It does require a bit of re-learning how to prompt. You have to be implicit about the stuff you want to see.

I've been fighting to get "dude sitting at the kitchen table" to not chop off his head at 16:9 for a good chunk of today. So far I haven't found the magic words...

@iam440 "subject looks at the viewer. the man keeps eye contact. the man contemplates existence. the man has a face expression of contemplation. the man has shaggy hair and thick eyebrows and a handlebar mustache."

I do know what you mean. I am being hyperbolic with this example, but like each of those sentences alone might give you his face. you just need to explicitly call things out. Essentially when i have a prompt that is cutting parts out, it's either because A) it doesn't mention it, only mentions the body or the action (like seated at a table), or B) it's using some weird description of the setting that would enforce my true subject to not be present. Sometimes this can be adjusted by putting the important part first (ie man sits at table, not table man sits at), or finding "anchor" phrases by cutting away sentences I am most unsure about.

the part about anchor phrasing really is key for any of these modern models. any time I have a really complicated prompt that isn't working, i try and compose it into the smaller blocks of different image generations to fine tune the prompt. I mean in this example, if the man sitting at the table was doing really complicated stuff, you might find you're losing the core point of the photo. If you can find a really simple version of the "a man is seated at the kitchen table reading a newspaper, god rays, glasses", and then compose that within the larger example, it's easier than figuring it all out at once.

The model is much more prompt adherent than anything else, not sure why you get this issue.

I've found this model to adhere to most prompts better than the default krea 2 turbo model

Don't confuse an issue in the model, with an issue in prompting. If you don't write down what you want to see, then sure: you don't get it.

Describe shoes: you'll get shoes and the bottom of a body. Describe hair, hairline, the space above the head, etc. you'll get that too.

Be EXHAUSTIVE, like you're writing a book, describing a new protagonist. Don't write just "a man sitting on a chair". Or, sure, you won't ever get what you're expecting.

@hawkfreedom1177 exactly, everything has to connect for it to look good

There is this Shot Size/Zoom Slider for Krea 2 that might help you out there: https://civitai.red/models/2785002/pornmaster-krea2-shot-size-slider

@crappypornthrowaway Uh, are you sure you don't mean "explicit"?

@PaperTyger No, implicit. Explicit prompting is easy. When I say implicit prompting I mean more, like, phrasing things in such a way that the model knows more about what you mean.
To give you an example, let's say I want a high angle view of a POV blowjob, but I keep getting lower angle views. Repeatedly saying "High angle, high angle" won't help much. But I can do things like say "the floor is visible behind her head", or remove references from things that make it prefer a lower angle. E.g. if I say "Takes place inside a luxury apartment", then the model wants to include things like furniture and windows, which require a lower angle.

In OP's case, if he's getting headless pics, implicit prompting is things like "She is looking at the camera", rather than "head included".

@crappypornthrowaway Fair enough! Thanks for the clarification.

Can the subject wear a hat?

Hard to help you without providing any prompts/configs

PERV

Try the DPO version. Its much better and has its own turbo built in basically. Use it at 8 steps with 1 cfg. The base is more for training on.

@Ada321 thanks for tips.. Agreed, after some more runs on the kroma base at 50 steps, it started to act odd and not produce good runs (as you say more for training). Got lucky at first. B/c i actually prefer RAW krea 2 a lot more than turbo (at 50 steps), I was thinking that logic with Kroma would hold, but I don't think it holds.. seems much more temperamental than the regular raw for image generation. anyway, thanks lodestones.. great work :) I will try the DPO. .(update: tried dpo and its awesome @! )

Euler A might work despite being riskier, and the SGM Uniform scheduler allows a model to work harder on an image because a latent still gets more noise in the middle steps instead of linearly falling off with Simple (and thus resulting in less sharp details). So I am with Euler A + SGM and then upscale with SeedVR2 7B, it fixes artefcats to some extent.

