Version 1.1 not working?
Did anything change for the workflow for 1.1 vs 1.0?
Used the standard workflow and 1.0 works perfect. When I change out the LoRA to 1.1, it passes through the reference picture without changing anything (like denoise was only set to .01 or something). Does anything need to be changed in the workflow for 1.1 use?
Edit: I went to take screenshots and now it works for some reason. Server must have needed a full restart or something. I already tried restarting earlier, but idk...
Edit2: It just started doing it again and nothing was changed with the workflow. Going to sound odd, but it does not seem to work on anything portrait, but it works in landscape with 1.1. The original 1.0 lora works with everything.
Could you share a screenshot? Easier than asking for every single parameter. Just censor anything you don't want shown.
I don't see anything wrong with your screenshot, so that's really weird...
The only thing I would change is resolution steps to 16 down from 64.
Maybe an issue because you're using two input latents? Could you disconnect the B latent and see if it works with portrait aspect ratios then?
Looks like that makes a bit of improvement and the image actually changes a little bit now, but 1.1 is still very picky. Seems like I might need to use one or the other of the LoRAs depending on what I am trying to do.
Very very odd. Just to test, here's a confirmed working workflow I just made for testing. Could you give it a try?
The image of her in the red shirt should include a workflow, but if it doesn't then you can download it from this pastebin i uploaded.
https://pastebin.com/X3VNnjZ2
It is working. I guess for some images 1.1 is still a lot more subtle than 1.0. I have only tried Turbo models though.
"Make her dress blue"
1.0
1.1
Could you send me the source image to test with? Maybe I find something useful :P
This image has proven to be one of the more resistant ones to 1.1
Thanks I'll do some testing :]
Huh, you're right. It really does struggle with that image for some reason.
I tested it with flux.2 dev as well to see if it was consistent across models, but it changed just fine there.
I was able to make it change with each seed when I changed the prompt to "change the color of her dress to blue" or "change the color of her dress to a vibrant blue" rather than "make her dress blue".
Worked on both RAW and Turbo models.
It keeps coloring her feet, but I think it's because the artstyle is inconsistent and blurs between a drawing and 3d model/figurine.
Here are two results that haven't been cherry picked.
Huge thanks to @CeciliaXCIX for the debugging here. And that finding is really valuable: "make her dress blue" failing where "change the color of her dress to a vibrant blue" works is a real instruction-phrasing sensitivity, not your fault. I have noted it specifically as a target for the next version (training on more varied phrasings of the same edit). For now, phrasing edits as "change X to Y" is the more reliable pattern. Thanks both of you.
Thank you all!