Thanks everyone for the Buzz ⚡ the feedback 😊 and for sharing your
creations 🖼️
❤️ If you enjoy my work and would like to support me, consider buying me a
☕ Coffee.
Switch to CivitAI.red to see all the example results and explore everything this workflow can do 😏
A production-ready ComfyUI workflow for Krea2 Turbo, built around a modular, toggle-based architecture that keeps complexity out of your way and full control within reach.
Write a prompt, load a reference image, flip a few switches and go.
⚠️ Jun 29 2026 Hotfix
Replaced a deprecated "Get image size" node (part of the masquerade node pack) with the new ComfyUI's native one. Please, check that the "ComfyUI-Krea2T-Enhancer" is enabled (enabled option turned on), if not, enable it to unlock the model.
💡The images with embedded metadata are stored by default in your ComfyUI output directory, inside a folder named CivitAI. These are the images you should upload to CivitAI so you don't have to manually enter the generation data yourself. Since that information is already embedded within the image itself, CivitAI will automatically fill in all of the generation data for you.
Main features
Three prompt modes: type freely, enhance automatically, or extract a prompt directly from a reference image
2nd pass support: refine generated images with a secondary denoising pass for extra detail
SeedVR2 upscaling: post-generation upscaling with configurable resolution and attention mode
Post-processing suite: noise grain, brightness/contrast/saturation correction, and sharpening, each independently togglable
CivitAI metadata embedding: automatically bakes prompt, seed, sampler, and model info (including LoRAs) into saved outputs
Modular toggle design: every optional feature is its own group, enabled or disabled via Fast Group Bypasser switches
Recommended LoRA (optional)
💡 The workflow already patches the model to remove its built-in censorship. This LoRA gives an extra push for prompts that still need it. Recommended strength: 1.5
COMFYUI SECURITY ERROR FIX
If you hit a security level error while trying to install some custom nodes from ComfyUI manager:
You need to temporarily change the security level in your config.ini file.
If you are using the portable version, the file is located at:
ComfyUI/user/__manager/config.iniChange the value to:
security_level = weak
IMPORTANT:
Do NOT forget to restore it back to its original value, you should restore it afterward for safety reasons.
Perform a hard restart of ComfyUI (completely stop it, then start it again).
Install the custom node(s) using ComfyUI manager and double-check that it is the official one.
Restart ComfyUI and check if the node(s) work(s) properly.
Once everything is working properly, restore your original security level in the config.ini file and restart ComfyUI one final time.
Detailed instructions are contained in the workflow itself:
Red nodes are instructions and useful notes.
Yellow nodes are configurable elements you can adjust to your needs.
Description
First version release
FAQ
Comments (91)
thanks for this workflow, very efficient. I must add some text description If I want the reference image to be truly analyzed, particularly regarding the body shape... but it's a very good job !
Thanks.
This workflow is awesome, truly appreciate it! Cheers.
Thanks for the comment ;)
This is the best Krea2 workflow, thank you!
Thank you ;)
Thank you! THIS is an amazing workflow. I really like the way you designed it. It is easy to understand, well placed and superb!
Thanks :)
Amazing workflow, thank you for sharing!
I found one small issue: the workflow uses "Get Image Size", but the latest LayerStyle registers it as "LayerUtility: GetImageSize", so it shows as a missing node.
Everything else works great. Thanks again for your excellent work!
The "Get image size" node is a core node, not the LayerStyle node. The node was added relatively recently and we don't need custom packs anymore. Thanks for your comment.
*EDIT: I now know what node you are talking about, it is the one inside the metadata sub-graph. It was part of the old masquerade nodes, I don't even have that nodepack installed anymore, I will replace it right away and upload the corrected workflow, thanks for the feedback.
@LatentHeart I'm still very new to this. Most of the fixes I make come from ChatGPT's guidance, and I only learned what Buzz is used for yesterday. 😄 Thanks for the great work!
Love the workflow and just now trying out Krea2. What's the best way to diversify outputs? While they are very good, I feel like a lot of the images are the same unless I blow the prompt up.
