模型介紹為中英雙語，中文使用者可以關閉翻譯。
--
PornMaster [Flux.2 Klein 9B] - image-editing
PornMaster [Flux2 Klein 9b base] t2i Workflow
PornMaster [Flux2 Klein 4B] High-resolution restoration Workflow
PornMaster [Flux.2 Klein 9B] - Anime to Realism Workflow
PornMaster [Flux2 Klein 9b kv] t2i Workflow
--
cosplay:
Pornmaster anime to realistic - Klein+ZIT
--
--
使用英文提示詞的精準度會更高。
主要針對基礎模型中的NSFW主題進行改良，沒有強烈的改變原始基礎模型的風格。
可用於文生圖和圖片編輯，並且文生圖和圖片編輯功能皆已可以正常生成NSFW題材。
使用自然語言描述提示詞時效果較佳，使用”danbooru”方式的短標籤提示詞效果較差。
Flux.2 Klein 要生成良好的圖片需使用較長的自然語言句子來描述，並且使用專業的攝影術語時可以得到質感更好的圖片，例如:光圈、快門速度、色溫、色調等術語。
對於姿勢也需要詳細的描述，例如:左手和右手放置的位置、手臂和手肘和膝蓋是否可見、手指張開或是五指併攏、大腿或小腿是否可見，左腳和右腳是否可見。
編輯圖片時，如果你要轉動人物的臉部角度和改變表情時，請先使用"Nano Banano" 或是"Seedream"、"gpt image"進行編輯，flux2 kelin對於人物臉部角度的改變和表情的改變效果非常差。
當與NSFW的LORA一起使用時，LORA權重可從0.05開始測試然後逐漸增加權重。
建議使用” Turbo lora”，4步驟即可生成圖片，將CFG設定為1時負面提示詞不會生效，但生成圖片的速度較快，將CFG設定為2時負面提示詞會生效，但生成圖片的速度較慢。
如果不使用Turbo lora時，參數設定的方式請參考原版的”flux2 klein 9b base”。
注意:V2版本不需要使用Turbo lora。
使用Turbo LORA時:
CFG設定1，然後4步驟，品質較低速度較快。
CFG設定1.5，然後8步驟，皮膚紋理較真實。
編輯圖片時建議使用此設置，CFG設定2，然後4步驟。
可以使用基於Flux.2 Klein base 9b所訓練的人物lora。
Using English prompts will result in higher accuracy.
Primarily focused on improving NSFW themes in the base model, without strongly altering the original style of the base model.
It can be used for text-to-image generation and image editing, and both text-to-image and image editing can now properly generate NSFW content.
Using natural-language descriptive prompts yields better results, while short tag-style prompts in the “danbooru” format perform worse.
For Flux.2 Klein to generate high-quality images, longer natural-language sentences should be used for descriptions. Using professional photography terminology can further enhance image quality, such as aperture, shutter speed, color temperature, tone, and similar terms.
Poses also require detailed descriptions, for example: the placement of the left and right hands, whether the arms, elbows, and knees are visible, whether the fingers are spread or closed together, whether the thighs or calves are visible, and whether the left foot and right foot are visible.
When editing images, if you want to rotate the character's face or change their expression, please use "gpt image"、"Nano Banano" or "Seedream" first. Flux2 Kelin is very ineffective at changing the angle of the character's face and their expression.
When used together with NSFW LoRA models, it is recommended to start testing with a LoRA weight of 0.05 and then gradually increase the weight.
It is recommended to use the “Turbo lora,” which can generate images in just 4 steps.
When CFG is set to 1, negative prompts will not take effect, but image generation is faster.
When CFG is set to 2, negative prompts will take effect, but image generation is slower.
If Turbo LoRA is not used, please refer to the original “flux2 klein 9b base” for parameter settings.
Note: The V2 version does not require using a Turbo LoRA.
When using Turbo LoRA:
With CFG set to 1 and 4steps, lower quality but faster speed.
With CFG set to 1.5 and 8 steps, skin textures appear more realistic.
When editing images, it is recommended to use this setting: CFG setting 2, then step 4.
The character lora, trained on Flux.2 Klein base 9b, can be used.
--
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_flux2_klein_9b_base_fp8_V1
https://www.runninghub.ai/model/public/2030515890512465922/?inviteCode=rh-v1304
PornMaster_flux2_klein_9b_base_fp8_V2
https://www.runninghub.ai/model/public/2030515880504856577/?inviteCode=rh-v1304
PornMaster_flux2_klein_9b_base_bf16_V3
https://www.runninghub.ai/model/public/2031377858077138946/?inviteCode=rh-v1304
PornMaster_flux2_klein_9b_base_fp8_V3
https://www.runninghub.ai/model/public/2031653325136990209/?inviteCode=rh-v1304
PornMaster_flux2_klein_9b_base_fp8_V4
https://www.runninghub.ai/model/public/2054800384882946049/?inviteCode=rh-v1304
--
text_encoders
· qwen_3_8b_fp8mixed.safetensors
vae
--
--
模型許可：
本模型請在遵循 black-forest-labs 公司非商業授權 (FLUX Non-Commercial License) 下使用。
This model falls under the FLUX Non-Commercial License.
