Shameless plug for my deviant art page https://www.deviantart.com/le-razzz
Follow me there for great Ai video content.
Realism Engine Krea 2 v2
This version is pretty in line with Ideogram 4 v5 knowledge.
Realism Engine Krea 2 v1
Due to high demand I decided to do a Krea 2 training. There's probably improvements to be made as I learn the quirks of the model.
The gallery represent pretty much what the V1 can do.
V5 it out! It fix a lot of the issues V4 had.
Not many concept added but the previous ones are reinforced.
Better styles, more realism even with euler which make everything faster.
The default settings for the Lora's are 0.8 and 0.3-0.5 for the unconditional.
If you like a cleaner image you can increase the unconditional value but it might start to look more plastic.
Added concepts:
sucking breasts
deepthroat
soaking wet pussy
V4. New concepts
no bra pokies
anal fisting
vaginal fisting
big hanging boobs
bondage
spread pussy
futanari
anal prolapse
I recommend .7 for the Lora and .3 for unconditional. Euler look more realistic now so you can get faster realistic generation.
V3 add a bunch of sexual position concepts
Here's the full list of all the concepts so far:nude woman
nude man
pov vaginal sex
pov doggy style vaginal sex
reverse cowgirl vaginal sex
sideview doggy style vaginal sex
doggy style anal sex
pov anal sex
reverse cowgirl anal sex
pov blowjob
blowjob
sloppy blowjob
handjob
titjob
vaginal dildo
lesbian pussy licking
huge natural breasts
face cum
hairy pussy
pussy insertion (name your object)
puffy nipples
big fake boobs
upskirt
ass spread
creampie
gaping pussy
gaping anus
self fisting
Note: There might be a bit of pussy and anus confusion sometime. In that case put a bounding box where the other hole should be.
V2 add men and a bunch of new concepts.
Don't go higher than .90 or it will start creating bad quality image.
Keep the steps at 18 instead of 12 as you may get generations error from time to time.
Sweet spot is between 0.50-0.90.
res_2s = best definition
res_2m = most realistic and faster than 2s
euler = fastest but less realistic
https://huggingface.co/RazzzHF/Realism_Engine_Ideogram_4/resolve/main/rei4_v2_workflow.json
V1 - First RE Ideogram 4 release. Trained on female DS only so far. Male is coming in the next versions.
Sweet spot is between 0.50-0.60. Higher will degrade the quality.
MAKE SURE YOU USE JSON PROMPTS.
res_2s res_2m is recommended.
Description
Krea 2 Version
FAQ
Comments (57)
breast sucking never work, without this lora krea sometime do, but with it, never. training data excluded?
Can you do one for boogu image edit please
yes boogu has potential even at 0.1 but it needs a v good lora
@2legsRises357 thats what im looking for. Im getting less anatomy error with it than flux
GOAT
Ideogram is better imo
real
Diversity is a good thing. We have options now and you lames still complain.
@Superelmostar675 diversity in this context means splitting the community and lowering gpu time spent on lora training for the superior model
Ideogram is better quality but Krea2 is perhaps a better experience (much faster, natural language, excellent NSFW understanding with just a single lora (this one https://civitai.red/models/2727829/krea-2-enable-nsfw-prompt-adherence-krea2-nsfw?modelVersionId=3066310)
So I love both models. One for sheer quality the other for ease of use.
@Prefo convenience will always win for the average lazy 1girl proompter, doesn't change the fact that ideogram is currently SOTA for local and krea is a slightly better klein 9b
@rtrash1440 exactly the json bbox prompting is so powerful
@rtrash1440 Enjoying one model does not stop one from enjoying the other. There are advantages of each model - do what suits the use case. No shame either way.
Ideogram is annoying to use and super slow, I could see it still being better though if you have really good hardware and don't mind the complex prompting
Love the krea one! If you ever add to your dataset please add more anal! nobody ever does anal anymore and its so mainstream now. thank you for your work
Yeahh! Anal, Anaaal, AAAAANAL!!! In every imaginable position, please!
anyone know how to train ideogram lora? thanks!
Impressed by Krea2's NSFW capability but it outputs SDXL like quality, it has an uncanny synthetic AI feel because I used SDXL daily for over 2 years. I know that the Krea2 team trained on a lot of synthetic data and its using an inferior Qwen Vae; imo this vae is the biggest culprit to the AI feel, qwen vae sucks and ruined my love for using Qwen image. I suppose there is a niche for that synthetic AI look as a lot of users here love to make that type of image, but I always prioritize extreme realism when it comes to NSFW so Ideogram on those terms is still just slightly superior as it does shadows and skin texture perfectly. Krea2 leans more towards the Z-Image Turbo and SDXL feel, a little too soft.
It's less apparent with some specific style like cinematic or amateur. But yeah it's mostly apparent and Hive Ai detector can see it pretty easily.