Are you using the prompt enhancer? if so, set the seed of the prompt processor to "randomize each time". Krea2 has an amazing prompt adherence, it is one of its strongest features, but it can be also its weakness. A combo of random seed in the prompt processor and a random seed in the global seed can help with that.
@LatentHeart thank you, that seemed to have helped!
Why is this workflow trying to dial out when loaded in ComfyUI?
If you mean why does the workflow make network requests, it is because it fetches data from CivitAI to embed it into the saved image (Read the post). If you don't need this feature, you can safely delete the "CivitAI metadata" group. If you are talking about the "notifications are disabled" message, it is because the system notification node at the end of the pipeline that you can safely delete too if you don't need it.
Like it but I do not know why the image turns out mostly the same. I think have turned every node that has a seed to random. I have 2nd pass enabled, extra nodes and seedvr2 upscaler enabled. Any tips? or is it just the model?
Really enjoying this workflow and all the options, but the one I can't seem to find is batches. Is it possible to do more than one image per seed?
If you make 2 images with the same seed, the same image will come out every time. So more than one image per seed makes no sense.
@LatentHeart It's called batching, look it up. You can make as many variations of 1 seed as your VRAM will fit, my dude. I like 4 options per seed, it's faster than doing 4 separate generations.
@evanneff163 I know what batching is, but again, if you batch several images with the same seed, the images are exactly the same. Now, batching with a random seed, that's another story; you can do that by:
1. In the "Main settings" group, where the latent is created, you can expand the "Empty latent image" node, and select the batch size there, or
2. You can directly batch the images using the built in comfyui's batching system, by selecting the batch size next to the "Run" button (the field that has "1" by default).
For either of those you need to select random seed, if you fix it, all the images will come out the same. This is because the diffusion process is deterministic, same seed, same parameters will always produce the exact same image unless you inject randomness into the diffusion process, but that's another thing. Let me know if I'm misunderstanding you and I can help with something else.
@LatentHeart You're wrong. I literally did it 5 seconds ago and got 4 different images from the same seed. It generated them in a single run cycle and not one of them looks the same. Different angles, poses, faces, etc. Some of them suck, some of them are perfect. idk what you're on about but you're wrong lol.
If I did the same seed again, I would get the same images. But these are all from one seed.
EDIT: Actually, my bad. Looks like batch nodes do incremental generations no matter what the mode is (fixed, random, increment), but if I drag any of the results back into comfy, the source seed is identical. It's much faster than doing 4 separate generations though, which is why I like to do it this way. I never looked into until I read your explanation (which makes sense). You're not wrong, but I'll leave my shame up anyways.
Hate to be negative, and I know you don't have any control over this, but I'm not at all impressed by the image to prompt functionality. what kinda goonbot don't know what pov is 😂
hahaha it happens. Check the generated prompt though, sometimes the prompt is right, but the model struggles with it.
@LatentHeart it was a pov missionary image, and the bot said she was sitting and lying down, and never mentioned "POV"
I definitely took some stuff from your workflow for my own though. Was worth the download!
Any tips to keep it realistic? I see that the seedvr2 upscaler makes the image less real. Any settings to change? Currently using the seedvr2_ema_7b_sharp-Q4_K_M.gguf at 4k with sdpa
I'm not sure if this will help in your case, but I'll share my personal experience. I use the same settings as you described, but the model is not Q4, but Q6. Absolutely photorealistic images of any subject.
Hoping for some advice, installing the nodes, I ran into some errors like the CRT node, saying that it is installed but still not working and when I enable the SeedVR2 Upscaler. Thanks
I often have to run the following in a custom nodes folder before it works:
pip install -r requirements.txt
very cool wf- thanks for sharing it.
Thanks for the comment.
anyway to fix resolution master. latest update of comfyui broke that node causing this whol workflow to be rendered useless
nvm my dam comfy ui install was busted xD
Why is not possible to export as API?
