it did only anime then suddenly it did photo too? LR version?
i had this on my main pc, whatever i prompted it did only like drawing anime style. (Krea2 turbo+kroma01 lora)
and then suddenly after ... idk it started doing phto as well. Even without mentioning it specifically. Very weird.
And it looks good despite Krea2 being that artsy model.
However here's the good news: It is GREAT, it gives Krea 2 full Chroma freedom (including adding the big N). Krea2+kroma is really good at anatomy and stuff, i've seen very few freak-outs. It handles complicated prompts (actually complex not one thing enhanced into oblivion), it looks good... New favorite.
The combo might be as good/better at not messing up anatomy as boogu image. And has the big N. And the image quality of krea2+kroma is deffo better. Well done. Of course OG Chroma did the same, despite being based on Flux Schnell, THE worst anatomy model (before Ernie Image). There must be one hell of a dataset (and skill) behind Chroma
.
Now im using it on my laptop, and it did the same again. All anime. Then one accidental megaprompt (failed enhance run) and suddenly photorealism. Curious.
Since the different loras aren't explained i tried rl and it does definitely NOT add the Big N. not sure what rl stands for but it seems to be the safe version?! It is also much bigger.
UpdateED update (i had an example without kroma it did not the kelvin cline bit...) :
my prompt of death seems to reach the point of how many things it can handle. Krea+kroma manages most things. But sometimes if add the cog in O it forgets the cat, if i dont, it does the cat. mostly. And the anime version seems to handle more of them. i've only seen one 6 finger hand. And arguably the anime arms are too long.
--
Expand: The prompt! it's silly. i just dumped more and more details one it to find boogus breaking point. Note it is not enhanced into a shortstory, its just details listed hard and fast.
young man age 18, average height, blue jeans, white tshirt, on the front a large abstract cogwheel symbol with "WE engineer" above and "Everything" below.
The young man stands with his arms stretched out sideways and up, the pose and his short slim fit tshirt expose the mans midraff and his low levis jeans' beltline and partly visible orange underpants band with partly visible blue print "Kelvin Cline" .
left and right of him stands another man each, each holding and lifting up one arm of the central man (realistically).
the left man (25, tall, slender, wiry, bald shaven) wears a tight elegant blue dress shirt with black/gold striped tie and a small golden "boss" print on the shirt pocket.
the right man (tall, slim, age 30, hairy body, short disheveled black hair) wears a dirty white ribbed tanktop under a dirty blue workers dungaree with a small white "NO BOSS" print on the front, a tiny baby tabby kitten balances on his biceps.
the central man looks emberassed at the left man, who returns a stern look.
In the back a large, messy but realistically equipped workshop. On the back a large dirty poster saying large in Red "DON'T" BUT the O on DON'T is replaced by a cog symbol; and below in slightly smaller, blue "Overengineer".
Expand: The prompt! it's silly. i just dumped more and more details one it to find boogus breaking point. Note it is not enhanced into a shortstory, its just details listed hard and fast.
young man age 18, average height, blue jeans, white tshirt, on the front a large abstract cogwheel symbol with "WE engineer" above and "Everything" below.
The young man stands with his arms stretched out sideways and up, the pose and his short slim fit tshirt expose the mans midraff and his low levis jeans' beltline and partly visible orange underpants band with partly visible blue print "Kelvin Cline" .
left and right of him stands another man each, each holding and lifting up one arm of the central man (realistically).
the left man (25, tall, slender, wiry, bald shaven) wears a tight elegant blue dress shirt with black/gold striped tie and a small golden "boss" print on the shirt pocket.
the right man (tall, slim, age 30, hairy body, short disheveled black hair) wears a dirty white ribbed tanktop under a dirty blue workers dungaree with a small white "NO BOSS" print on the front, a tiny baby tabby kitten balances on his biceps.
the central man looks emberassed at the left man, who returns a stern look.
In the back a large, messy but realistically equipped workshop. On the back a large dirty poster saying large in Red "DON'T" BUT the O on DON'T is replaced by a cog symbol; and below in slightly smaller, blue "Overengineer".
examples:
This is actually krea2+kroma (workflow is in the img, but it is basically the base template one). Every prompted detail is there. basically perfect.
boogu image edit base+turbo lora) gets it right basically every time. (i use the edit model but its done witout any inputimage).
Boogu image (edit)turbo noticeably less so. While Krea2 turbo seems fine (although it makes basically identical images every time while boogu base+distill has much more variety. more details and examples: https://huggingface.co/Comfy-Org/Boogu-Image/discussions/9#6a66da44ead94c6648d0648c
Boogu image 0.1 base+turbo lora:
Expand: and just for the laughs, this is flux.2 klein 9b, it got the cat position right but at what cost... (ignore the filename, it was 9b):
Expand: This was the one i did accidentally without the kroma lora, so its pure (obnoxiou sfw) krea2.
i had this on my main pc, whatever i prompted it did only like drawing anime style. (Krea2 turbo+kroma01 lora)
and then suddenly after ... idk it started doing phto as well. Even without mentioning it specifically. Very weird.