--
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
V2版本不需要使用Turbo lora。
也可透過使用-1權重的Turbo lora，去除加速功能。
V2版本可以堆疊使用更多的lora。
當與NSFW的LORA一起使用時，LORA權重可從0.05開始測試然後逐漸增加權重。
CFG設定1，然後4步驟，品質較低速度較快。
CFG設定1.5，然後8步驟，皮膚紋理較真實。
編輯圖片時建議使用此設置，CFG設定2，然後4步驟。
可以使用基於Flux.2 Klein base 9b所訓練的人物lora。
--
The V2 version does not require using a Turbo LoRA.
The acceleration feature can also be removed by using Turbo Lora with a weight of -1.
The V2 version allows stacking and using more LoRAs at the same time.
When used together with NSFW LoRAs, it is recommended to start testing with a LoRA weight of 0.05 and then gradually increase the weight.
With CFG set to 1 and 4 steps, the quality is lower but the generation speed is faster.
With CFG set to 1.5 and 8 steps, the skin texture appears more realistic.
When editing images, it is recommended to use this setting: CFG set to 2 and 4 steps.
Character LoRAs trained based on Flux.2 Klein Base 9B can be used.
FAQ
Comments (70)
Hello, I gott his error with the V2:
"ERROR: Could not detect model type of: C:\ComfyUI\ComfyUI\models\diffusion_models\pornmasterFlux2Klein_v2.safetensors "
Do you know what my proble could be ? I tested it with your workflow and had no problem with v1
Thanks for your work ;)
awesome model, best available yet
Did it work for yoU?
@rlm77dd ofc
thanks for the mode, i just used 25k buzz on it.
please add bf16 or fp16 as well
"ERROR: Could not detect model type of: C:\ComfyUI\ComfyUI\models\diffusion_models\pornmasterFlux2Klein_v2.safetensors " Getting this error. are you sure you uploaded this correctly? I see another person with same issue
edit: it doesn't even work in the workflow that you linked. can you message me on discord to send correct version so I don't have to pay again? leex2579
@rlm77dd For users encountering the "Could not detect model type" issue, please update comfy ui to the latest version.
@iamddtla is there any solution to same problem on forge neo?
You might try downloading the model again; there might have been a problem during the download process from Civitai.
This situation can be resolved by updating the COMFYUI ontology to the latest version. My old version of the ontology cannot recognize it either
This is so good, thank you for making this. Works great wit image editing as well.
v2 isn't working for me how are you running it?
Nicely done and better than Big Love.
I tested both version and prefer the v1 more than v2. V2 changed the face/style too much and V1 has better style retention.
Hopefully for next v3 author could make some adjustment to bring back the v1 vibe. Will stick with v1 for now.
And for those who experienced the "Could not detect model type" issue, please update comfy ui to the latest version.
Are you referring to the character Lora when you say the face change feature? Or is it a feature in the editor?
@iamddtla I mean the image edit workflow with klein. When performing an edit with this lora enabled, I noticed that the face/image will shift from the original image. The V1 lora does better and it preserve much more details from the original image ,and I noticed a bit more shift with V2 lora.
@lyn444675 Are you using nsfw lora? In my own testing, versions v1 and v2 showed the same stability when editing images.
@iamddtla No additional loras except for v1 I have turbo lora enabled.
Not sure what you mean by stability here. For me the the look differences are significant for all my use cases. V1 will produce the output that resembles more of the original look, while V2 will produce ones with slightly different hairstyle, facial structure which are subtle on their own but they add up to be quite noticeable.
@iamddtla I think what you can try is ask for AI to repose a character (with large pose change it is more obvious) completely, with different face and body angle. and then you can tell the differences. Hope this helps with debugging.
@lyn444675 Perhaps you could share a tooltip you used when editing a character, showing how to change the angle and pose of their face, so I can test it out?
@lyn444675 I believe that when editing characters, "Kelin" is far less stable than "Nano Banano" or "Seedream" in terms of face rotation. The best approach is to first edit the face angle using "Nano Banano" or "Seedream," and then use "Kelin" to edit the 'nsfw'.
Can you make Q4_K_M or NVFP4 for those who only had 8GB VRAM?
Is there any negative prompts that reduce the typical body horror, especially on edit mode? Have you figured any? I take it that klein is still evolving, but it seems like approx 70 % of images generate either missing/extra limbs, amputees or adjoined body parts, twisted poses or at worst, something straight out of Lovecraftian horror stories.
I guess Klein is very good if you wanted to create some eldritch horror without much prompting effort.
I feel that the turbo lora adds more to the issue, by leaving some limbs unrendered for example.
