@iamstreamin
Hi! I'm glad you like the workflow. However, it's difficult to help you without seeing your entire setup.

Regarding point 1: The plastic skin look comes from Z-Image itself, specifically from the generated RAW image. The SDXL pre-definer should easily fix this issue. I actually don't have any problems with the plastic skin look. It looks very natural in my images. You have several options to counteract this. Search for "Realistic Skin" LoRa here on CivitAI. You can use one for Z-Image itself and load it directly into the first LoRa loader. Alternatively, you can download an SDXL "Realistic Skin" LoRa. SDXL/Pony etc. offers a wider selection and it makes more sense to edit the skin details within the Pre-Definer. Then add this to the LoRa loader of the pre-definer. Start with a strength of 0.55-0.7 and test it. The third, and in my opinion, best option is to first test a different SDXL model. Thats the checkpoint that is loaded by the pre-definer. The results vary greatly depending on the model! This has the most impact.

Also check your prompts. Use something like "natural skin texture, subtle pores" etc. in your main prompt and also within the pre-definer prompt.

Point 2: If the face appearance of your RAW image changes and you want to keep it stable, first check where the change occurs! That's precisely why there are different Image Compare nodes inside the workflow. This gives you full control and allows you to track down where an unwanted change is happening. I suspect this is coming from the PreDefiner. Check this using the smaller Compare node to the right. If it's happening there, lower the Denoise value in the PreDefiner or try disabling (bypass) it temporarily to see if the change stops. You can also experiment with the CFG value. You should lower that as well.

If you have any further questions, feel free to contact me via DM here on CivitAI or directly on Discord. The small comment sections here aren't designed for extensive support like this. On Discord, you could send me images, which would make it much easier for me to help you. You're welcome. 🙂

@koalabearr Hello Koala, thanks for your message. It's difficult to help you without seeing your entire setup.

If you haven't changed any of my basic settings, then it's basically impossible for you to get only pixels as output. Therefore, I suggest you reload the workflow into ComfyUI, select your model, the VAE, and the clip encoder, insert your prompt, and test again.

@xiaobai_solon Hi, thanks for your feedback. I really appreciate it. 🙂

1. The SDXL pre-definer in v5 was primarily added to fix the typical NSFW issues with ZIT. It was requested by several users who wrote me. I've already received a lot of positive feedback about it. The downside, of course, is that the pre-definer also alters the face itself, not just the parts you want. If you want to keep the face stable from the RAW image, you'll need to bypass the SDXL pre-definer.

You can also try lowering the denoise and CFG values, for example, denoise to 0.10 and CFG to 3.5, but this won't completely eliminate the problem. If you're using a ZIT character Lora and it's essential for you that the face remains stable, you should simply bypass the node as well as the face and eye detailers. Or, insert a ZIT face detailer with your Lora between the last detailer and the upscaler. That would restore the face. Alternatively, you could mask the face so that it isn't altered by the pre-definer.

Or you could simply use v4 or v5 Pro with Z-Image Detailers instead of v5 Pro with SDXL Detailers.

2. No, I don't currently plan to release an anime/2D style version of this workflow. Iam more specialized in photorealism. But what exactly would be the difference? You can generate anime images just as well with this workflow.

3. A ControlNet version of this workflow will be released at the beginning of the new year.

4. Interesting idea with the WD14 tagger. I'll try it out as soon as I have time for that. If it doesn't slow down the workflow too much and it works well, I might include it.

Thanks again for your feedback. 👌

@rlewisfr346 Thanks. Glad you like it. 🙂

1. Yes, the detailers are doing great. But the results vary widly depending on the setup. My stock settings are always a very good starting point, but you'll need to experiment and test the values. Try different SDXL models for the pre-definer. Experiment with scheduler/sampler, cfg, denoise, steps, lora strength, etc.

You can also try a different ZIT VAE file. 👉🏼 UltraFlux-v1 VAE file In many cases, this yields better results and more details for your ZIT raw image. The same with the ZIT model. Try out the RedZimage AIO model model and compare the results. Iam using it quiet often these days. You can also use "realistic skin" LoRa's and put them in the LoRa loader of the pre-definer.

2. If you wanna use ZIT char LoRa's you really need to crank down the denoise of the pre-definer or disable it. Because with denoise set to 0.20 it will change the face. I've don't uploaded a v5 version with ZIT pre-definer. All versions are using SDXL as pre-definer model. But feel free to try out the v4 version. The v4 is working with ZIT as pre-definer model. 👍🏼
Alternatively, there's the v5 Ultimate Edition. This includes Lora Loaders for all six detailers. With the ZIT version, you could load the character LoRa into the face detailer so that the facial features are restored after the pre-definer.