And it looks good despite Krea2 being that artsy model.However here's the good news: It is GREAT, it gives Krea 2 full Chroma freedom (including adding the big N). Krea2+kroma is really good at anatomy and stuff, i've seen very few freak-outs. It handles complicated prompts, it looks good... New favorite.
The combo might be as good/better at not messing up anatomy as boogu image. And has the big N. And the image quality of krea2+kroma is deffo better. Well done. Of course OG Chroma did the same, despite being based on Flux Schnell, THE worst anatomy model (before Ernie Image). There must be one hell of a dataset (and skill) behind Chroma
.
Now im using it on my laptop, and it did the same again. All anime. Then one accidental megaprompt (failed enhance run) and suddenly photorealism. Curious.Since the different loras aren't explained i tried rl and it does definitely NOT add the Big N. not sure what rl stands for but it seems to be the safe version?! It is also much bigger.
Make sure to use natural prompting. Don't forget to start the prompt with things like "A real life photo of" or just "A photo of". For anime or cartoon start with "An anime/cartoon artwork of". A prompt full of tags can confuse the training data of the model. The RL model works nice between the 0.6-0.8 spot, but it takes a life to load (at least for me). RL means "reinforcement learning", I think lodestones create it after undertrained reports.
i tried many prompts and it didn't. And then suddenly it did and now it does it with every prompt. The prompt in my added examples (one is a anime one) is just plain facts listed. Not quite tags but certainly no flowery 19'th century novel.
And you see how Krea2 turbo+kroma01 does it well (in terms of prompt details followin, but also image quality is imho ok, i had to skimp on steps a bit bcecause laptop.
RL is full on sfw. More so than base flux2. But its early days. I'm actually blown away by the basic kroma one (more intesrested in the nsfw capabilities that the style thing). It can do what Chroma1hd can but the base is so much better.
i tried many prompts and it didn't. And then suddenly it did and now it does it with every prompt.
RL is full on sfw. More so than base flux2. But its early days. I'm actually blown away by the basic kroma one (more intesrested in the nsfw capabilities that the style thing)..
I understand, the style adherence is definitely something that kroma needs to improve. I also had issues to generate a specific style that the base model already now, such as realistic/photorealistic 3D and another aesthetic styles. The nsfw of kroma is awesome but needs some additional details for the skin texture and another places like... Boobs 😳, the skin texture can be improved but Qwen Image vae definitely sucks for this.
i'm in the lucky (?) position that boobs are not too important. And chroma is one of the very few nsfw capable models that does THAT as well (and very well) witout crowbaring in female details everywhere when not prompted.
i'm in the lucky (?) position that boobs are not too important. And chroma is one of the very few nsfw capable models that does THAT as well (and very well) witout crowbaring in female details everywhere when not prompted.
Nah, I mean the adherence in general (including males and mature mommies). An Interesting fact that I've noticed testing is that the LoRa strength impact the uncensor rate. For example, I type "a big fat woman/man with huge boobs" (lmao), the model (Krea 2) can simply ignore the boobs size and make it smaller than I've prompted, just because the LoRa strength is 0.6.
but does it improve the texture, since i assume it has never been really trained on them? I also has some good and more diverse results with krea2 raw+ distill lora, even at lora at1.0, low steps and 1 cfg. But the difference in Quality is not quite as big as with boogu base+turbo vs og turbo model..
I guess a community finetune, especially a big one for all one like Chroma, will always a be bit of a compromise but here will not be a major t2i model that does nsfw despite the hopefull asking comments on every release. The public backlash would just be too great. So we'll see how good Kroma gets.
but does it improve the texture, since i assume it has never been really trained on them? I also has some good and more diverse results with krea2 raw+ distill lora, even at lora at1.0, low steps and 1 cfg. But the difference in Quality is not quite as big as with boogu base+turbo vs og turbo model..
I guess a community finetune, especially a big one for all one like Chroma, will always a be bit of a compromise but here will not be a major t2i model that does nsfw despite the hopefull asking comments on every release. The public backlash would just be too great. So we'll see how good Kroma gets.
Yeah, I mean, realism is not really important for me. I'm just typing this because I've seen some feedbacks complaining about this ""issue"", but come on, even the latest GPT Image 3 isn't really realistic as real life (it looks like deep fried and blurry lmao). Anyway, I'm happy for we have Kroma.