Carefully read the model's instructions. If you are missing the "left hand," write "left hand visible" in the prompt. Do not use negative prompts to correct the anatomy. Klein must describe in detail the location of your limbs and whether your hands and feet are visible.
pornmaster's models can be very good but most of the time they require a very convoluted setup to function. Thankfully the loras and illust/noob models are always easy to use.
I have best results with 2 cfg but nothing in the negative text prompt. Generally all anatomical problems are due to the positive prompt, specifically, two sentences overlapping and getting confused and printing the anatomy for both at the same time. The CFG>1 will help it resolve, but not always.
@LetTheBassDrop this is a klein problem, all of the NSFW content isn't weighted deeply yet
@happylittleteapot Oh yeah, and this model is exceptional at what it does. I've had great results. I was just speaking on his past models. Sometimes you DO need to actually read what these people write in the description.
@iamddtla My message came off a bit flippant and rude I think - I absolutely LOVE your models and didn't mean to come off any other way. I know the word convoluted can have negative connotations but in this context I simply meant to convey that it's important to at least take a cursory glance at a model's description. This model is fantastic and I'm enjoying it very much. You're one of my favorites and I always look forward to your creations.
Many models have this problem. The solution is to first generate the desired image using a relatively stable N.S.F.W. model, then convert it to a depth map, and finally use the Flux.2 Klein editing mode to reference the depth pose map.
i can't get it to work on Wan2gp, the text encoder fails, even with the one you provided, might be just the program's fault
Excellent! This is great, works great for edit too.
Thank you!
@panorios how is the face consistency with edit.
@nikolaibloom805 Premium, does only what you ask (***), rest of the image is the same.
I can't run it on forge neo, is there any solution to this problem?
ValueError: Failed to recognize model type!
I haven't tried using "klein" on Forge Neo; you can send the program's errors to GPT.
thank you so much for these models, I just want to point out that I get way better results when not using the turbo lora you are suggesting for v2.
hah, I saw you said v1 only requires turbo lora my mistake.
Wanted to say thanks for V1. For some reason I can crank the CFG on this higher than on regular base without degrading the image (at least in some circumstances). Downloading V2 now.
THIS MODEL ARE GREAT!!! NO NEED ANY ADDITIONAL LORA NOT EVEN THE TURBO ONE.
SUGGESTIONS: EULER A - 4 STEPS | RESOLUTION - BASE FOR 1:1 4:3 ETC....
V2 is giving a failed to recognized model type error in a updated comfy - V1 works great however, so left scratching my head. Even tried your Pornmaster WF as well as the official klein WF and sadly error persists.
Curious if anyone has ran into also with comfy?
You might try downloading the model again; there might have been a problem during the download process from Civitai.
@iamddtla one of the first things i tried. 3 times time, error persists. I also did suspect it being a server side glitch but no actual way to test against that sadly.
i'll wait a few days and try again and update this message if error goes away. nothing gets fixed over a civitai weekend so won't bother till after that lol.
这种情况是需要更新到COMFYUI 本体到最新版就可以解决，我的旧版本体也不能识别
ERROR: Could not detect model type of: ComfyUI\models\diffusion_models\pornmasterFlux2Klein_v2.safetensors
You might try downloading the model again; there might have been a problem during the download process from Civitai.
@iamddtla
@iamddtla Ok, now it works, but works far worse then V1. Maybe Turbo Lora was and issue.
我也是这种问题，下载了三次升级了comfyui,现在还是不能用
@Eva033 應該是comfyui的bug，你可以先使用v1版本。
I must admit, this is a fantastic checkpoint. It can generate characters of different ages and in various poses, and the facial features remain unchanged, there are no obvious anatomical problems, and the effect is extremely realistic. I look forward to your continued updates to this project.
It‘s a simple merge Base Model + 30% SNOFS 1.1 + 30% Unchained V2
@CupRunethOverWithLSD Thank you for telling me this. I'd also like to know what LORA the image quality is based on?
@hanekawa0917792 It really is only the Base Model and those 2 Loras, nothing more.
@CupRunethOverWithLSD thanks
ValueError: Failed to recognize model type! on V2 sadly
I don't know why you're getting this error. Version 2 worked perfectly fine with comfyui. You might try downloading the model again; there might have been a problem during the download process from Civitai.
@iamddtla Still same problem after re-download but V1 is working so, that's strange
same error on Forge Neo
I'll create a v3 version and see if it improves the problem.
Okay good to hear I wasn't the only one with this error, Was wondering what I did wrong. Hope it will be fixed.
@Yuriji @Neo2323 @sladderklittan152 Updated ComfyUI again to stable release branche, and ok, now it works, but works far worse then V1. Maybe because of merged speed-up lora.
@eliont The acceleration feature can also be removed by using Turbo Lora with a weight of -1.
It's likely a bug caused by the comfyui update; I myself am also unable to read the official original version of Flux2 Klein 9b base fp8.
This situation can be resolved by updating the COMFYUI ontology to the latest version. My old version of the ontology cannot recognize it either
Amazing model really outperform however can you share BF16 version for us with 30series card. Ty
Even better would be INT8, fastest speeds on 30XX cards