Thanks again for your feedback. Wish ya a happy new year! 🥂

@collector82 Hi, thanks for your kind words. I'm glad you like the different versions of the workflow.

The ControlNet version will be released in January. However, I'm not a fan of releasing things online that aren't 100% functional. That means I'd rather test for an extra day or two than risk any complaints. Quality over quantity. I'll also continue working on the manual. There's still room for improvement. Better instructions automatically mean fewer questions and more success for users.

Feel free to contact me anytime here on CivitAI, on Ko-fi or on Discord.

I also wish you a happy and healthy new year. 🙂

@ayushwesker Hey, I'm glad you like the workflow.

1. The main reason for the SDXL pre-definer is the NSFW fix. This was requested by several users and implemented in v5. Check example pictures to see what it's doing. It also provides a certain diversity in the faces and results in better skin detail. This is also the disadvantage when using a ZIT char LoRa. In this case, simply disable the pre-definer or just use v4. v4 works with ZIT as pre-definer. Or use the v5 Basic version. This one is without a pre-definer and will not change the face.

2. No, there isn't a detailer just for skin. At least not that I know of. But you can use "realistic skin" LoRa for ZIT and SDXL. Both are compatible with this workflow. However, you'll need to experiment a bit with settings like LoRa strength, denoise etc. to get the best results.

The v5 Ultimate Edition is offering more options. E.g. 6 detailers instead of only 3 and separate LoRa loader for every detailer. See preview screenshot for that and read the changelog provided in the workflow description. For example, you could insert the character lora into the face detailer to recreate the face after the pre-definer. There are truly many different possibilities with the various versions of this workflow. You just need to take the time to try them out. It's not a one-click wonder.
If you have any further questions, feel free to ask. 🙂

@svenson2020648 Hey, thanks. I'm glad you like the workflow.

And you're right. Sometimes the teeth might not turn out perfectly. It's the same with the fingers and toes in Z-Image. This is due to the model, not the workflow, but it's very easy to fix.

1. This should absolutely not occur with portrait or half-body shots. Why? Because there are enough pixels available for the mouth area so that the Pre-Definer or Face Detailer can correct it. This means the problem lies in too few pixels. The first fix is ​​to increase the resolution. This will automatically lead to a better result.

2. The second solution is even simpler. We ensure that Z-Image creates the raw image flawlessly from the start. Since we don't use a negative prompt here at ZIT, we tell it what we want in the positive prompt. Simply write something like "perfect white teeth", "flawless white teeth", "sparkling clean smile" or "bright polished dentals" etc. in your main prompt. Phrases like these will automatically ensure that teeths always look perfect and natural. Try it out.

I will include this point in the troubleshooting section of the manual.pdf.

If you have any further questions let me know. 🙂

@svenson2020648 No problem, I'm glad it worked right away. Have fun with the workflow!

@PatrickChen Hi Patrick, glad you like it. But why you don't read some of the last comments here? I've already answered this question two times. Your question is also answered in the v5_Ultimate_Manual.pdf file you received with your download. There is a troubleshooting section inside this ebook. 😉

Short answer: You have two options here. 1. Just disable the pre-definer. There is an toggle on/off switch for that. 2. Use the ZIT Detailer version and put in your lora into the lora loader of the face detailer. This should bring back the face after the pre-definer.👍🏼

@Seikkailijapoika Interesting suggestion. You are the first one that is asking for a penis detailer update. 😉
I personally don't need it, but if there's a real need for it, then I'll add it to the feature list for v6. 👍🏼

@Seikkailijapoika Cause some more ppl asked for it, this feature is integrated in v6 now. Release in some days.

@emmasteadman This workflow (with all its features) is running without any problems on 8gb GPU's. So i suggest to use a smaller SeedVR2 model and read the workflow description.

But if you say its just freezing then there is something wrong with your ComfyUI setup. It should give an OOM (out of memory) error. So check your ComfyUI setup, update it to latest version including the custom nodes and the frontend package.

@emmasteadman Nope. No problems so far.
But maybe you just think that its frozen but its still downloading the model file? Check console window for that!

@mendoras231 When some nodes are still red they are not installed. After installing you should completely close your ComfyUI and reload it. Not just using the restart function within the Manager.

@mendoras231 Then there is really something wrong with your ComfyUI installation. You should start from scratch with a fresh install.

@Daaraen There is no need to google anything. All custom nodes are listed (and linked) in the workflow description. You can easy install them via ComfyUI Manager. You dont need to do any manual installs or typing any commands in the terminal window.