With the right settings, you can get high quality photorealistic images from Krea. Use the raw model with the turbo lora at a low strength, with Ostris' Rebalance lora instead of the rebalance node. 16 steps with a cfg of 2, using euler and beta57 is giving me amazing results.
@ObsidianDreams where can I get Ostris' Rebalance lora?
@nonemo13 I got it on the ai toolkit discord. I'm not sure if he's posted it anywhere else
the rebalance node is only good to unlock the nsfw part of the base model. Having the rebalance node active while using this lora gave it that typical AI look with my tests. Without it it totally improved in terms of realism ( you don't need it anymore anyways, because the lora unlocks all the nsfw concepts).
I'm using the turbo model with 10 steps and 3k resolution and always add some stuff like 'overexposed footage' to get more natural amateur like images. Yesterday I thought it's meh, but now I really love it. Especially because it has that general nsfw knowlegde it makes more complicated stuff work easily. no more dicks growing out of stomachs other other bodyparts on blowjob poses, because the model already knows where to place that dick, regardless of pose!
@WhatTheGuy the rebalance NODE does give it an AI look, but Ostris' rebalance lora doesn't.
I agree with you on the Qwen VAE, no idea why KREA used it instead of the amazing flux 2 VAE
@ObsidianDreams ah okay, will try that !
@ObsidianDreams Where can I find that rebalance Lora?
Gemini told me this lora has been mirrored on HuggingFace. I have not verified it yet. https://huggingface.co/lrzjason/QwenImage-Rebalance/tree/main
@garyhoyer2013242 it’s not for Krea 2. Thanks anyway.
Now I’ve found that Mystic Lora gives much better prompt adherence and realism. Give it a try.
@johnsmith20241210 AI Toolkit discord
This 'obvious AI look due to training on synthetic images' seems to be a continuing trend designed to placate law enforcement/politicians. It might also be for gaming the image quality rankings.
@KYoshino Developer claims that Flux 2 VAE didn't exist soon enough.
Softness, in my experience, is not the main problem with images when it comes to them lacking realism; it's exaggerated contrast and exaggerated sharpness; basically, the fried look.
is there any good system prompt for this type of images?
WOW! the Krea 2 version is absolutely amazing! So many complicated stuff just work. For example blowjobs in cars had the dick growing out of all kind of strange body parts before. now it just works! The very best I have ever seen so far! Only part where I see potential for improvements is fluids. Everything else is top notch!
Soon™
really good for v1 hope to see ur updates like ideogram4
Im the only one noticing how krea 2 is low key unhinged ? You can train lora so far but model does heavy lifting...
krea is insane
Prompt enhancer can extend the magic.
@aising23 that's what i've been using or prompt adherence lora
holy shit this is the gg lora i've been waiting for
works perfectly with character loras too ( i set to 1.25 if needed)
euler, beta 57, 12 steps using krea RAW with the kijai turbo lora at 0.6
turbo lora 0.6
realism engine 1.0
character 1-2.5
generate at 4MP, its worth the wait
hey, which kijai turbo lora?
Same question - I don't see any turbo lora from Kijai
@electricpickle thanks bro
what is this character lora you are talking about?
I’ve just started using Krea with the Turbo version and haven’t tried character LoRAs yet. Do you think I’ll need to switch to the Raw version for that?
@electricpickle yes this
@idrogeno821i just prefer the quality using the RAW with turbo lora
krea2 best shit ever
Didn't have much luck with Ideogram and it was super slow, perhaps a skill issue here, but this feels like Ideogram but consistently good and fast this btfos SDXL slop finally I think if you are smart with other LoRAs and in-painting.
If I had more VRAM, I had interesting idea a while ago to code up OpenCode/Codex but for images, just prompting until VLLM says it has image that adheres to prompt, and maybe use impainting or LoRAs or even other models, etc. Just until it gets the desired result. This model combined with the insane amount of SDXL fine-tunes and LoRAs seems perfect for that idea, maybe someone richer than me can make it all work, a uncensored Gemma model would be a good VLLM for this
if you can sample one at a time you can sample both. There are several iterations of this, some were mentioning running a Hermes Agent instance and communicating with it however you wanted. It'd run the LLM, generate a comfy workflow, and run it, providing you an output.
For some reason Kirea 2 version really struggles with pubic hair, and it takes wet pussy a bit too literally everything is wet, has bias for clothing on guys, doesn't do cowgirl well. Otherwise however, insanely impressed, it's crazy this quality of image generator would have blown people away if some major lab released it proprietary not too long ago, but I have it open source* on my dinky little RX 6800 somewhat uncensored
weird, i often got the opposite, only pubic hair that i can't remove.
That's kind of how it's been for a while. Open weights are about / at most a year behind closed api. Commercial models are only in business because most people are too lazy / technically challenged to run open models, and the obscene prices for good pc parts thanks to our ferengi business culture. And probably moreso hype generating funding where demand is nonexistent.
At least with llms, the api has a somewhat reasonable free tier, and competitive open models really are server class. Not so with imagen.