This workflow is not designed to be run in a single simple one straight line pipe line.
I love the workflow, but I'd really appreciate it if someone could tell me how to generate my characters. So far, I've only managed to create random girls. I'd love to be able to add a reference image for the swap. Does anyone have a node already set up?
Use a character LoRA.
what are your speed times? rtx 4080 12gb vram + 64gb. no lora takes about 40sec, with loras goes 60-80sec :C
RTX 4090 + 64GB in RAM, 42 seconds per image for a full 4K generation with 2 sampling passes with a base resolution of 1080p 16:9; or 12 seconds without the second pass and upscaling. This times are once the models are loaded. This model is insanely fast.
Very good workflow, a bit confusing to utilise and some parts it's completely unnecessary to me... but it gets the job done.
Better then the other workflows i used to far.
Thanks.
Hi for some reason i have this:
# ComfyUI Error Report ## Error Details - Node ID: 127:188 - Node Type: ClownsharKSampler_Beta - Exception Type: AttributeError - Exception Message: AttributeError: 'NoneType' object has no attribute 'shape'
Any suggestions?
1. In the 4 main sub-graphs of the workflow, expand them and check if there's something not enabled / bypassed if it doesn't have to be, for every active subgraph, check that nothing is bypassed inside it; for every bypassed sub-graph, check that nothing is enabled inside it.
The four sub-graphs are:
- Prompt processor (Prompt group, although this sub-graph should be always enabled)
- ClownsharKSampler 1 (Main sampler)
- ClownsharKSampler 2 (2nd pass group)
- Image Saver (CivitAI metadata group)
2. Check that the models set in the Models group are the same as the ones listed in the MODEL LINKS note, if you used one or more you already have, check the sha256 hashes to verify they are exactly the same.
3. Check the workflow image in the post, compare it with your workflow and see if you haven't manually bypassed something by accident.
4. Update your custom node packs and ComfyUI.
@LatentHeart Thanks for rep, I found what the problem was. I tried another model. For some reason, this one - KREA2 GPT 逼真版 Grand PUSSY Truth | MIST2 - KREA2 GPT逼真版 | Krea 2 Checkpoint | Civitai give me error. TY for workflow.
Quick question. I want to try using Krea2 with BBOX from Ideogram (I know the coordinates need to be inverted). Is it actually as bad as many people say with krea2?
@KaindeMort Sorry, I have litterally 0 experience with Ideogram 4 and I got no idea how the prompting works in that model.
I'm trying to preserve the subject's appearance as closely as possible in image-to-image generation, but it never seems to match 100%, even after disabling the second pass. Is there anything else I should adjust or enable to better maintain the subject's identity?
This is not an edit workflow, this is image to prompt, not image to image. For perfect character ID I can recommend an edit model like Qwen edit, or using an image to image workflow in combination with the Krea 2 Identity edit LoRA:
Krea 2 Identity edit
Great workflow, but please tell me what node is popping up the notifications so I can kill it with fire.
I don't know... maybe the SystemNotification node? the one with a red box that reads "You can delete this node if you want"? hahahaha
@LatentHeart Where TF is it? I wouldn't ask if I could find it. I've gone over every inch of the workflow.
Deleted it from the JSON now that I know what it's called.
@wild_RPG It was next to the final Image preview node, under the SeedVR2 upscaler node, it has a red note over it hehe
I feel you, it's so annoying if you switch between multiple workflows... and the node it is indeed "SystemNotification"
Another masterpiece! Keep up the good work! 👍
Thank you so much 🙏
Looks excellent without the seedvr2upscaler, but just one problem. while manually inputting 2160 3840 (9:16) resolution using resolution master the generated characters are a bit vertically stretched. how can i get rid of that?
You mean 4K directly from the get go? yeah, I've noticed that too, but I think it is an issue with the model itself, it seems to me that at ultra high resolutions like that it starts to break, it can certainly generate the images but the spatial structure starts to get finicky, I wouldn't cross the 1440p resolution. I haven't tested none standard resolutions though, you can probably fix it by using similar dimensions but in a more squared shape, maybe that could help, but again, haven't tested it.
