Overview
ComfyUI workflow using Phr00t's excellent Qwen-Image-Edit-Rapid-AIO.
Works amazingly well for NSFW tasks when using any of the NSFW checkpoints such as Qwen-Rapid-AIO-NSFW-v11.1.
The NSFW checkpoint already has a bunch of NSFW LoRAs baked in, so no LoRAs are necessary for most NSFW generations:
But it works well with other LoRAs such as Anything2Real or Next Scene, which typically won't respect NSFW images when using the base Qwen-Image-Edit-2509 model.
Description
FAQ
Comments (99)
can u upload the lora plz tq
No NSFW LoRA is needed because they're baked into the base model
this is great thank you. is it possible to have it edit just specific parts of an image rather than the whole tthing? it keeps making some odd changes sometimes, and occasionally stretches the entire model in the image
Hmm could you give some examples of the prompts causing these issues? It works well for me, I can edit just the hair, top, bottom, etc. without apparent issues.
The breasts they generate are all very small. Is there any lora that increases the size of the breasts?ty!
Hmm by default the model tries to "imagine" what the appropriate size the breasts should be given the input image. You can try overriding it by specifying the size of the breasts in the prompts, perhaps even emphasizing it like "(huge breasts:1.5)".
A CLIP and VAE are inside the checkpoint itself. Why load them separately if you can connect them from the "Load Checkpoint"?
Oh, my bad. I was re-using my existing workflow with Qwen Edit 2509 (not AIO) and didn't realize that the AIO checkpoint already has CLIP and VAE baked in. I've replaced the workflow with a version using the built-in CLIP/VAE.
Thanks for point it out!
I'm not sure this works with next scene v2, I added the Lora, prompted to rotate the scene 45 degrees and it didn't move an inch, were with the default next scene workflow it always works, using the latest rapid v11 version.
That's weird, what prompt are you using?
It works fine for me when I prompt like "Next Scene: The camera pans 45 degrees to the left" using Next Scene v2 and Rapid AIO v11.4.
And for simple camera rotations, I find Qwen-Edit-Multiple-Angles LoRA (https://huggingface.co/dx8152/Qwen-Edit-2509-Multiple-angles) work better than Next Scene LoRA.
@wildkrauss that worked thanks!
一直出现重新连接
How do I connect and load image nodes?
Sorry, I'm not sure what you mean. The image node (Load Image) is already connected, so all you need to do is click on "Choose file to load" button to load your image
no, there is a SDXL Empty Latent Size Picker
@wildkrauss Am I missing something? I can't seem to find the "Load Image" node, and the image1 input on the TextEncodeQwenImageEditPlus node is also unconnected on my end.
I'm in the same situation.
I can't find the Node for loading images.
I tried running it, and image generation worked.
However, loading existing images isn't possible because the Node is missing.
If you have a solution, please let me know.
It works now, thanks! I had to manually add the Load Image, VAE Encode, and ImageScaleToTotalPixels nodes and connect them as shown in the screenshot.
@arbsec @kot0ha @asray225 Oops I think I've uploaded the wrong workflow (Qwen Image instead of Qwen Image Edit). Can you download the latest workflow and test again?
@wildkrauss thank you!!!
I think you uploaded the wrong workflow. This one seems like it only generates images. It does not edit.
Oops you're right! I think I've uploaded the wrong workflow (Qwen Image instead of Qwen Image Edit) by mistake. Can you download the latest workflow and test again?
Why is the image I generate just a bunch of pixels?
What else would it be?
Hello great workflow , just wanted to ask is there a way to keep consistency when changing scene or action , it seems like when I just change the clothes it works great but as soon as I do a bit more changes it changes the face. I tried different prompt like keep identity or keep face but it still changes a lot. Thanks
i have the same question
So far only black images. Advice?
hi! Got similar problem, have you solved it yet?
@895812718396 turn off sage attention maybe
@bessu216 I'm sorry if that's a stupid question, but how to do that? I think I haven't done any actions to enable sage attention on purpose, or is that on by default?
@895812718396 on the bat file that u run comfy , open it with edit and remove the sage attention if you have
You don't need to turn off sage attention. Just insert the Patch Sage Attention KJ - Node between the model and lora or whatever comes. set it to cuda.
