Some comments and issues

[agragregra]

1. In some photographs has degraded colors, tones, etc. when expand images.
2. GGUF models are noticeably worse in images result quality than Safetensors.
3. A request to add an option to select Klein 9B during server installation.

[Acly]

1. You need to be more specific, show images and comparisons. Also check the "Color Match" option in Diffusion Settings
2. This is normal if you use a smaller quant (use less memory, but quality is worse)
3. Klein 9B is not publically available (requires sign in and agreement) - not planning to circumvent that for now

[agragregra]

1. You need to be more specific, show images and comparisons. Also check the "Color Match" option in Diffusion Settings

1. Sorry i chech other version, same issue in ALL versions. "1. In some photographs has degraded colors, tones, etc. when expand images."

[fointypinger]

That's the colour band issue I identified here.

Are you using Flux 2 Klein 4B? If so, try using 9B instead. Instead of "Expand", use "Add Content" with the prompt "Expand the image to fill the canvas". I've found this works perfectly.

[agragregra]

2. This is normal if you use a smaller quant (use less memory, but quality is worse)

Thank you. Yes, i have to download Safetensors manually for real work.

[agragregra]

3. Klein 9B is not publically available (requires sign in and agreement) - not planning to circumvent that for now
   If so, try using 9B instead

but we dont have "flux-2-klein-9B-outpaint-lora.safetensors", I looked for it, it doesn't exist, 9b manualy downloaded can generate images only

[fointypinger]

You don't need an outpaint Lora for F2K 9B. Just expand the canvas by the desired amount, make a selection slightly overlapping the edge of your image, select "Add Content" and use the the prompt "Expand the image to fill the canvas".

As an aside, F2K is such a capable model, I'd recommend only using Loras as a last resort. Most of the time, the desired result can be arrived at solely through better prompting (and in less time and with higher quality).

[agragregra]

You don't need an outpaint Lora for F2K 9B. Just expand the canvas by the desired amount, make a selection slightly overlapping the edge of your image, select "Add Content" and use the the prompt "Expand the image to fill the canvas".

As an aside, F2K is such a capable model, I'd recommend only using Loras as a last resort. Most of the time, the desired result can be arrived at solely through better prompting (and in less time and with higher quality).

Thank you so much

[Acly]

You can try without outpaint lora - the lora isn't strictly needed, it just helps make it more reliable.

Flux 2 Klein is bad at maintaining color, it's not a problem that can be completely solved I think. The default setup with 4B/Lora/Color-match works for many images out of the box in my tests, but some require to experiment a bit (9B can help, slightly more steps can help, maybe do one side at a time.

It's not entirely clear from looking at the output, but if you think there is a issue with color match (like it makes things worse or no change), please post the original .kra file as it is before the expand. That way I can try it.

[fointypinger]

I'm pretty certain it's just a quirk of 4B. I've found 9B + Add Content + Prompt works perfectly every time; might it be possible to rework the "Expand" option for 9B so it worked in this manner automatically? Expand doesn't really function as intended with 9B as it currently is.

[rktvr]

yeah, 9b craps all over 4b in terms of quality and capability. doing dozens of comparisons etc, it's better to get a gguf of 9b than the full 4b model.

[agragregra]

yeah, 9b craps all over 4b in terms of quality and capability. doing dozens of comparisons etc, it's better to get a gguf of 9b than the full 4b model.

I wonder if there's a technical way to log in to huggingface from the plugin interface, for example, by pressing a button, if shadow unlogin downloading is prohibited.