I've being trying to use your "CivitAI metadata" automation in my own workflow and i don't understand how to make the "tools" and "techniques" set up automatically as well.
Can you please help me out, it seems that it cannot be set up automatically or i'm doing something really wrong because i've tried a bunch of stuff and i didn't succeed.
My workflow: https://pastes.io/2YZEpKjy
As far as I know, CivitAI's metadata parser doesn't look for those fields within the image metadata, so even if you inject them, they simply won't be parsed, so for now, we still need to input them manually. I hope they include those fields in the future, and video metadata too. I could be wrong though, but if it is possible, I don't know how to do it.
I very much like this workflow, mainly because it (mostly) just works. (A color correction node is broken for me rn, but im not using it anyways.)
However, one other thing that doesnt quite work for me, are random seeds. No matter what i try, when i set the seed to random, even when i put in a seed myself, after a few generations it will ALWAYS converge to "1003949590422403" for me, and then stay on that, forever. The only thing i can do right now, is to choose increment the seed instead of randomize it. Im not sure if this is a problem on just my end, but its very interesting and id love to figure out how this happens. Ofc computers are never truly random, maybe something with how this seed node generates randomness is broken (at least now, maybe after an update, idk tho). Ill try to use another seed randomizer node and just feed that into this node, maybe that will work for now.
Usually i make my own workflows, but this one was so complicated that i said screw it, ill just give it a go. And it worked, even tho the problem i always had with comfyui becomes apparent already: It, or some custom nodes, update so frequently, that something WILL break in a few weeks time, and already did. The probably biggest negative about comfyui as a whole.
It happened to me once, but only once, I input a seed my self and restarted comfyui, and that fixed the issue, in my case it was flipping between two seeds, It took me a while to figure out because the image changed but it was just flipping between two values. Anyway, replacing the global seed node with another one is easy, just connect any other seed node to the 3 seed inputs:
- The one from the 1st ClownsharKsampler
- The one from the 2nd ClownsharKsampler
- The one from the CivitAI metadata group (if you are planning on using it)
@LatentHeart Using the rgthree seed node instead and just connecting it to those 3 points works for now. :D
@Bigboyblaziken Awesome ;)
stunning workflow ! but missing only 1 thing , the ( krea2-identity-edit) please add this to your workflow ! it's a must
Thanks, I'll see what I can do ;)
Ugh... No one can ever just make a simple workflow. Everyone tries to fit 5,000 features into every workflow, making things overly complicated.
This is a very basic workflow though. If you need an even more basic one, you can use the ComfyUI's template from the templates menu.
I get what you mean, but thats the main reason to use comfyui imo. To have a workflow that does multiple things at once/one after another. The positive is tho, that a well made workflow (like this one), just works. It should be as easy as using any other workflow really, if you wanna change things around a lot, any workflow you didnt build yourself tends to get complicated fast.
For basic normal images a template workflow should be enough, or using another, easier to use webui entirely, like forge neo.
@LatentHeart Maybe so. It just looks complicated from the images, I guess. It just gets annoying when I'm looking for something basic and can never find a workflow that doesn't try to do it all, especially considering that I'm a noob to ComfyUI.
@Bigboyblaziken I've never heard of ComfyUI Imo. I may check that out. I have a little experience using ComfyUI, but I'm only familiar with the basics. This one looks complicated, but maybe that's just my lack of experience talking. But thanks for the suggestions.
@JayC123 Wait, youre looking at this comfyui workflow, without knowing about comfyui? Ok, then it makes sense that it looks complicated. Comfyui is complicated, and not comfy at all. Workflows like this try to take most of the work away from you, so you can just "simply" make images. But yeah, you wont be able to do anything with this workflow, without using comfyui.
