Master Z-Image-Turbo in ComfyUI: Speed & Quality Guide

Z-Image-Turbo is pretty quick when it comes to image generation. I like to use the rodent method for workflows. This means putting things neatly into different colorful little groups because I think it makes workflows both easier to follow and to update.

Z-Image-Turbo in ComfyUI Nodes 2.0

You might have noticed something called Nodes 2.0, which changes the look and feel of all of these nodes. At the moment, it's possible to do a quick toggle via the Comfy menu. As it's new and resistance is likely futile, I tried to make this first workflow as 2.0 friendly as possible. This was more difficult than expected given the number of bugs and missing features in Nodes 2.0.

Model and Sampler Setup

Here we have the model loader first with Z-Image-Turbo selected in a typical diffusion model loader. For even more speed, you can use torch compile. There is also a Model Sampling Aura Flow node to play around with. With a shift value of three, there's actually no difference to when the node is bypassed. I couldn't find a good use for it, but it's there if you want to explore.

For the clip loader, it's using Quen 34.B.

Display and UI Issues in Nodes 2.0

Both of these text nodes, the prompt and the note box, can randomly change size. Sometimes it happens, sometimes it doesn't. They can even vanish off screen. Switching between versions can cause layouts to change unexpectedly. If you want to change the color of your notes, that doesn't work at the moment either.

The fast group mute or bypass nodes get squished up in Nodes 2.0, which is why I added each of them individually instead. In the sampler group, the get and set nodes from KJ nodes can't really be used in 2.0, so I replaced them with explicit links. As Z-Image uses CFG1, there's a basic guider in there with no negative prompt at all.

Seeds, Steps, and Limits

If you'd like to use a fixed seed whilst testing, Nodes 2.0 doesn't have any of that. Control after generate is missing, so you can't change it to fixed. That's why I added a node from ComfyUI KJ nodes that gives you the ability to actually have a fixed seed.

As Z-Image-Turbo is fast, you can get away with as few as five steps, but it is happy going higher if you want to do 10, 15, or 20. The max resolution is 2048x2048, so I wouldn't try going much above that to start with.

It's also worth pointing out that if you use any custom nodes at all, you'll get a warning triangle about unsupported nodes highlighted red. There are usually no nodes actually highlighted red. Typically, everything runs okay. You'll just have a few display issues like with the fast group bypasser and the get set nodes not letting you pick or set values.

The generated image in this case is great. We get an anime rodent holding a postcard, and the text is nice and clear. Depending on the settings you use, such as the sampler or scheduler, you can get slightly richer colors and more details, but I like to start with something close to the defaults.

Image-to-Image and Inpainting in Z-Image-Turbo in ComfyUI

You can't do image editing yet, as they haven't released that model. You can do the usual image to image by using an image latent.

Denoising Controls and Schedulers

In Nodes 2.0, the point key doesn't work when you try to select a denoising value. You'll be stuck using the plus and minus keys, which is slow and annoying. I added a note with some information about decent denoising ranges, which vary depending on the scheduler. DDM, for example, requires much higher values than linear quadratic.

Image Loader Slowness

The new image loader is slow. You can still pick an image you've already got somewhere, but it isn't as fast as it was before and could leave you confused. It sometimes looks like nothing is happening. The more times you click on it, the slower it gets.

Nodes 2.0 issues aside, the resulting generation from Z-Image is pretty decent, managing to create this from an utterly unrelated input image. She has a nice hat, is in the same position as the person, and some areas have become more anime and forestlike, even if the overall image style is still photographic.

Inpainting Example

If you want to change a specific area of an image, the inpainting group shows an example of that. I masked only her face, so it limits the changes to that area. This is managed by the Set Latent Noise Mask node, which is the only addition. Everything else in the image stays the same, but it changes her little rodent face.

You can vary the denoising value here. If you go all the way up to one, it will change a lot. If you lower that slightly, it will keep some of the original features of her face.

Style Tests with Z-Image-Turbo in ComfyUI

Painting and Classic Art Styles

First up is a prompt for a painting style. A fairly simple one with a cat and a dog in a matte oil painting style at 1280x1280. The result is very good. The cat is on the shelf in the background and the dog in the foreground, so it followed the prompt request well. There are small details that shift it from a full photographic style, such as the lightness around the edges and what almost look like brush strokes.

Given that it did so well, I tried a much shorter prompt. Lady with an hermine is a famous painting, but can it take the essence of that, add a little fire and some moonlight, and change the style to that of a claymation figurine? The answer is yes. The lady seems to have a tail, so it isn't perfect, but that can be changed easily by generating again with a new seed, and that only takes a couple of seconds.

I tried the same prompt with a small change at the end to give a pixel art style. Once again, she has a bushy tail. Or it could be a simple mistake. Although AI never makes mistakes, does it?

Pencil, Charcoal, and Low Poly

Some models don't handle colorful pencil sketches very well, but Z-Image does. I like the way it added the white border as if it's been done on paper. There are plenty of lines to indicate a pencil art style, and this time she appears to be tail free.

