My First Krea2 Turbo model!
V2 usage:
20 steps, CFG 1, er_sde/simple should give great results!
If you like the model, please post some of your creations !
FIXED THE CONVROT INT8 VERSION!
A note regarding user @zokamuk :
Unfortunately, this user chose to leave a negative creator rating without ever contacting me about the issue they claimed to have with my latest model.
Their direct messages are disabled, so there was no way for me to troubleshoot, explain, or resolve the problem beforehand.
I stand behind my work and I'm always willing to help users who encounter genuine issues.
Constructive feedback is welcome; silent negative ratings without any attempt at communication are not.
To anyone experiencing problems with one of my models: please reach out first. In almost every case, there's a solution.
Description
V1 was a little too grainy and had some issues following prompts.
Should be fixed in this version!
FAQ
Comments (15)
v2 is really a whole different model at this point, loses a lot of older styles
It is, but it is still capable of creating older styles.
which loader do you use to load the int8 version plz ?
I dont use INT8 versions, so I have no idea :D:D sorry :)
Honestly, I was a little skeptical at first, but the prompt adherence really has improved dramatically. I have no idea how you managed to make this possible, but it’s seriously impressive. This is a great checkpoint! 😃
I converted the BF16 model directly to INT8 Conv, and it runs fast on a 16 GB VRAM GPU. I don’t know why that works so well either, but it does.
Thanks @larflarf
Why didnt you take the int8 version I made? :)
@CornmeisterNL I didn't know that existed.. Since you mentioned in your comment that you don't use int8, I thought only bf16 existed.😭
What's the difference between INT8 and fp8?
The short (non-technical) version is that FP8 doesn't work great on 30-series GPUs. You don't get the speed benefit from using a lower precision.
If you've got a 40/50 series, FP8 and INT8 will run a similar speeds (with FP8 usually having slightly higher quality) but on a 30-series, INT8 is LOTS quicker.
@pedro1148393387 Hey, thanks man!
I'm getting an error when I load this model on my krea 2 work flow, (ValueError: Unknown quantization format for layer blocks.0.attn.wq)
I don't really understand why.. I'm using the v2 int8 model, i assumed it would run just like other krea 2 models I've been testing..?
Hehe it should :)
@CornmeisterNL The INT8 weights were fine, but all 224 quantization records were missing "format": "int8_tensorwise". I added that field to each record in a copy of the model without changing the weights or tensor positions. ComfyUI could then recognize and load the INT8 format. Whack.
@HenloFLY Im on a holiday, but when back home, Ill take a look at it!
@CornmeisterNL I did fix it, but I think it may be my fault, I can elaborate if anyone needs if not, I can now run it without editing any of the model.