[Redefine] Seed Differences in Results

Hi all,

I’m having a strange issue with Gigapixel. In the preview window, everything looks great — especially areas with people. The faces and details look refined and acceptable. But when I export the final image, it looks very different. The people appear oddly deformed — not just softer, but with strange artifacts or changes that don’t resemble what I saw in the preview.

It feels like the final render is using a completely different model or settings than the preview. I’ve tried various models (Standard, High Fidelity, etc.), toggled face recovery, experimented with scaling, and tested different export formats, but nothing seems to help.

As a very hacky workaround, I’ve resorted to taking a screenshot of the preview area, then pasting that into Affinity Photo to composite over the original image. It’s obviously far from ideal, but it shows how much better the preview looks compared to the export.

Has anyone else seen this or figured out a way to fix it?

I’m on Win11, RTX5080 and GP 8.3.3.

Thanks in advance!

Deformed heads and faces …

Yes I’ve seen it in the last few days with old monochrome postcards and could not find any options in Gigapixel AI 8.3.3 or Photo AI 3.6.0 to get close to the OK Previews.
My system is Windows 10 PC RTX 4070 Super
I will look for some examples later this evening

At least I’m not the only one on this boat :sweat_smile:

Here’s one example — in theory, we should expect a 1:1 match between the preview and the final output, but as you can see, that’s definitely not happening here.

Preview:
image

Output image:
image

1 Like

:rofl:

It looks like they’ve finally resolved making everyone look like Topaz Girl.

I’ve never had that issue with the non-generative models, only with the newer generative models (Recover and Redefine). With non-generative models, preview matches output. If your having problems with the non-generative models, then you’ll need raise a ticket with support, so they can work through it with you.

1 Like

Thank you for your insights!
Being new in this forum, do you know if the moderators open the ticket or I have to do something for it?

The link for the contact form is https://docs.topazlabs.com/#contact-us

Or the email address is support@topazlabs.com

1 Like

Thank you, I sent them an email. If possible I will keep things updated here if something comes.

Hi mate, I’ve the same problem in any image. When I do small previews everything is ok, when I calculate full preview I got a lot of deformations. Did you find a solution?

Hi,

I reached out to support and got this response:

The issue is that the preview uses one seed, and another seed is generated when the export is run. This can result in an unexpected output.

I will submit your ticket to the engineers and email you back when we have an update that addresses this.

So it’s definitely a known issue now — hopefully we’ll see a resolution soon.

In the meantime, my workaround is to generate a bigger preview, take a screenshot, and do some compositing. :melting_face:

Thank you. They have to understand that it is not a problem of different seed, it is a problem of quality of the ai drawing when doing full preview. There’s a clear problem, their response seem an excuse to me.

1 Like

What is the size of your original image? I just posted a fix that may apply to your issue as well. In short, start with a higher resolution image. I had the problem when my pre render image was 1024x512 but the issue disappeared when my start image was 2048x1024, Good luck.

Not sure if this solves your problem but…

I have been repeatedly frustrated that the preview and final render bear no resemblance to each other. Preview looks great but the render is over processed.

The issue is consistent but is fixable by changing the original pre-render image to a larger size.

1 Like

My starting image was 3400x1700. I followed your advice and I changed the resolution to 13.600 x 6.800 using photoshop. Then I loaded the image again on Gigapixel (Redefine 1x Creativity 3) and everything worked much better! Thank you for your workaround, it can help until the team find a real fix.