Z-Image-Turbo in ComfyUI! (Patreon)

Video - https://youtu.be/0AHIltRW1Ak

Overview

Z-Image-Turbo is, as the name suggests, very fast! Its also very good at following the prompt and generating images in a wide variety of styles.

Lots of different options available including:

Text-2-Image

Image-2-Image

Image Inpainting

HiRes Fix

Split Sigmas

Prompt Enhancer

You can use many of the standard schedulers and samplers, and it can generate images up to 2048x2048!

Some styles aren't quite as a good as others, and it is doesn't know quite a few common creatures and characters, but given the size and speed I'm not complaining!

It is also FOSS, making it one of the best models out there at the moment.

Workflows

A whole bunch of workflows for you to play with, and while I suggest NOT using Nodes 2.0 for the time being, the workflows with "Nodes2" in the title use fewer custom nodes and don't include the "Get/Set" options. This is because in Nodes2 mode, you can't actually pick the get options making it a bit annoying to use.

The Nodes2 versions also EXCLUDE the prompt enhancer too, so if you're looking for a more simple experience, then those are a good choice.

The most basic one is "ZImage Turbo Nodes2 v1". This excludes advanced features such as HiRes fix, Image compare and the prompt enhancer, leaving you with the simplest possible Z-Image experience. It is OK to use in "Nodes2" mode, but there are still various bugs such as menus "popping under" other nodes.

The "full-on" experience is "Z Image Get Set Upscale". This includes HiRes Fix, Get/Set nodes for fewer noodles, the prompt enhancer and image compare options... however, it is best used WITHOUT Nodes2 enabled.

Model links are available in-worfklow:

It uses the Flux VAE, so chances are you may have that already.

Torch Compile can be used to make generations even faster:

If you've opted to NOT install triton, then simple delete/mute/bypass this node.

Due to Nodes 2.0, the fast group bypass options are now split out individually:

This is because in Nodes2 mode, anything with more than 1 option just gets compressed into an unusable mess.

Nodes 2 also seems to like RANDOM numbers more than anything, hence I've included an integer constant node. If you want to use RANDOM numbers again, simply disconnect it like in the image below:

The HiRes fix settings you can play with and the actual values that work best will depend a lot on the chosen scheduler. The note suggests some decent ranges, which also apply to the Image-to-Image option as well:

The prompt enhancer uses their text by default, but depending on the LLM used you may wish to alter the system prompt to suggest it also looks at the input image (if you're using one)

This way you can sort of get "Image Edit" capabilities as it can "see" the image when using a vision model.

Have fun!