32 comments

[ 3.0 ms ] story [ 48.4 ms ] thread
Very good, not always perfect with text or with following exactly the prompt, but 6B so... impressive.
Did anyone test it on 5090? I saw some 30xx reports and it seemed very fast
We've come a long way with these image models, and the things you can do with paltry 6B are super impressive. The community has adopted this model wholesale, and left Flux(2) by the way side. It helps that Z-Image isn't censored, whereas BFL (makers of Flux 2) dedicated like a fith of their press release talking about how "safe" (read: censored and lobotomized) their model is.
It‘s mainly due to system requirements that Flux.2-dev doesn’t get same usage as Z-Image. A 5090 needs about a minute to generate an image with a basic workflow with Flux.2-dev. But prompt adherence and scene/character consistency in edit mode is (way) ahead of Qwen-Edit-2509 if you ask me.
Z-Image seems to be the first successor to Stable Diffusion 1.5 that delivers better quality, capability, and extensibility across the board in an open model that can feasibly run locally. Excitement is high and an ecosystem is forming fast.
I’ve paired my Z-Image Turbo with SeedVR2 upscale, running on a RTX3060 12gb, 32gb sysMEM, generates in 40sec. I’m holding out for Z-Image Edit that is a larger model, once that is out… going to be interesting. Oh and to train your own ZIT LoRA, takes 5hrs for 3000 steps. So fast.
I've done some preliminary testing with Z-Image Turbo in the past week.

Thoughts

- It's fast (~3 seconds on my RTX 4090)

- Surprisingly capable of maintaining image integrity even at high resolutions (1536x1024, sometimes 2048x2048)

- The adherence is impressive for a 6B parameter model

Some tests (2 / 4 passed):

https://imgpb.com/exMoQ

Personally I find it works better as a refiner model downstream of Qwen-Image 20b which has significantly better prompt understanding but has an unnatural "smoothness" to its generated images.

One thing I noticed is that you tested it with very short prompts; Z-Image Turbo really likes long prompts, and recommends using an LLM as a prompt enhancer, even providing a prompt template. [0] I have had pretty good look using an English translation that was posted on Reddit [1] with Qwen3-4B-Instruct locally sometimes modified somewhat for particular tasks; it seems biased to adding text to some images as-is) with short prompts.

[0] https://huggingface.co/Tongyi-MAI/Z-Image-Turbo/discussions/...

[1] https://www.reddit.com/r/StableDiffusion/comments/1p87xcd/zi...

Just want to learn - who actually needs or buys up generated images?
It's amazing how much knowledge about the world fits into 16 GiB of the distilled model.
i have been testing this on my Framework Desktop. ComfyUI generally causes an amdgpu kernel fault after about 40 steps (across multiple prompts), so i spent a few hours building a workaround here https://github.com/comfyanonymous/ComfyUI/pull/11143

overall it's fun and impressive. decent results using LoRA. you can achieve good looking results with as few as 8 inference steps, which takes 15-20 seconds on a Strix Halo. i also created a llama.cpp inherence custom node for prompt enhancement which has been helping with overall output quality.

We have vLLM for running text LLMs in production. What is the equivalent for this model?
I'm particularly impressed by the fact that they seem to aim for photorealism rather than the semi-realistic AI-look that is common in many text-to-image models.
What kind of rig is required to run this?
My issue with this model is it keeps producing Chinese people and Chinese text. I have to very specifically go out of my way to say what kind of race they are.

If I say “A man”, it’s fine. A black man, no problem. It’s when I add context and instructions is just seems to want to go with some Chinese man. Which is fine, but I would like to see more variety of people it’s trained on to create more diverse images. For non-people it’s amazingly good.

It would be more useful to have some standards on what one could expect in terms of hardware requirements and expected performance.
I've messed with this a bit and the distill is incredibly overbaked. Curious to see the capabilities of the full model but I suspect even the base model is quite collapsed.
Dude, please give money to artists instead of using genAI
Unfortunately, another China censored model. Simply ask it to generate "Tank Man" or "Lady Liberty Hong Kong" and the model return a blackboard with text saying "Maybe Not Safe".
This is an issue with your provider. You need to download the model.

It generates an image of a tank and the statue of liberty for those prompts.

We talked about this model in some depth on the last Pretrained episode: https://youtu.be/5weFerGhO84?si=Eh_92_9PPKyiTU_h&t=1743

Some interesting takeaways imo:

- Uses existing model backbones for text encoding & semantic tokens (why reinvent the wheel if you don't need to?)

- Trains on a whole lot of synthetic captions of different lengths, ostensibly generated using some existing vision LLM

- Solid text generation support is facilitated by training on all OCR'd text from the ground truth image. This seems to match how Nano Banana Pro got so good as well; I've seen its thinking tokens sketch out exactly what text to say in the image before it renders.

All the examples I tried were garbage. Looked decent -- no horrors -- but didn't do the job.

Anything with "most cultures" were manga-influenced comic strips with kanji. Useless.

>manga-influenced comic strips with kanji. Useless.

Are you sure it was Japanese? Because the model is Chinese so it's likely to output Chinese (it happened in my testing).

I‘m wondering: Is it faster or slower when spread across two GPUs (RTX3090)?
This model is awesome. I am building an infinite CYOA game and this was a drop-in replacement for my scene image generation. Faster, cheaper, and higher quality than what I was using before!