Sign In
homemodelsimagesvideospostsarticlesbountieschallengeseventsupdatesshop

Z Image Turbo (ZIT) Anime Lora Training Experience

39

CitronLegacy's Avatar

CitronLegacy

Jan 3, 2026

training guide
zitanimelorasz image turbo
Z Image Turbo (ZIT) Anime Lora Training Experience

Introduction

Happy New Year, everyone! I hope you’re ready for a great year.

Lets learn some "new tricks in 2026"! (That kind of rhymes... right? 😅)

I've trained over 100 ZIT Loras, mostly anime, so I wanted to share what I’ve learned about Z Image Turbo (ZIT).

TL;DR

  1. Yes, you can train ZIT Loras on tags instead of captions. No caveats. It works.

  2. Here’s a link to examples of my ZIT loras and the settings https://docs.google.com/spreadsheets/d/1-C3P6h83RuMhW5xCjtWRGVEfxzPrVtKLDzHZSTr0s3M/edit?usp=drive_link.

    1. All datasets use tags and are basically the same as my Illustrious datasets. The only thing I change is the number of image repeats.

    2. Look at the Status column. Note that I have documented loras that are failures. Mostly these failed due to small datasets or small training steps. (For example, BOTWStyle_ZIT failed by v2 was a success because had more steps).

    3. Here are all the config files I used for AI-Toolkit (https://github.com/ostris/ai-toolkit) to train these loras. Full transparency you have everything, except the actual datasets 😜

      Configs: https://github.com/citronlegacy/ai-toolkit-zimage-config-generator/tree/efc14a25cf65b81fb9676d3271718ce55ce0fa5d/example_configs

  3. You don't need a guide for training on ZIT, just train it the same way as Illustrious. You only need to change the the training configs/settings and possibly image repeats.

  4. Need help writing prompts for ZIT? Just use the prompt at the bottom of this article and have an LLM write the prompts for you!

Long version

Disclaimer

There is more than one way to skin a 🙀... hmmm... more than one way to train a CAT! 😸 I'm not telling you this is the only way to train ZIT. I'm telling you this is how I do it 🍋, and it works if you'd like to use this strategy.

I think it's a mistake when people tell you there is only one correct way to train a Lora, because those people are missing out on a chance to learn/improve. I've made over 3,000 Loras, but I still have more to learn/improve.

Please comment if you have tips or tricks for ZIT. We learn faster together! Team work makes the dream meme work!

Intro

Last year, Z Image Turbo (ZIT) was released, but there isn't any documentation on how to train loras? At least, there were no guides when I got started. 🤷‍♂️

In 2025, I made over 100 Loras (mostly anime), so I wanted to share my findings!

Findings

Z Image Turbo (ZIT) trounces Flux—lol, Rest in peace Flux. (Well I've heard its based on Flux (maybe?) so this is an invalid comparison 🤣)

I really struggled to make anything (anime or realistic) on Flux, but realistic things are a breeze on ZIT. Even a rushed or lazy dataset produced nice realistic ZIT Loras.

I'm not an expert on realism; my forte is anime or drawings, so that is my focus in this article.

Anime on ZIT

ZIT is good at generating very high-detail art.

ZIT understands natural language and knows a lot of "normal" and "non-anime" concepts, so if you write a nice paragraph, you can really get high-quality depth of action, scene, motion/emotion.

I think "natural language" is a funny term because the average person does not write the fancy-quality paragraphs that are needed to prompt ZIT/Flux/whatever.

How to write ZIT Prompts

If you are struggling with prompting ZIT, the solution is simple: ask AI!

I'll give you an example prompt that you can give to an LLM to have it write good prompts for you. (See bottom of article)

With this prompt in a chat session you can now either

  1. Describe the image you want using lazy language or even tags.

  2. Upload an image you want to copy and ask the LLM to give you a prompt to make that image but with whatever changes you want. (As of Jan 2026, you are able to upload at least 10 images per day to Grok. I use Grok way more than ChatGPT for writing prompts)

Conclusion

All this said, I'm not a ZIT evangelist knocking on your door to tell you about your new future. I love Illustrious, and I hope that Illustrious (or something similar) is always relevant. (Illustrious costs way less VRAM than ZIT so thats nice). I'm going to keep making Illustrious as my main focus.

ZIT is new, it's fancy, and there is no documentation on how to make ZIT Anime Loras --> until now! 🍋🎉

Prompts

As promised, here is a prompt you can use to have an LLM write a natural language prompt for you.

Please modify and upgrade this prompt to fit your needs. "Use it, break it, fix it, trash it, change it, mail - upgrade it. 🎵"


### Prompt Guide to Paste into a New Session
You are now an expert prompt engineer for Z-Image Turbo (the fast open-source image model from Tongyi Lab).

Key rules for crafting the best possible prompts for Z-Image Turbo:

1. Use long, detailed, coherent natural language descriptions that read like a narrative or story. Avoid short "tag soup" lists (comma-separated keywords without structure). The model responds much better to full sentences and paragraphs.

2. Structure the prompt in logical layers/order:
   - Start with the main subject, action, pose, and appearance.
   - Describe the environment, background, and setting.
   - Add atmosphere, lighting, mood, and camera angle.
   - End with style, artistic influences, and quality/technical boosters.

3. Seamlessly mix in tags and quality boosters, especially toward the end or integrated naturally. Useful boosters include:
   - photorealistic, ultra-detailed, 8k, masterpiece, highly detailed, sharp focus, intricate details, best quality, cinematic lighting, volumetric lighting, depth of field, etc.
   - Specific art styles: in the style of [artist], cyberpunk, studio ghibli, greg rutkowski, etc.

4. Do NOT use negative prompts — Z-Image Turbo does not support them.

5. Aim for 100–300 tokens (roughly 75–200 words). Longer, more descriptive prompts almost always yield better results.

6. Optional tricks that often improve results:
   - Start the prompt with a role or context (e.g., "A stunning photograph of..." or "An epic fantasy illustration of...").
   - End the entire prompt with a period (full stop).
   - If enhancing an existing prompt, you can think step-by-step first to refine it.

Example of a strong Z-Image Turbo prompt:

"A beautiful woman with long flowing silver hair stands confidently on a misty mountain peak at golden hour sunrise, wearing an elegant flowing white dress that catches the wind, dramatic volumetric god rays piercing through clouds behind her, intricate details on fabric and hair, fantasy atmosphere, in the style of Alphonse Mucha and Greg Rutkowski, photorealistic, ultra-detailed, 8k, masterpiece, sharp focus, cinematic lighting."

When I give you an idea or scene, always write 1–3 optimized prompts following these rules. If I ask for variations, create different versions (different angles, styles, lighting, etc.) while keeping the core structure.

39