euler + simple - the key is your prompt - describe every detail .. every.... detail... and if you don't wanna go insane have an LLM do it for you lol - i highly encourage you to create a template text document that structures from head to toe every detail a character requires - and then instruct the LLM to fill in any missing details you omit from your initial concept - then go on to environment and style - style being like the camera type used , the focus , the lens, and leave that towards the tail end of the prompt and you'll have a god tier image.

template order of importance: Character A , Character B if applicable , character details head to toe , clothing head to toe details , Actions , Environment, lastly Style. in that order.. you'll probably have to go back and forth with your LLM to create a substantial full fledged non-bias dynamic template that is a swiss army knife it'll take you a bit but once done you'll have it for future uses and your prompting capability will go through the roof...

The best wf I found after many tests..

https://drive.google.com/file/d/1T2oBB_yBUn5SzlB4EHaTkFpISzLA1Tiu/view?usp=drivesdk

And then the OP (Claudi_Orellana834) deleted their account

For me the secret with Chroma was that it needed a lot of steps for acceptable male anatomy. What is a lot of steps? How about up to 512? But some of the results were very realistic considering that it was based in a weak model (Flux Schnell).

@ss9999 Wtf man, who want to run 512 steps, how long are you'r generations lmao, just use loras

@OnyxAI Have you ever tried to generate three naked males and have them all be rendered accurately and as realistically as possible? Even two can be struggle with models like Chroma. What I have consistently found is that "high quality" male genitalia are only rendered with relative quickness when the realism has been dialed back in favor a look that can be rather easily identified as AI. Otherwise, there are always a bunch of flaws. Slow generation time is worth it if you get a high-quality result, versus generating dozens or hundreds of "meh" results. Been there, done that. I started with SD 1.x. Everyone seems to be in love with Krea, for instance. However, when compared with Moss, Krea likes to "AI-ify" everything, much of that due to its built-in censorship. Moss, by contrast, is unfortunately too small a model to really do realism well. Some tricks can be used with models intent on "Ai-ifying" to get a more real look but those results nearly always have bad flaws. Ideogram seemed fairly promising for a small model but I don't have the time/energy to deal with its arcane prompting. Still waiting for a truly uncensored image model that will beat Wan 2.2 in realism.

@ss9999 Man, you're calling wan 2.2 realistic, you're living in the past wtf, wan is so trash in 2026, zit was already way ahead, but was erased by krea. I'm sorry that you never used any good workflow on them, but Krea is by far away the best model all time for now, and moss will probably not beat krea 2 in terms of realism.
It's simple, Krea 2 prompt adherence is awesome, you never got any body morphing or anatomy issue, what do you need more?.
For real I feel sorry for you because you are missing so much..
The only downside when you have bad char. loras is the contrast on skin like you told, but this is easily solvable with a good dataset / training and a good workflow.

I still dont know what you call " realistic " in terms or render for 3 mans, but I just launched a batch of 10, and they all came good, with any imperfection on any pictures, 1mn generation 4k pictures. I don't understand what you mean but any pictures I want to generate come good 90% of the time. The last 10% are textures or small details issues, but this is due to my personnal sampler actually, if you generate with a more classic sampler this is not going to happen. Feel free to come private if you want me to send the results

@OnyxAI I've been doing AI since SD 1.x and I have seen more results generated on this site than most people have. My opinion isn't based on a lack of testing. I have tested SD 1.2–1.5 (plus finetunes), SDXL (plus many finetunes of it), Flux 1D (plus finetunes), Flux (the human-centric finetune BFL labs released... can't remember the name), Chroma, Chroma HD (plus finetunes), Chroma Radiance, Zeta Chroma, Hunyuan Video (whatever version came out around the time of Wan 2.1/2.2... maybe 1.5), Wan 2.1, Wan 2.2 (and its various substandard NSFW finetunes), Wan 2.2 "reasoning" lora and checkpoints, some lousy image model I can't remember the name of, Klein (plus finetunes), ZIT and ZIB, Ideogram, Krea 2/Kroma (only a little so far for Kroma), MiniMax H3, and MOSS/Kromoss. I skipped Ernie, LTX (whatever version that was so big until MM3), and Qwen because they didn't seem worth the bother.