Really late reply but yes I solved it by dissabling sage attention
When the prompt involves complex poses, actions, or scenes, the generated images often struggle to maintain facial consistency. Is there any way to solve this?
That appears to be a known issue, but here are some suggestions from the HuggingFace community:
* start with the scheduler/sampler mentioned in the README (doublecheck !) - this one has bitten myself a few times ...
* start with a simple prompt
* do not change too many aspects of the image at once. Try to change clothing (or outpaint) in one pass to get the person you want, and then use that as the input and change the environment.
* add something like 'keep face as is' , ' professional photography' , 'ultra detailed' either at the beginning or the end of the prompt
Hi! I uploaded an image, and only a totally gray image was generated. I'm not familiar with this situation, do you have any idea?
Update news. I found an error in workflow that caused this. TextEncodeQwenImageEdit
Storage size calculation overflowed with sizes=[4569050144367968256]
Why do all the images I generate only have noise? Do I have to use Qwen-Rapid-AIO-NSFW?
Which version of the AIO checkpoint are you using? Some versions seem to only work with certain combinations of sampler+scheduler.
do you know how to fix oil skin in aio when i use it it always make skin look fake
Which version of the AIO checkpoint are you using? I've noticed that some versions tend to give a more "fake" look to the skin. IMO the v11.4 and v17 are the best so far for realistic skin.
After a few simple fixes, it works fine. (It was working fine to begin with.) I'll upload the result image.
Glad to hear it's worked out!
what specific changes? please tell me. can you maybe share your workflow?
Due to slight changes in facial features and body shape, I used a consistency Lora.
hello，i got the same problem, could you tell me which lora did you use？
You are the answer.
My hero, I have a question
Why do you use Anything2Real or Next Scene loras?
Specifically Next scene lora is just for scenes
Why do you use them??
Oh, they're only used for specific occasions when I want to change NSFW images (anime or artwork) into realistic photos, or to make a logical "next scene" which I can use with Wan 2.2 FFLF (First-Frame Last-Frame) to create controlled animations.
You can ignore them if you don't need them.
@wildkrauss then should I know why next scene lora is good for making realistic photos??
Just experience??
@catpowercity604 you would only use the Next Scene LoRA when you want Qwen to imagine a "next scene" based on the original image.
For example, let's say the original image is of a woman wearing a skirt. You could use the Next Scene LoRA and prompt it with something like: "Next Scene: The woman lifts her skirt and smiles seductively at the camera". Then you could use original image as the First Scene and the modified image as the Last Scene in a Wan 2.2 FFLF workflow to create a video starting with the first scene and ending with the last scene.
But if all you want is to edit the original image with minimal or no changes to the pose, background or expression, you don't need to use the Next Scene LoRA at all.
@wildkrauss Thx my hero God bless you
I am unable to download RES4LYF. Says incompatibility.
Same here, tried almost everything
Tried it for hours too but I fixed it. I was using the standalone exe of ComfyUI. Now I switched to the portable version with the old custom node manager and was finally able to install it.
I also fixed it with the portable version, deleted the desktop one and downloaded the portable one, exactly the 0.10
So, i am using the v19 of the checkpoint. For me the images look very artificial and missing much realistic detail compared to normal Qwen2512 Text2Image where i can run 50 steps and 4 cfg. When i try here to raise the steps its breaks the image at above 15 steps. Also hair on legs and stuff look highly arteficial, almost "mosaic" like. Also many many deformations in hands, feet and legs. Is there a way to get much more realistic looking images? i am using the recommended sampler/scheduler for v19 inside a basic KSampler.
This is amazing, I didn't know that wen edit even exists... yes im new, learning, and this is mindblowing
cam some1 help me? Every generated photo has moles on the breasts and genitals. How can I remove them? I'm struggling with this 😣 everything to both the negative and positive prompts to exclude moles and pubic hair, but it still includes them anyway, "I'm also loading a LoRA, but it keeps the base image's face. Why is that? 😵
Same here; I am experiencing the same problem too. Also, Qwen cannot generate breast sucking or huge breasts at all. Its Huge breasts are large breasts by every standard.
