Are you using any other loras? If yes disable them.

Make sure you connected everything correctly (including lora loaders). Cond and uncond need separate loaders so if you did that make sure you connected them correctly.

Use correct strengths (0.8 + 0.3).

My suggestion is to start with a fresh correctly setup WF with only realism engine lora and see if you still get bad results.

@Unscathed7928 Thanks. No I've been using the provided workflow as is without any other LoRAs. The outputs are not bad generally speaking, they're fine to very good. It's only the subjective appearance of the women is not exactly my taste. I was just wondering if there's any best practices to face prompting with Ideogram and/or this LoRA.

Yeah getting the same results so far. Lots of puffy fake plastic surgery botox lips.
Prompting it doesnt really correct it either. Not really sure whats triggering it because none of the examples have it which is weird

@fatberg_slim I mean if you're looking for generic AI face like the one in your pfp then this model prolly wont do that without a dedicated lora

@razzz fp8 but I think its token related. When I prompted for a completely different scene/woman it went away. Also noticed blonde/beach brings out a generic face so maybe it just overfit on some concepts. Its a great lora regardless tho

I find that mixing Anime training into a realistic checkpoint always makes realism worse, just use the right tool for the right job.

@J1B That's kinda the point??????????????????????????????????????? Yeah use the right tool for the job, use the right LoRA for your image

Gemma-4-E2B-Uncensored-HauhauCS-Aggressive

@stasisirene I tried that, but I don't think my system prompt is good enough. The anatomical results are horrific. Will you share your system prompt?

@koalabearr I just copy and paste a working JSON in the chat, then I ask in the same message to generate a prompt using that structure, specifying what I want. It works fine. If you don’t like where the bounding boxes are placed, you can move them manually

@koalabearr Most have mentioned the ones provided by the developers. I simply wrote my own that included an empty JSON object, then one filled in, and my own descriptions for the fields and the model. I definitely had to "beg" it to make sure it wasn't hallucinating bbox values and to know what size/ratio of photo to create, to make sure its prompts weren't too verbose, and to make sure it always kept the same structure as the provided JSON blob (cause it kept hallucinating its own keys forever lol). But with those three things I reinforced, it eventually consistently gave me blobs that worked. I used a Qwen-based model locally. Creating my own context kept the token length much shorter, and much more focused.

It's not that the outputs don't work, I'd just rather not run an LLM alongside ComfyUI and everything too. Perhaps with a bigger LLM base model (I think I used a 20+B param model) the generated prompts would be more creative and nicer. But I just manually create the JSON with the KJ prompt builder instead. My own prompt writing tends to be better than the LLM anyway, it takes all of 2 mins max to make a full blown prompt.

If you struggle with ideas, that I find is the best to have the LLM quickly hash out those in more human readable text (so you can use it on any LLM), then you can go and fill in the prompt builder yourself.

Okay, thank you very much, I'm really looking forward to it!

I have serious troubles in general with clothing. That's really all that's seemed to get worse for me as the versions have increased, but even other lora combos don't really work. Some prompts, with some specificity, I can consistently reproduce creative clothing. But the majority of the time, especially now, most of my people are always nude. Even when told to have clothes or even with the boxes and everything. I'm beginning to think maybe not including colours might be the reason some of these prompts fail, as that would really force a distinction. In worst cases I just mess around with bounding boxes and words until it increases the success rate. Definitely a worse experience than other previous models when it doesn't "just work".

@makiaeveli try setting it lower. It actually helps with some skimpy but still sfw clothing. Like if you want alot of cleavage.

@Starry_Eyes That used to work really well. I'd keep it at 0.4 for the softcore and 0.7 or so for the hardcore. But even putting the lora on both models I'm getting the same results of showing both nipples at both strengths. Obviously the other versions and stuff are a solution, I'm not trying to complain lol.

Realism engine lora already exists for Zit,Zib and F2k

@razzz how would that work exactly? really intrigued by the idea but isn't this just for merging checkpoints or full models rather than merging loras and models together?

This model is extremely powerful and it can do stuff that wasn't even in the lora dataset.. you just need prompt it correctly. You do can literally ANYTHING you imagine if you put enough time into it. I've posted an example of a thighjob in the gallery.

You can add 3 boxes whatever they contain to bypass the safety filter

@aramintastudio it works! thank you so much!

better that than the lora gets removed from civitai (like the one from sexgod)

@Aecroo 😲