So, yes, clearly I have very little knowledge of any of this and must rely on your chiding to set me right. In terms of how I define realism: I define it as looking at the result and having it be as true-to-live (least amount of uncanny valley) as possible. That's what realism is, after all: real.

The overwhelming majority of the things generated here are at a low standard of realism, including things people call realism. Flat people floating in their environments is a hallmark of much of this, something ZIT was notorious with.

Krea 2 does not beat Wan 2.2 at all, unless it comes to fighting all of its baked-in censorship (blandness, for one thing) and wasting a lot of time in the process. I don't do 1girl content. I generate male-specific content. Lightweight models very likely do better with 1girl (and 2girl). Krea 2 isn't worthless but it frankly really sucks because it's designed to fight the user instead of cooperate. When a model gives more lifelike results with too few steps to generate non-deformed/non-flawed results than it does with more steps that's a thumbs-down from me. Wan 2.2 is the opposite. The more you feed it, the better the results are.

The last point I'll make is that I've seen people post results here that were made with over 200 nodes. If a model is so horrible that it needs huge piles of custom nodes and other trickery that's not good.

@ss9999 Just come private man, I'll repeat myself, but generate males naked with a perfect anatomy is not hard, you just need a lora to unlock krea, a clip prompt, and a sampler, nothing more. my basic workflow is like 15-20 nodes only, if we ommit all the post processing part. moreover, it comes with a custom sampler, so it's 5 nodes just for this. really I don't understand you, you only need 10 steps to generate amazing pictures, and tell that wan 2 beat krea is really showing that you didn't explored the models enough. Trying all the models and checkpoints is cool, but that doesn't mean you're using them well. Just come in private, and show me what you call " realistic " and let me generate any picture for you with my prompts so we will see and agree on what's wrong here

@ss9999 Have you tried @lonecatone23 's Krealism fine-tune? I agree that base krea2 has serious issues. but this IME it's a step above its peer fine-tunes, takes krea2's potential and actually delivers quality realism.

@ravemry9 lonecat is really talented but I don't like use his work, always too complex for results that can be achieved with way lighter workflows

@OnyxAI I can't argue with that cause tastes differ and I don't know your preferences, I'll just say that yeah his workflows are complex but IME his workflows always deliver and they're not complex for no reason it's to achieve the quality output he's shooting for.

TLDR i don't agree his results can be achieved with lighter.

@ravemry9 This is just tastes so. I can't find any realistic picture exemple from his krealism model. But I assume it's probably because professionnal look is mainly linked with " ai look " because of all mainstream models, and most of the time achieving realistic results is in my opinion getting results that look like casual smartphone shots more than perfect professionnal composition pictures.

Civitai gives us no special treatment and generally never replies. If anything, I'm surprised they let it work with the on-site gen. ¯\_(ツ)_/¯

There is a krea2 realism slider lora that works as well. Though often just prompting that its a real color photograph is enough.

You should be using the OPD version of the model for generation.

The base model here is for LORA training and the like. With the base model selected you can see that Lodestone (model author) wrote about it in the small "about this version" section on the right, below the download button:
```

for training purposes, train your lora here then apply it to the turbo one

DO NOT GEN USING THIS MODEL! it will yield crappy result
```

OPD here is the turbo!

Lodestone should probably put this somewhere more visible in the main card :P

@nowantspammo ouch, thanks, I actually used OPD but made the comment under the wrong model...

Loras trained on the Krea2 raw base model work on every finetune.

@demonkingoftyranny Thank you!

If it drifts, up the weight.