模型介紹為中英雙語，中文使用者可以關閉翻譯。
PornMaster Krea2 Turbo T2I Workflow
PornMaster Krea2 i2i Workflow
PornMaster Krea2 Uncensored lora
Krea 2 styles: ComfyUI-NO8D-controls
如果某些被強力審查生成不出來的主題可以搭"Krea2 TextFusion Refusal-Reduction LoRA"一起使用，也許能夠生成。
官方原版的krea 2 turbo模型在描述某些臉部特徵或攝影風格的提示詞時，會將人物變成2d或半寫實風格，可透過"Realism slider lora"來改善.
"Realism slider lora"可以將動漫角色的lora變成寫實風格.
我已經在我自己訓練的角色lora上進行測試，可以良好的還原人物樣貌並且生成nsfw的內容。
Krea2模型使用英文提示詞來生成圖片的精準度會比使用其他語言更高。
Krea2模型如果使用中文提示詞時，人物會自動更偏向亞洲面孔。
有時會在不需要出現精液的情況中，生成的圖片會自動包含了精液，使用”krea2 raw”的工作流，即可填寫負面提示詞，並將cfg設定為1.5，負面提示詞即可生效。
你也可在負面提示詞排除：2d、Anime.
If some themes that cannot be generated due to strong censorship can be used together with "Krea2 TextFusion Refusal-Reduction LoRA", they may be able to be generated.
The official Krea 2 Turbo model will turn the character into a 2D or semi-realistic style when describing certain facial features or photographic style prompts. This can be improved using the "Realism slider lora".
"Realism slider lora" can transform the anime character Lora into a realistic style.
I have tested it on my own trained character, Lora, and it can reproduce the character's appearance well and generate NSFW content.
The Krea2 model generates images more accurately when using English prompts than when using prompts in other languages.
When the Krea2 model is given prompts in Chinese, the generated people will automatically tend to have more Asian facial features.
Sometimes, images may automatically include semen even when it is not intended. If this happens, use the "krea2 raw" workflow, where you can enter a negative prompt and set the CFG value to 1.5. This will make the negative prompt take effect.
You can also exclude it by negative prompts: 2D, Anime.
text_encoders
qwen3vl_4b_int8_convrot.safetensors
vae
我的其他檢查站和Lora。
My other checkpoints and Lora.
--
我已將模型上傳至runninghub。
I have uploaded the model to runninghub.
Pornmaster_krea2_turbo_fp8_v1
https://www.runninghub.ai/model/public/2070906642396434434/?inviteCode=rh-v1304
Pornmaster_krea2_turbo_fp8_v2
https://www.runninghub.ai/model/public/2070906642396434434/?inviteCode=rh-v1304
Pornmaster_krea2_Turbo_int8_convrot_v2
https://www.runninghub.ai/model/public/2075976521210408961/?inviteCode=rh-v1304
Pornmaster_krea2_Turbo_bf16_v2
https://www.runninghub.ai/model/public/2078120164255870977/?inviteCode=rh-v1304
不要問我如何安裝的問題與comfyui如何使用。
客製化lora:我的Telegram ID:@iamddtla0620
我的微信ID:iamddtla
我的line ID:iamddtla
時間寶貴，請直接說明來意，例如:訂製lora.
Do not ask me questions about how to install it or how to use ComfyUI.
Customized lora: My Telegram ID: @iamddtla0620
My WeChat ID: iamddtla
My Line ID: iamddtla
Time is precious, please state your intention directly, for example: customized LORA.
Description
感謝我的朋友幫我製作了"int8 convrot"版本.
V2版本合併了由我自己訓練的Uncensored lora。
V2版本比V1版本更加靈活，並盡可能地保留Krea2原本官方模型的風格。
V2與角色lora的相容性良好，並且與其他nsfw lora的相容性更好，能夠堆疊更多的LORA一起使用。
複雜性愛姿勢的解剖學比V1版本更良好。
提示詞的編寫方式與v1可能無法完全相同，需重新嘗試編寫提示詞順序和提示詞的權重。
比v1版本更好的保留原版模型中內建的人物。
Thanks to my friend for helping me create the "int8 convrot" version.
Version 2 integrates an Uncensored LoRA that I trained myself.
Compared to Version 1, Version 2 is more flexible while preserving the original style of the official Krea 2 model as much as possible.
Version 2 offers excellent compatibility with character LoRAs and improved compatibility with other NSFW LoRAs, allowing you to stack more LoRAs together with better stability.
The anatomical accuracy of complex sexual poses has also been improved compared to Version 1.
The prompting method may not be exactly the same as in Version 1. You may need to experiment with the order of your prompts and adjust prompt weights to achieve the best results.
It retains the built-in characters in the original model better than version v1.
FAQ
Comments (44)
Congrats on getting the INT-8 ConvRot file up. Did doing a run on runpod or something similar help? I think people will find a well made Int8 ConvRot will outperform. The only thing that might be better and small size would be a mxfp8. However only blackwell like the 5090 can properly utilize it. Thanks for getting a better file up. Keep up the good work. I am waiting for it to come out of early access as I don't have the buzz for it...
The memory shortage problem can be solved by having a friend with a larger graphics card memory.
Awesome... Thanks for doing the upgraded file and thanks to your friend as well.
I bought V2 FP8 version, why should I pay for V2 int8 version again? Did I miss any steps?
The int8 format generates images faster than the fp8 format, but you don't necessarily need to purchase the int8 version. . int8 and fp8 are the exact same model.
v2 has major regressions for some of my prompts. 11k is expensive for something seemingly inferior... v1 is great though.
