What is this?
This is a simple workflow for generating high quality, realistic images at high resolution using Krea 2. There's also an optional full-turbo version of the workflow, which is not suitable for realism (or creativity) but is handy for some things. Below in this post there are also some tips & a lot of info about the model.
The sampler & lora settings in this workflow also improve the facial expressiveness of people from Krea 2. There's an explanation of how/why in the info section below. It's not perfect, but it's the best we can do until finetunes come out.
Otherwise, the sampler settings are geared towards sharpness and clarity - but you can introduce grain and other defects through prompting or with loras. It also does anime / digital artwork / whatever images well, but you may want to bypass the second sampler for that.
All the images attached to the post were generated directly with this workflow with no further editing.
Nodes & Models
Custom Nodes:
RES4LYF - A very popular set of samplers & schedulers, and some very helpful nodes. These are needed to get the best outputs, IMO.
RGTHREE - ( Recommended ) A popular set of helper nodes. If you don't want this you can just delete the seed generator and lora power loader nodes, then use the default comfy nodes instead. RES4LYF comes with seed generator & lora nodes as well, I just like RGTHREE's more.
ComfyUI GGUF - ( Optional ) Lets you load GGUF models, which for some reason ComfyUI still can't do natively. Once installed, you use the "Unet Loader (GGUF)" node to load the model. If you're not using any GGUF models you can just skip this.
Required Models:
Important Note: If you can, you should use the Int8 Convrot version of the model (unless you want higher quality using BF16). The Int8 Convrot model is almost 2x as fast to gen with, and is the same quality as FP8. Massive free speed boost. You will need to update your ComfyUI, support was only added early July 2026. You will also need an NVIDIA GPU, and CUDA version 130 or higher.
Main model: Krea2 RAW B16 / FP8 / Int8 Convrot | or | Krea2 RAW GGUFs - It is strongly recommended that you use the RAW main model with the turbo lora at 0.6 strength instead of the Turbo main model when making photo-real images. It gives WAY better results, and the only downside is that it takes a bit longer to gen. Gen times are already pretty short, so that's not a big deal.
The main workflow assumes you're using the RAW model with the turbo lora, and the settings will be very bad if you use the turbo main model instead.
Even the 'full_turbo' workflow still uses the raw model, seeing as you can just set the turbo lora to 1.0 strength and then it does pretty much the same thing as the turbo main model.
Turbo Lora: Rank 64 Turbo Lora - Using this with the RAW model at ~0.6 strength is better than using the Turbo model. The only real downside is speed. Even then, if you're in a hurry you still have the option of upping the strength to 1.0, which makes it just like the turbo model. Gen times are only 50% longer when it's at 0.6, so it's not really worth it to use full turbo IMO.
Anti-Censorship Lora: 2 Vector Bypass Lora - You should use this even if you're doing SFW stuff. More detail is below, but essentially this will massively improve prompt adherence, facial expressiveness, character detail, and numerous other things. There is no downside as long as your sampler settings are good (which this workflow takes care of for you). Do not use other bypass loras , they go too far or cause degradation of quality; this is the only one that works properly.
Text Encoder: Qwen3 VL 4B - Use the BF16 one if you can. Some people say text encoder quality doesn't matter much & to use a lower sized one, but it does matter and it affects quality.
If you're using a GGUF text encoder for some reason, swap out the "Load CLIP" node for a "ClipLoader (GGUF)" node.
VAE: Wan 2.1 FP32 VAE - This gives you sharper, clearer images than when using the Qwen Image VAE. There is no downside. It works because the Wan & Qwen Image VAEs are almost identical, and the FP32 precision improves the quality.
There is an alternative VAE you can use that's even sharper, but it has drawbacks so I've detailed it in the info section further down.
--
That's the end of the required info, you can stop here if you just want to use the workflow!
--
Info & Tips
Alternative Sharpening VAE
The Wan 2.1 Upscale2x VAE gives you even sharper images than the Wan FP32 VAE (it's VERY noticeable), but it sometimes introduces extra artifacts into the image and it also amplifies existing ones. It's up to you whether you think it's worth it or not, I personally think it's good for some images and bad for others, so I just output both and pick whichever turns out best.
Here's an example image using the normal Wan FP32 VAE: https://ibb.co/fGtZwdW8
And now the same image using the upscale2x VAE: https://ibb.co/RTj2DjVw
It's not in the workflow by default. To use it, you need to grab the ComfyUI VAE Utils node set and use the "VAE Decode (VAE Utils)" node instead of the regular VAE decoder. Then you also need to downscale your image by 50%, because this VAE decodes the image at 2x resolution (which is why it's so sharp). This pic shows what the setup should look like: https://ibb.co/XcrXmpr
What About Non-Realistic Images?
I still recommend using the raw model with the turbo lora at 0.6 strength for this. This is because the raw model is much more creative than the turbo model; you'll get better variety this way.
However, the second sampler is now optional because you may not need the extra detailing step anymore - you can just bypass it and it'll work fine. You can also change the scheduler in the first ksampler to sgm_uniform if you want an alternative look, but it's up to you. Just don't forget to change it back to beta if you're doing realism again ;)
Full Turbo Workflow?