@Bigboyblaziken I mean, don't get me wrong, I use ComfyUI. I've used the Z Image workflows, Wan SVI Pro workflows, etc.; I've even made custom workflows using AI to build the workflows for me. I actually had ChatGPT build a Krea2 workflow since I couldn't find one that I was interested in, and I'm currently adding to it. But I don't know how to put workflows together myself, nor navigate complex workflows that other people have built. I could take the time to learn the ins and outs of ComfyUI, but I'm not gonna lie, having AI build the workflows for me has made me pretty lazy about doing that.
@JayC123 Ooooh, i was confused about the "I've never heard of ComfyUI Imo. I may check that out." And then i realized, you replied to me using "imo". Imo means "In my opinion." Its nothing related to comfyui. :'D
@Bigboyblaziken I actually just realized that 😂 ...I thought that was some secondary ComfyUI platform or something.
Idk if ClownsharKSampler node got an update, but now every generation using this workflow throws an exception at me:
TypeError: krea2t_enhancer_wrapper() takes from 4 to 6 positional arguments but 7 were given
Edit: okay, it looks like the latest unstable version of ComfyUI is causing this. Reverting to 0.28.2 fixed the issue.
It seems an error with the Krea Enhancer node though, maybe updating that node could fix the issue.
@LatentHeart yeah, I was confused because the interface was pointing to the KSampler node, but the error message specifically said Krea Enhancer. Anyway, it looks like Krea Enhancer hasn't been updated recently, so it might be some bug introduced by native ComfyUI.
The same problem. I reinstalled comfy and got the error
I had gotten this as well, looks like it was introduced in the last version of ComfyUI. Reverting to an older snapshot solved the problem, if anyone else runs into it.
How to increase batch size?
Expand the "Empty latent node" in the "Main settings" group, and input your target batch size there.
Excellent workflow! Question if anyone's seen this issue: I'm running into an issue where the workflow just hangs at ~58% in the ClownsharKSampler if any Lora (like the linked NSFW one!) is enabled. The workflow just stays hung there and never completes or throws an error. Has anyone else seen this?
Alright so after some additional testing I noticed that the krea2 model recommended in the workflow links is the fp8 version, but the e.g. NSFW lora is a bf16. After installing the bf16 version of krea2 from the civitai page the workflow was able to complete without hanging, albeit at the longer processing time the bf16 model requires. So it may just be that the model and the lora need to match on that axis.
The fp8 works just fine without loras enabled. Has anyone been able to run that lora with the fp8 model? It'd be great to be able to use those loras with the faster load/process time of the fp8 model.
Great workflow! everyworks fine. Not sure what that Resolution Master do, because it seems like i cant get it to work. So i went and use the Simple Image Size instead >.> Just hoping that Advanced image size can be changed to maybe something like Base Dimension Empty laten image Wx H. Of course it can be changed manually by other editors, overall this is a really Great Workflow where it contains good lora stacker, easy noise editing, and upscale/comparison. Thank you for your good work and time, keep it up! Looking forwards v2
Thanks for the comment. Yes, there's an empty latent image node next to the resolution nodes, it is collapsed but you can expand it, then disconnect the with and height inputs and you can manually input the dimensions you want if that's your preferred way.
@LatentHeart you got a node where i can easily do this but along with like 9:16 x pixel? i remember seeing a node like this for example(dont mind the numbers): portrait /1:4 , landscape/ 2:6, ultra wide / 4; 8, something like that. Or it can just be like 480x480 portrait, 768 x 1032, fullbody.
@RottenFlesh7 That's what the Resolution master is for, in that node you can select presets of standard resolutions, translate aspect ratios, magnet adjust to specific multiples, etc. I'm sorry you are having issues with that node though...
Loras are not working for me.
This is my model: krea2TurboOfficialComfy_krea2TurboFp8.safetensors
Hi LatentHeart <3
Your workflow works perfectly. Would it be possible to add a negative prompt, or is there no real benefit to doing so?
The workflow is designed to work with the turbo version, so no standard negative prompt. You can adapt it to work with the Raw version though, then you can add a negative prompt in the ClownSharKsampler subgraphs.