The prompting method may not be exactly the same as in Version 1. You may need to experiment with the order of your prompts and adjust prompt weights to achieve the best results. v1 has a higher probability of generating sex images even when they are not required to do so compared to v2.
@iamddtla Now tentacles have balls. Deepthroating doesnt work anymore, at all. And since I cant put negative prompts, it's like impossible to fix.
@Zelrikriando I believe the inability to use the "deep throat" feature is due to a problem with your prompt word writing; I was able to generate an image in my own testing.
@iamddtla If you would like to share your prompt, I would be interested. also I am available on discord(handle: zelrik)
@Zelrikriando https://civitai.red/images/136398618
@iamddtla So it looks like my LoRa's have more destructive interference with v2. The link you sent me gave me some workaround ideas though.
@Zelrikriando Alternatively, you can find a "deep throat" version of Lora created by someone else and use it in conjunction with a low-weight version to solve the problem.
Ya bro 11k is just insane
So far testing V1, is one of the best best models!, the prompt adherence somehow is even better than other models, the realism and details are so good, awesome job, I hope V2 is as good or even better
Can you share FP16 variants?
When I run the BF16 version, my computer just freezes. I haven't found any other storage space where my friend can upload the 24GB BF16 file to me.
@iamddtla I may be able to help. What specifically is the problem. Are you looking for something like a dropbox where he could store the file for you so that you can download it?
@heartsinaustin Thank you. My friend has purchased space on Google Cloud. I'll wait for him to release the BF16 model to me when he has time.
@iamddtla Thank you for the BF16 variant. <3
由于没有负面提示词，没法用之前sdxl的方法，画出来的袜子一直默认都是半透明，强调了效果也不是很好，怎么才能稳定的出不透明的厚黑厚白丝袜呢
之前用你做的 Klein 9b 跑 realistic i2i 真的超對我胃口，體驗非常好 想問一下，如果是 Krea2 的 i2i，大大的評價或推薦度如何呢？
I haven't tested the krea2 i2i workflow yet, but I will release the workflow if krea2 it2 works well.
After testing most of the NSFW Krea2 checkpoints, I concluded that this one is by far the best.
Too broke for V2, but loving V1. Any tips on how to keep it from putting random liquid or semen on skin?
Perhaps I trained too many semen images; next time I'll reduce the number of semen pictures.
@iamddtla Second question, if I may. Any advice on getting variations on angle, clothing, etc. when running a batch of the same prompt? I am getting really wonderful results, but if I don't change the prompt the images all basically look the same. I've done some stuff with wildcards for angle, etc. but it always stays very close to the same thing.
@DingleDorf This is the nature of TURBO; it generates very similar compositions unless you use a BASE checkpoint.
@iamddtla Good to know! Thank you!
Sometimes, images may automatically include semen even when it is not intended. If this happens, use the "krea2 raw" workflow, where you can enter a negative prompt and set the CFG value to 1.5. This will make the negative prompt take effect.
It's raining cum, make it stop.
fp16 please ..
When I run the BF16 version, my computer just freezes. I haven't found any other storage space where my friend can upload the 24GB BF16 file to me.
V1 fp8 is fantastic, splendid work dude.
I feel like there's too much pov stuff in training materials. Like it seems that over half of the scenes the model explicitly wants to do pov view. Leading also random pov penises spawning on scenes that are meant to have only two subjects for example; or if trying to do something that is taken from front (let's say front view doggystyle with focus on woman's face) it either forces it to POV shot doggystyle or spawns the penis also in front.
So probably would be nice that on further models to have it be somehow more explicitly triggerable if wanted. POV items seem to be good quality though. But if I am not prompting for it I don't want it either – same problem has been in other nsfw models, however most of those have negative prompt (what I don't think KREA2 has afaik, but if I am wrong, please correct me) making it easy to write a list of things that one doesn't want to see instead of seed hunting which itself is tedious and annoying.
May also be faceless man or man's face out of frame etc... that somehow triggers it often.
You should show the images you generated and the prompts you used so that we can determine whether the problem lies with the prompts or the model. I have seen many example images that can generate images from multiple perspectives.
@iamddtla yea, I generated some (few images) without my LLM enhancer/lazy prompter and it isn't that big of a deal actually when I do completely manual ones, although manual ones seem to not have much variety obviously due to default poses, faces, hairstyles etc... from base model and due to how turbo models in general are; if not detailed – which is why I even have the LLM in between to begin with.
I need to look better my local LLM generated prompts to determine what is it there that is triggering it so often for the model – from reading those hastily they didn't seem to have anything obvious that would be trigger. Which is making me think that model has either of much of it in training data without explicit trigger or there are some random – quite common combinations of words that are triggering it (such as in sex scenes "looking directly at camera", "man with his face out of frame" in certain situations).
I will experiment a bit more at better time to determine which is it and come with some examples.
damn you wont even need lora for this, except if youre going for a different style/aesthetic. good job
有時會在不需要出現精液的情況中，生成的圖片會自動包含了精液，使用”krea2 raw”的工作流，即可填寫負面提示詞，並將cfg設定為1.5，負面提示詞即可生效。
你也可在負面提示詞排除：2d、Anime.
Sometimes, images may automatically include semen even when it is not intended. If this happens, use the "krea2 raw" workflow, where you can enter a negative prompt and set the CFG value to 1.5. This will make the negative prompt take effect.
You can also exclude it by negative prompts: 2D, Anime.
Just came here to say, Amazing work, man. My character Lora stays consistent in pose, and the expressions are flawless at this checkpoint. V1 is awesome; I just wish there was another way to buy V2. Would jump on it instantly. :/
Love that your checkpoints dont fight the character LoRAs. Absolute best checkpoint.