✦ Arthemy Comics Krea2 ✦
I've baked parts of this filter-bypass in oder to improve the facial expressions: https://civarchive.com/models/2746817/krea2-filter-bypass-fedor
If you ever worked on a model, you know how it is hard to play with fp8. The process this time has been even messier than usual (so much so that my custom suite for tuning Krea models almost fell apart in any possible way it could), but the result is really fun to work with, if you want a western comics aesthetic or western cartoons.
As always, I've "trained" this model by manually changing its internal value while checking some live outputs at every step - tuned and equalized its tensors and then built highly specific positive and negative LoRAs trained to solve some of its issues.
This is really fun to play with but, I want to work a lot more on the TextEncoder side of things - it feels like it's not providing the emotions I'd like to showcase in my images, it's a little stiff, but I think I can solve that... somehow, somewhat, somewhen?
Let me know if you encounter some bugs or inconsistencies (I'm happy to fix and analyze the errors, but I can't fix what I don't know that it's broken) - and have fun! ^_^
___________________________________________________________________________________
1_____Size: Around 1024 x 1344
2_____Prompt:
Quick Prompt
Western comics style, bold ink outlines, hatched shadows, [expression], seen from [camera angle], upper body portrait, dutch angle, [pose].
[Character description][clothing][accessories].
Background: [setting]. Lighting: [light source, color and shadow].3_____CFG: 1.0
4_____Steps: 9
I hope you like my models and I can't wait to see what you all are going to create with it!
Description
Fixed facial expressions
Now in BF16 format
This is a total model fine-tune so, it's not possible to convert it into a LoRA.
FAQ
Comments (19)
no fp8 for version 1.1?
I can share that too if you want, for me it was just faster to load than the fp8 in this format.
I have RTX 3060 with 12gb vram, so 23gb is too much for me
@mariocax have you tried? You might be able to use it anyway with comfyUI
@Arthemy would be nice to also have the fp8 version, since everything over leaks in ram, and it slows down alot.
@Emenimuno @mariocax fp8 and int8 added! ^_^
@Arthemy you're awesome! TY
you are the best
@Arthemy apprently, you're right. I can use BF16 with no problems, just tested it out on a 16gb vram. ComfyUI is really nice.
Very nice update. Seems to follow prompting better. Keep up the good work!
There were a few things that I just had to fix. The previous facial expressions were just flat.
1.1 indeed perform much better but size increase trippled the loading speed in my 5080
Would you consider releasing the raw (base) version as well? It does help creativity when used with Turbo lora at 0.6 strength.
I can do that, but it might be a little overkill to use the RAW version, have you tried different versions? ^_^
Don't take this as a criticism, I'm just trying to understand everything better. This model has no trace of the comic book aesthetic if you don't specifically prompt it heavily in that direction (realistic western comics style, bold ink outlines, etc). So it makes me wonder 2 things: How much of this look is coming from the checkpoint vs coming from the prompt and also if this is the case, why not make it a lora rather than a whole checkpoint? Surely if the checkpoint itself doesn't have this look heavily baked into it, you should be able to recreate it pretty well with a lora?
Hey there!
First of all: I will take it as criticism, but there is nothing wrong with that, in fact: I love being criticized because it's giving me an important feedback on how to improve or how to explain what I'm doing in a better way.
Before responding to your questions, I have to explain how I work:
Instead of training a LoRA that provides a specific aesthetic, I start from the base model and I change its internal values manually in order to slowly make the "comicbook aeshtetic" the kind of aesthetic I was searching for. This means that this model might not be able to produce realistic photos, because I've tuned it to amplify the concepts tied to a 2D comicbook aesthetic.
To do that, I wrote a few test prompts and, for example, if "purple skin" was generating a character with a normal skintone on some parts of their body, I tried to recalibrate the values until the output showcased the correct skin-tone on each part of that character.
Imagine doing that for many concepts, tuning it based on the kind of outputs that you're trying to achieve on the whole structure of the model.
The output will be a "full-model" that has been recalibrated to work best in a specific area, but it's still able to do everything else, because overriding what "realistic" mean, might affect its ability to make your "Comicbook outputs" more realistic, if you want to add that keywords to the prompt.
"If the Checkpoint itself doesn't have this look heavily baked into it, you should be able to recreate it with a LoRA"
Unfortunately not, because by tweaking the whole model, if you try to extract the differences, you'd get a 20GB~ LoRA (the model is almost 90% different).
"How much of this look is coming from the checkpoint vs coming from the prompt and also if this is the case, why not make it a lora rather than a whole checkpoint?"
Just because a model is able to create everything, it doesn't mean it's not better in a specific area. I could make a LoRA, but I find it much more interesting to work on the whole structure of a model (since these modern models have seen enough material to be already able to produce any specific aesthetic).
Also, LoRA and Fine-tuning rarely match exactly what you're trying to achieve. By manually shaping the whole model I can see how and where it's changing, making sure that the model is improving in one area without affecting its ability to be flexible.
Said so:
Even though my reasoning might be correct - as a model creator - I understand what you mean as a user and you're 100% correct, that these full-models are overkill. I'm already working on a "Model / CLIP" tuner that will transform these models, not into a LoRA but into "Modifiers" for the original model. This means that this model can become a simple JSON of a few KB worth of patches that will change the values of individual blocks and sub-blocks of any model. ^_^
I'll let you know when it will be ready!