Same here
My lord AIO model version is 21v
Do you have any plan to update this workflows?
Is it normal that the rest of the details in an image lose detail/ become blurred? Is there a way around it? Sorry, this is such a complex program, and it's a miracle I got it working at all lol
Great overall with only one issue, sometimes it generate nasty spots on photos, not sure it from workflow, lora or base AIO models. The spots only on bodies and clothes, not on background.
im getting"os error 1455" on the load checkpoint node, and it says that "The paging file is too small to complete the operation." What should i do?
google how to add a page file. do the ammount of ram you have x1024 and then do x1.5 and x3 of that # for the min and max page file values
So, the size of the resulting image is decided by the ImageScaleToTotalPixels node, and the resolution that you can input in the comfyui-image-saver node it's just the saved meta data, correct?
Worked Great for me. Still experimenting on different photos.
it's so powerful，thx for sharing it
is there a LORA for NSFW image edits around somewhere? I'm using the nvfp4 edit model so I would really like a LORA only for this..
help i keep getting reconnecting after starting to run the workflow
You need to increase your virtual memory, ideally to 50GB or more.
add paging vram and run with --lowvram works for me
I got Stucked on "restarting backend to apply change" for the clownsharKsampler_Beta, please anyone can help me?
hello, help, I have used it many times and for the past few days it has stopped working. I work with ComfyUI in the cloud (VAS AI). I went back to ComfyUI version 10 and even using the same workflow it only generates a black image. I changed the checkpoint to other versions AIO nsfw v23, v18.1 and it still doesn’t work. Does anyone know what the problem could be?
I was able to solve this problem by installing the nightly version of each node.
@scastel25441 Can you give me a list of those nodes please?
preview image doesn't work??
When using the latest Qwen Image Edit Rapid AIO V23 to generate characters, I encountered skin noise and small black dots. After switching from ClownsharKSampler to comfyui's built-in KSampler and applying the recommended euler_ancestral/beta settings with steps=4, the noise reduced—though I'm unsure why. Will you continue updating this workflow? I used Deepl for translation; please bear with any inaccuracies.
I'm wondering if I should switch to 11.1 or maybe this workflow needs to be updated somewhere? I'm on v23. Penises look really weird, like they're made out of plastic.
Otherwise this works really well.
It is not the workflow that's causing it.
I have run into the same issue, it can be resolved by using this lora at 0.3 strength, make sure you are using the word "PENISLORA" at the beginning of your prompt. Just know sometimes the penis can look backwards if the the camera is looking top down at the penis (I think the dataset was with penises viewed if looking at the man) https://civitai.com/models/1476909?modelVersionId=2510724
That lora does indeed fix it. Thanks
@MysticalOne The link not found. Would please check it, or tell me the exact name to search.
@Starfish88 qwen_penislora_v2_000021000.safetensors
Hash (sha256) ee303b20aad39cdbf6db694eeffc6259a8f297acc538739948a51a51ac74982f
Can a mask be set for areas not to generate? This keeps changing the hair and face so it turns out to be someone completely different
Seconding this, I'm encountering issues with the generation replacing portions of the initial image that I don't want to change.
missing nsfw lora loader, looks like I need to have a masters degree from Harvard to understand this workflow. Does not work
If you're missing nodes you gotta install them manually, either by using; git clone %git URL of node pack% or actually manually installing, this workflow is the simplest workflow I've come across and it does exactly what its suppose to lmao
Once I downloaded all the Reqs it worked great with the suggested prompt. Imagine lost a little sharpness, but was definitely true to the subject.
I am using AIO ver23 // But with this workflows... // it can not make pussy hair well // looks very artificial // Any body recommend me good prompt or other methods plz.... T.T // it only makes well clean pussy...
I am also having this problem
@microsoftccsl001 Do you solve this problem??
Im getting these weird Scan lines?
I know flux had this issue once, how do I fix this?
I do see in your Example you are also getting the stripies "Scan lines" in your output.
Have you found a solution for this yet?
It works! Thanks.
But it's too slow, and the original image gets distorted. Still, that's okay.
Has anyone figured out the black dots fix yet?
Using this is just generating black boxes as an output. What gives?