Charcoal sketch often comes out far too clean. Here it has a slightly smudged look as you'd expect with charcoal. Perhaps a little too much detail, but it has that dirty gritty feel.

Low poly 3D render looks for a blocky result, which is what it delivers. She is tail free once again and has a comfy thing to sit on whilst keeping her warm.

Cubism Trials

This is where most image generators fail. Cubism. Z-Image falls at this hurdle. Far too much detail for my liking, although it has shifted toward a painting style.

I tried KL optimal, which still isn't very cubist. It does have more painting-like qualities with brush strokes and a more damaged look. I suggest playing around with those settings for yourself.

I tried a different sampler with even more steps going up to 15. Still not very cubist, but tending more that way. The light around the moon becomes more blocky along with the rest of the night sky.

Woodblock at High Resolution

I changed the prompt to a woodblock style rodent rogue purchasing his supplies at an ancient market stall and bumped the resolution up to 2048x2048. The result is very impressive. There are a great many pouches, ideal for keeping seeds in, and the image looks great overall. Ready to go adventuring.

Long Prompts, Multiple Characters, and High-Res Fix

My rodent mage is looking to purchase some seeds on board a spaceship being served by a blue-skinned floppy-eared alien. With DDM at 10 steps for clarity, that seems to have done the trick. They each have their correct outfits. There are some delicious-looking seeds and even a nebula outside the spaceship portal window.

With DPM++ 2S and linear quadratic at 12 steps, it's also decent. She has double ears and some slight blurring around her face, but it's still an impressive result. Back to the default with eight steps shows similar outcomes.

High-Res Fix Workflow

Time to add a high-res fix. I start with a generation passed to a high-res fix group along with a latent upscale of 1.5. Using six steps with a denoising of 0.46 adds a whole lot more detail. Various features change, but you can adjust the denoising value to taste. You can really see the change on skin and facial details and the nebula outside, and the rodent fur gets a nice boost.

Split Sigma and Text Fixes

With a classic split sigma workflow at default settings, there is an issue with the Cheese Explorer text. She has only one set of floppy ears this time. The high-res fix adds a load more detail and even fixes the Cheese Explorer text. Lots of nice fur and robe texture detail in there too.

Prompt Enhancement in Z-Image-Turbo in ComfyUI

I added a prompt enhance feature using an English version of the prompt that they provide. If you're looking for their prompt enhancing template, it's available on their Hugging Face page in that bit of Python. The original is in Chinese, so I used Google Translate to turn it into English. It's using a local O Lama powered LLM Gemma 34B. That allows you to be even more creative with your prompting.

A novel image style of forms without forms, nonlinear shapeless shapes crystallized to create a wonderful scene knitted on the clothing of a rodent sorceress sitting on her throne in darkness. The preview shows the actual prompt, so we should expect to see a rodent sorceress seated on a throne of stone wearing a dark gray tunic with thorny vines on it and a sign reading Veritas, and the overall scene dominated by crystalline shapes.

The initial image is okay, but the sign and her facial features are a little unclear. The high-res fix looks much better. The crystals look amazing. The sign gets fixed and she even has a hood on those robes along with a much clearer face.

Tarot Card Example

Picture your dream question in your mind, because the large language model will do the generation for you with three cards atop a purple velvet cloth. The reading gives three cards. The text is nice and clear already, and the high-res fix adds nice details to each of the images. The Hermit gets some cool robes. It changes Fortune to Fortuna, but that can be changed with denoising.

Where Z-Image-Turbo in ComfyUI Struggles

Mythological creatures is one place it falls short. The cyclops test often trips models up. The first generation has a great-looking face, and the high-res version adds more details, but he still has two eyes, which is very un-cyclops-like.

Trying a satyr, a goatlike creature often playing the pipes of Pan, Z-Image gives something that isn't very mythological at all. It's a good-looking image, and the high-res fix adds lots of nice detail. It just isn't what I was looking for.

I asked for a random creature and the LLM chose a griffin, a legendary creature with the body, tail, and back legs of a lion, the head and wings of an eagle. The first result isn't bad. It's a decent painting style with nice wings. The high-res fix adds great feathers, but the face becomes that of a lion instead of an eagle. Less griffin-like, even if the overall image aesthetic seems a little better. Still, we can't have everything all at once.

Final Thoughts

Z-Image-Turbo in ComfyUI is fast, flexible with steps and resolution up to 2048x2048, and strong across many styles with clear text and solid prompt following. Nodes 2.0 brings useful changes but also bugs, missing seed controls, flaky UI behavior, and slow loaders that you may need to work around with KJ nodes and careful wiring. Image-to-image and inpainting give good localized edits, high-res fix meaningfully improves detail and text, and prompt enhancement can boost creativity. It stumbles on certain mythological specifics and cubism, but overall it delivers impressive results with short prompts and quick iterations.