@TheRealSmartie Thanks. And don't worry, i don't forget it. But my to-do-list is very long... 🫣

@elysiumfinest That depends on why it's displayed in red. Either you're missing the node or the model file. Check your console log.

@bosstne821 Hi, thanks for your comment. Did you read the manual that came with the download? If not, take a look and scroll down to "troubleshooting."

And yes, of course the pre-definer changes the face. That's one of its purposes. It's meant to introduce more face variations so we don't always have to work with the same boring ZIT faces. However, the main reason for the SDXL pre-definer is that it fixes typical NSFW issues and adds more detail for example the skin. And it does that really well. There's no function where we can say "please only change the background" and that wouldn't make any sense.
You can lower the strength (e.g. Denoise to 0.10), but that will still change the face a bit. Or you can try to use the positive and negative prompts in the SDXL pre-definer for that. Try adding "face change" or similar to the negative prompts. But i never tried that out. The best option is to disable it if you want to keep the face from the ZIT RAW image. Or you can work with a character LoRa and feed it into the Face Detailer after the pre-definer. In that case, you could leave the pre-definer enabled.

But please explain why you need the pre-definer for "background details"!? I can't really picture what that means. Especially since ZIT is better than SDXL or PONY in terms of background details.

@bosstne821 20 generations only for testing? That is nearly nothing. I made thousands of test generations before I released v5. 😉

You need to take the necessary time to get a feel for how to prompt correctly and use the workflow. Twenty attempts won't be enough. Using Grok is working fine for ZIT prompting. Gemini is also working great. But you need to tell them at the start of the fresh session that you need "natural language prompts for z-image model without negative prompt and at least 100-200 words" then you will get the right prompts. Don't use ChatGPT for ZIT prompting. It will not work that well. AI generations are always try and error. This workflow is very powerful but you need to learn. After some hundreds of generations you will get a feeling for it.

You can also try out some LoRa's to get a sharp background. e.g. Amateur photography LoRa's and load them into the first LoRa loader for the base ZIT image. Should work.

And of course you can use both. ZIT and also SDXL/PONY character LoRa's. That's the reason why you got 2 different versions. 👍🏼

Thanks! Glad you like it. Feel free to share some of your best creations in the gallery below. 🙂

@ionelbolintineanu707 There is no problem with the upscaler. And this error message has nothing to do with these workflows. It's a problem with your python and cuda setup. Search on google for that error message. There you can read how to fix that.

@ionelbolintineanu707 No worries! Just a small tip for the future: When working with ComfyUI and complex AI workflows, it's usually better to stick to stable releases rather than 'bleeding edge' versions. Most nodes and dependencies are optimized for Python 3.10/3.11/3.12 and CUDA 12.1/12.4/12.8. Using version 13.0 is very experimental right now and often leads to the library errors you're seeing. Switching to a stable environment will save you a lot of headaches! Also with many other workflows. You should start with a clean setup and the right versions and all will work fine! Best of luck getting it sorted.

@ionelbolintineanu707 I totally understand. Custom node dependencies can be a real pain!

Here is a tip to still enjoy the workflows with the upscaler: Just set up a second, clean environment. You can simply download the portable ComfyUI version and put it into a separate folder. Keep that one on a stable stack (like Python 3.12 and CUDA 12.8). This way, you can keep your 'experimental' setup for your specific nodes, but use the stable one for high-quality rendering and upscaling without any errors. It's much easier than trying to fix broken library links in a 'bleeding edge' setup.

Thanks for the kind words, and I'm glad you're enjoying the work despite the technical hurdles!

@ionelbolintineanu707 Glad it's working now. Have fun with it and feel free to share some of your creations on the workflow gallery below. 🙂

@elysiumfinest Thanks. glad you like it. 🙂
Yes, I have exactly that. An upscaler based on this workflow – it's actually my daily driver upscaler. I’m planning to release it soon, along with several other workflows I’ve been working on. Stay tuned! 👌

@leviathan814373 Hey! Thank you so much for the feedback! I’m glad you’re enjoying the Pro version results! 👍🏼

If you like the depth and control here, you’ll definitely love the Ultimate Edition. It takes the whole logic and refinement to a higher level and includes 3 more detailers. Keep up the great work with your renders! 🙂

@kevfactor859 You can use this workflow for Z-Image Base Model. There is nothing wrong with it as it uses the same Text Encoder and VAE. You simply need to download the new model and select it in the model loader. To see how the Base model performs, just increase the steps in the Sampler to 30-50 and set the CFG to 4-7.

Every file you need you can find in the workflow description. They are all listed there.

