really great idea. It would save me a lot of time if I could do that. Right now I am switching to ComfyUI (easier than opening the small plugin window) to adjust my prompt and select the layer.
If I could do that + type directly into the Photoshop plugin for my prompt, it would be so great :D Maybe it's custom Positive Prompt node so it can be displayed in the Photoshop plugin window?
Also, for me I have many workflows saved: It would be great if I could maybe mark which ones I use often so I don't have to search for the ones specific to SDPPP.
Thanks for the great work! I'm planning on to use your nodes with https://github.com/erosDiffusion/ComfyUI-enricos-nodes/tree/master and use the image I quickly edit in Photoshop to send into this workflow.

ComfyUI allows you to bookmark a workflow. I'll show them in high priority next version.

I saw your reddit post and openAI workflow, great work!

That would be great! Yes, I do bookmark my workflows so that would work.

And yes, for the other point, when I use Generative Fill in Photoshop, or if I drop a larger image in, there is often pixels outside of the canvas. It seems your node picks up the entire image, not just the visible canvas. The problem is when I generate with that larger image and send it back to Photoshop, it gets resized. I'm not sure if this is easy to fix, though.

[zombieyang]

forgot to fix in 1.5.1.
I'll reopen it to remind myself.

[zombieyang]

The out-of-canvas issue:
I think it is because you link the boundary input with a layer boundary output?
I haven't think about it before, but maybe it is a good case to return the entire image when boundary is specified as layer boundary? If someone want to get what he/she see on canvas, the document boundary should be used instead.

Or actually you did not use the layer_bound? That should be a bug that I didn't reproduce.