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
New concepts added
FAQ
Comments (48)
really loving V3, but V4 has problems now. V4 seems to be overtrained. It alters characters now. Also general stuff like having the woman sitting inside the bathtub while blowjob doesn't work anymore. she's sitting on the ground in front of the bathtub now. V3 was better doing that <- I'm really loving V3 ! Maybe you can fix V4?
EDIT: It was user error, had my CFG at 5, turned it up to 7 and the big boobs came back, they still look a bit goofy like a bad boob job type tits. I'm guessing there are a lot of fake tits in the training data.
Im having a lot of trouble with big natural boobs in v4, either they're small or comically huge. Could be user error and probably just more trial and error experimentation to dial in the prompt.
I notice a lot more extreme hardcore concepts implemented into V4, probably why I'm having difficulty with my older prompts that worked fine in V3 that were mostly softcore. I feel like the addition of the new extreme concepts made the lora unstable. You can only push a lora so hard or stuff it with a number of multiple concepts before/until it starts to break and you have to resort to a full fine tune of the 9 gig model. I think its starting to reach that limit with the amount of concepts it can understand. I feel like its forgotten some concepts that worked in v3 and replace it with the concepts of v4.
v2 by far gives me the best results.
Great work! By far the best LOra for realistic bodies! V3 and Lenovo are wonderful together. For now I like V3 more than V4, but I have to test it more since I just started using it. for sure I can't use the same lora strength
V4 samples went a little.... off the path.
I'd like to see an expanded V2.
whatever you did in V4 made my gens worse. V3 and even V2 are better. Just giving feedback. Love the loras! Thank you for your work, your loras are the only reason I can have fun with ideogram right now
Thanks. I tried something different for V4. From my testing is was more stable. But I see a few feedback that it might not. I'll probably go back to the old way in V5.
Thank you, this is a great addition to an incredible model. Would you be able to add more control over the vulva? Like knowledge of inner labia, pubic mound, clitoris, clitoral hood, etc? If I'm missing something please feel to give any tips in this area. Thanks
Anybody had success in fine control of the ethnicity of female subjects in Ideogram4, please do share your experiences.
it is a little diffcult as changing the ethinicity can sometimes break the art style/medium, and it may require some trial and error
@Zachamos777 Thank you got some success using sample hints from Claude.
I have never had a problem but I believe I only ever reference their ethnicity in their bounding boxes alone. I'll generally specify the location in the other places, like "background" or "desc" or "aesthetics". It must have done enough to keep it separate. Although I will say I simply say "They are {American|Chinese|European}" at the end of the bbox descriptions.
All of the images I have with good bounding boxes work great with V4. However, any time a bounding box is too big, placed wrong, it does its best to match the box. This was the first time getting 6 fingers and toes, for example.
V3 started doing this as well. In other models we might say that it is overtrained, trying too hard to jam anatomy in that doesn't fit. But in this case, we're asking for the anatomy to be in an incorrect place. I suppose it still means it's overtrained in some capacity, but I don't exactly know which. Like for example, if people want more creativity, you'd want less bounding boxes. But with more bboxes, and correctly placed, the outputs should improve. Which I honestly feel like V4 is improving the outputs.
This should get interesting, is all I'm getting at. Maybe if an image could have more than one JSON prompt, or if it was randomized throughout the training set. Some get more bboxes, some get less.
I came to the same conclusion with my testing, I didn't get as much issues as reported because of the accuracy of my bounding boxes. Once I give the model more liberties, there's more chance for errors. I like your idea of DS duplication with low and high bboxes, something to try.
@razzz in my experience training with multiple captions per image (especially when auto-generated from very different tools) does in deed help to make a LoRA more universal.
But so far I haven't trained Ideogram4, so I don't know how this is behaving and what's working well and what isn't.
Using V4 I can easily generate a woman, but the nude man placed next to her is sadly missing his penis, although he has balls.
Happens in many different prompts.
It's an issue that I'm checking on solving. Meanwhile you can mention the penis in the bounding box of the man to get around it.
this happens to me almost everytime too, not just with this model but with most NSFW models except for Chroma lol
I get floating penises instead of attached to body.
Yeah, not just Ideogram: F2K9B, ZiT, etc. Detached penis, often being held at arm's length. Frustrating trying seed after seed to maybe get one.
how do you guys generate the proper description json inside the worflow? i'm using an external script in pyton made with Claude for that, but is much less accurate in compare your default prompt? thank you, impressive lora by the way.
Thanks. I'm using Gemini but any capable LLM can do it. You can use the official ideogram system prompt to make sure to get the most of it.
https://github.com/ideogram-oss/ideogram4/blob/main/src/ideogram4/magic_prompt_system_prompts/v1.txt
thank you very much razzz :) I will try it. cheers :D
Honestly just write them. I'm finding most of my natural language prompts are already one step from JSON anyway, almost each sentence corresponds to a character or object. For sentences that are multiple, I just leave in the description. If I were to rewrite my "Ideogram 4 Prompt Generator" context, I'd tell it to ignore json entirely, but instead just give me the sentences that correspond to the inputs for me to manually create each bbox. I found even if most of my bbox values were somehow not entirely hallucinated, they'd still be weird.
The best IG4 lora so far! Can we please have innie pussy next version? 🙏
Thx. Not in the next version but the one after.
V4 is underrated
Thanks for the model! The only issue I've run into so far with v4 is the rotation. Almost every time it generates the anatomy of penis with the wrong orientation — it seems like it only knows how to render one side. The tip always looks like it's being viewed from above. Even when you generate it from below, the tip still appears as if it's seen from above.
I have this problem as well. v3 is much better in this regard.
BBL, wide hips concepts🤌?
Hell yeah
Yes please bubble butts ftw!
Civitai added ideogram as base model, please change it.
I won't. There's a reason why no one with a NSFW Lora did the change it to Ideogram. You won't find any nsfw Lora under ideogram model category ever.
@razzz Why is that the case? Why would Civit do that?
@busy_ad6969 They are forced to.
@busy_ad6969 Licensing terms!
@busy_ad6969 Didn't know that...what a cucked model lmao.
How do you write your prompts without the KJ nodes?
I use an LLM with the official magic prompt from Ideogram https://github.com/ideogram-oss/ideogram4/blob/main/src/ideogram4/magic_prompt_system_prompts/v1.txt
@razzz what llm is sophisticated enough AND does nsfw?
@cobaltpixiv520 Gemini through API is fully uncensored or Grok.
@razzz Wait I have been using both free and paid Gemini API and it has never managed to even caption my nsfw dateset O_O
@cobaltpixiv520 You doing actual API call or you're using their UI ?
@razzz actual api from cmd
@cobaltpixiv520 Make sure the safety settings in the payload are all set to BLOCK_NONE.
@razzz Wow it knows bobs and vagene!
man this shit so fucking realistic. W dev team