@DerDaAgropesca Hey! Thank you for the kind words. It’s great to hear that the 'Basic' version is providing such a smooth experience for you – that’s exactly why I built it: clean details and less frustration. 👍🏼

@replicantje230 It seems you did not read my post from yesterday with the title "✅ Z-Image Base Compatibility" in the comment section. 😉

@replicantje230 I wrote my thoughts in my comment yesterday. Z-Image Base is not meant to use it as a productive checkpoint for great photorealistic results. It's not trained like ZIT. It's for pro users who wanna re-train / refine models. Z-Image Turbo is still the way to go! Very fast and great output quality. In combination with my workflow and SDXL Definer and all the detailers you will always get high quality outputs. It doesnt matter if it's for SFW or NSFW. 🙂

If you wanna use the base version you need to wait for some fine tuned checkpoints. But i dont would waste my time with that, when it takes 4x or 5x the time generating an image with 30-50 steps. But time will tell.

@replicantje230 Everything is possible. But it makes no sense for NSFW because ZIT is censored. SDXL isn't. Thats the reaosn why the workflow works the other way around.

But i will release a img2img Upscaler + Detailer workflow in two different versions. There it will be possible.

@Airenn Why not download and test yourself instead of asking such questions? 😉

1. Of course v5 is the same great output quality than v4 because its based on v4. I also wrote that in the desription and changelog. But i've deleted the "best quality" on v4 now, because it can be a bit confusing. v5 is better overall because it offers more options and the same quality!

2. One version is working with SDXL for the detailers, while the other version is working with the ZIT model for that purpose.

3. Read the description: I've linked my recommendations there. But it doesn't matter what my favorite is. It really depends on the usecase! There are so many good SDXL/Pony/Illustrious checkpoints out there. Test them yourself and decide whats working best for you.

Hey Helliana, thanks for your kind words. Glad you like this workflow. 👌

v6 is coming soon... Feel free to post some of your best images in the gallery below.

Happy rendering! 🙂

@akgm Hey, glad you like it.

For your question: Just fix the seed on the k-sampler! Every fresh seed gives you a new image variant.

@akgm Yeah, it will take some days to get into, but when you know how it works it starts to get a lot fun! 🙂

@mampunez753 You really ask me if "this is normal"? Sure, I specifically designed it so you can take a nap between every generation. 😄
Just kidding. Fix your Drivers. 15min is pure CPU rendering. Or did you render in 20K?
These workflows take between 30 - 90 seconds on a real GPU setup.

And if you want actual support, first tell me which of the five different v5 versions you're even using and send me a screenshot of your settings or the image. But do it via Discord. This comment section is not a support channel, as I've mentioned multiple times before. It's not built for that.

@incubusinmyhead Thanks. glad you like it.
I've already created it. But I don't know yet when I'll publish it here.

@iamaritomix545 Hey, glad you like it.
To your question: Just read the other comments here. I already answered this question a lot of times. Short: You can just disable the pre-definer or go down with the denoise value for it. Or you use the ZIT detailer version and put in your ZIT character lora into the face detailer. But Lora Injection is only supported in the Ultimate Edition.

V6 will be released in some days. This new version will have a face lock feature for the pre-definer. So the face will not changed anymore. But only available in the Pro and Ultimate Edition. Not in the free Basic Version.

@iamaritomix545 v6 Pro Version will be $20 and Ultimate Edition $35.
If you grab the v5 Ultimate Edition now (only 25$) you will get the v6 as free update in a few days.

@iamaritomix545 Just read the workflow description. Everything you need to know is listed there. Or take a look at the changelog article:
https://civitai.com/articles/25914/
And i already answered your question. But again: 😁
You can just disable the pre-definer or go down with the denoise value for it. Or you use the ZIT detailer version and put in your ZIT character lora into the face detailer. But Lora Injection is only supported in the Ultimate Edition.

V6 will be released in some days. This new version will have a face lock feature for the pre-definer. So the face will not changed anymore. But only available in the Pro and Ultimate Edition. Not in the free Basic Version.

@iamaritomix545 The links are also in the WF description. But here you go:
https://ko-fi.com/s/fa20e6d018

@iamaritomix545 You can message me on discord or here on civitai direct message when its officially released. I still need to edit the ebook manual vor v6. So i think 2-3 days.👍🏼
If you dont care about the updated manual just message me on discord and i will send you the v6 files there.

@ztyian Hello! You can just disable the pre-definer with the one click toggle. Or if you are using a ZIT character lora, you can inject this one into the lora loader of the face detailer that is working after the pre-definer. So he will recover the face. But this feature is only available in the v5 ultimate edition.

In the next days there will come v6 with face lock feature for the pre-definer. 🙂👌