You'll lose the creativity of the raw model by using it, but that may not matter to you at all depending on what you're doing. Or maybe you just need the speed.
As mentioned earlier, the full turbo workflow is set up for making non-realistic images, like anime / concept art / digital paintings. It only has one sampler because you don't need an additional detailing step, and you don't really need the benefits of a high-noise schedule either.
Euler/sgm_uniform is my general go-to for non-realistic images, and it holds up pretty well for Krea 2. I haven't tested it extensively though so don't take my word that it's the best sampler/scheduler or anything.
Otherwise, the only difference in the workflow is that the turbo lora is set to 1.0. You can also just use the turbo main model with the workflow and drop the turbo lora entirely, but then you're storing two main models for no real reason.
Krea 2's Facial Expression Problem: Censorship
This is the big one.
Basically, there's a lot of discussion going around about how Krea 2 doesn't do a very good job with facial expressions; characters lack expressiveness, and seem to have "dead eyes" a lot of the time. Smiles don't reach the eyes, that sort of thing. It's nearly impossible to make someone look angry, fierce, or anything more than mildly annoyed .
This is a very common problem with distilled models (i.e. turbo models), but in Krea's case it's mostly because of ridiculous censorship. The developers heavily censored Krea 2 against whatever content they arbitrarily decided was 'harmful', and in doing so they lobotomised their own model. It knows how to make an angry face, it just won't do it because it was collateral damage during the lobotomy.
You literally can't make people smile properly in most images with Krea 2 due to the censorship. That's not an exaggeration, try generating someone with a natural, realistic smile. Dumbest thing I've seen in years.
Luckily you can partially bypass the censorship using a simple lora, which you should use even if you're doing SFW stuff . It just makes better images, period. Some people say it also reduces the detail in the images, which is true - but this is actually just because you need to cook them a little longer. That is to say, if you have good sampler settings it's no problem. But it only works up to a point.
This workflow recommends using the bypass lora at 1.0 strength, but sometimes you need to go higher - even for SFW prompts - to get what you need. This isn't good because it degrades the image quality, but that's censorship for you. We'll need finetunes to properly decensor the model. This goes for SFW stuff too, remember - you will have a really hard time making a person look angry, even with the bypass on.
If you can't tell: I'm really annoyed about this and you should be too. The fact that you can't make someone look angry, happy, sad, etc completely ruins the model for a lot of applications. Literally unusable for so many things. All because they don't want your delicate little child brain to see blood or titties. Luckily, finetuners and lora makers will probably save the day
You can also use pornographic loras at low strength (~0.4) to increase prompt adherence even for SFW prompts. Yes, you heard that right: the censorship in this model is so stupid that you can get better SFW facial expressions and general model performance by using porn loras. No joke, I genuinely have porn loras on for most of my SFW generations.
Here's an example where I'm trying to get a strong, fierce expression from a sprinter using the words "She's frowning and snarling with effort" in the prompt.
Using the filter bypass at 1.0 strength it straight up refuses: https://ibb.co/WWV64GzM
It's better (still not good) with the filter bypass at 6.0 strength, but notice the image quality has suffered: https://ibb.co/mCnHq1mF
And... here it is with the filter bypass at 1.0 strength and PORNOGRAPHIC LORAS enabled at ~0.5 strength: https://ibb.co/2YCV1j9Z
Notice that the quality of the one with porn loras hasn't degraded at all, while also adhering to the fierce expression prompt better. I had to cherry pick 10 gens each just to get the first and second pics (which didn't even do a good job), but the porn lora one I only needed 3 gens - and all three of them were usable.
If this isn't a great example of why censorship is stupid then I don't know what is. This model would be god-tier if it wasn't intentionally broken by the devs. We can only hope that finetuned checkpoints can bring back what it lost.
Another area of improvement; it turns out that the model gives slightly better facial expressiveness in the earlier high-noise stages of generation - which means faces are more expressive when images are undercooked. But undercooking your images isn't good of course, so you need to finish cooking them one way or another. This is where a dual sampler set up comes in handy. More on that below.
Lastly, the raw model with the turbo lora at 0.6 strength is a bit better at facial expressions too. All of these tips combined are very helpful, but you'll still struggle with very intense facial expressions for the foreseeable future. Still, at least we can make people smile now (you can't do that with the censorship).
The 2 Vector Bypass Lora
This lora bypasses the censorship in the model, and is superior in every way - even for SFW images. It does reduce the detail of the image, but you can get it back by using noisier sampler settings, and your images will ultimately look better . I recommend using a strength of 1.0 at all times. If you need more censorship unlocks, use more loras instead of increasing the strength of this one.
It works by amplifying two specific vectors during generation (hence the name). This lora is the minimum you need to bypass the censorship, and therefore it's the best one . All the other ones change more stuff than they need to or are way too strong, do not use them. Don't even use the 3 vector one by the same author, just use the 2 vector…
