Instagram: https://www.instagram.com/synth.studio.models/
Buy me a☕ https://ko-fi.com/lonecatone
This represents many hours of work that I give away for free. If you enjoy it, Please make sure to 👍like and post to my workflows. That is how you support me. Also, feel free to ⚡tip 😉
Simple Krea2 Dataset Creator
Uses Identity Edit technology for a consistent image
Creates a prompt tag set with each image
Dozens of seperate prompts to choose from
Dual image 🖼️ option for better variance between facial expresions
comprehensive notes and tips.
Blackwell attention fix node: Deals with old architecture on newer machines. (Can be deleted if not needed)
The folder is included in the zip file. The entire folder gets placed in the custom_nodes folder.
https://civarchive.com/models/2806105/krea2-identity-edit-rtx50xx-blackwell-attention-fix
Repo is here: https://github.com/lonecatone23/comfyUI-BlackwellAttentionFix
也有中文说明
I take a lot of time to get these to work. If there are issues, please let me know so that I can correct them.
Description
FAQ
Comments (17)
Works beautifully!!
if anyone else is wondering ; The blackwellattentionfix custom node files can be found in lonecat 大佬's other workflow zipfile
https://civitai.com/models/2803688/lonecats-krea2-identity-edit-w-optional-rtx50xx-attention-fix
Yeah.. I added it (V1.0.1). My dog was being needy and I forgot. 🙂
Thanks though.
@sandpies V2 is much, much better
@lonecatone23 that was quick!! making me want to train more loras now
@lonecatone23 imagine time will come when AI will replace your dog :(
I was wondering about this... so this is a crash that happens with this workflow, not a crash caused by the identity edit lora itself?? Meaning, I need that fix to use just this single wf?
@ArtiOtaku The node prevents the common xformers crash on NVIDIA Blackwell GPUs (RTX 5090 / 5080 / 5070). I ran numerous workflows with it and tried swapping out nodes. I have had several people report the issue to me, that's why I made it. This workflow contains the stock set of nodes recommended for the LoRA by its creator.
It is an Identity edit issue and specifically related to brand new high end machines. This will run on 90% of the machines out there without issue, so most people will not even notice. You can delete it if you do not need it.
@lonecatone23 is the RTX pro 6000 also affected?
@lonecatone23 understood, so in case of a crash, we can use that, but most likely, we won't... thanks, will give it a try for sure.
@ArtiOtaku Yup. Let me know. I always apprecaite input so I can make things better
@schlemihl I'm affected when you said that 😭😭
It's the same sm_120 architecture, so I'd assume so.
You can always give me your card and I'll find out 🤔
Great Stuff. The true GOAT move would be to add some expression randomization to it that is also added to the caption files. expression is the same in all images which might burn it into a lora.
Don't get me wrong the result is already great and the krea quality makes them almost better than real images.
Hmmm, Ill change a prompt list and see how it cooks. I think that's part of the identity edit cons though.
Try messing with v1.0.1 if you uploaded it earlier. You can make tweaks better now
@klapperklaus V2.0 forthcoming. I made soem pretty big changes
@lonecatone23 in Facial expressions prompt list: it is listing expressions, but does not print it in tagger files. like it does print all the rest, but completely removing words that describe expressions from txt files in tag saver :D
for example:" Sitting on steps outside a building, elbows on knees, denim and sneakers, face showing a casual easygoing smile, soft afternoon light, urban architecture behind." this is the prompt. now this is tagger txt: "Sitting on steps outside a building, elbows on knees, denim and sneakers, soft afternoon light, urban architecture behind." not sure if that is intended.
In the end all images ended up with same blank expression, so guess not intended :)
@blhll Yeah, That's technically impossible.
Lower the reference boost.
You’ve set up a workflow that works really well—great job. The only issue I see is with the naming process: The first image created has a different name than the image itself; that is, while the input image is named “character name,” the text file is created as “name_00001.” This causes a mismatch between the image and the data. Perhaps there’s a setting I’m missing that could fix this—I’m not sure. Other than that, it works perfectly.