Z-Image non-Turbo

[knoic]

Describe your use-case.

During training with comfyui/z-image base bf16, the images become corrupted and show snowflake-like patterns when sampling.

What would you like to see as a solution?

Normal Training and sampling

Have you considered alternatives? List them here.

Normal Training and sampling

[Garland677]

same here, need zimage base support

[hameerabbasi]

How many steps did you use? Z-Image base requires around 50 steps.

[Garland677]

How many steps did you use? Z-Image base requires around 50 steps.

yes, 50 steps generates exactly same as 16 steps, currupted.

[Calamdor]

int8 is not working, use bf16 or float8

[knoic]

int8 无法正常工作，请使用 bf16 或 float8。

Thanks, bf16 is running smoothly.

[niceguy4]

It seems like the finetune option works for training the base model with some adjustments (@Calamdor 's suggestion). I think an issue is with ComfyUI running the finetuned models. I have issues, too. All black images on the trained model but no issues with the base model in ComfyUI.

I can also generate images (inference) using the finetuned model using musubi-tuner.

Possibly a ComfyUI issue or my dependencies are borked.

edit: this was confirmed by the Comfyui team as of yesterday. Comfyui is unable to generate images using finetuned models from the base model.

[dxqb]

edit: this was confirmed by the Comfyui team as of yesterday. Comfyui is unable to generate images using finetuned models from the base model.

full-finetunes? or including LoRA?
please link the issue

[dxqb]

Currently known settings; will add a preset when things are more clear

• use the Z-Image Turbo preset
• change base model
• remove the transformer override (De-Turbo)
• change quantization settings; details here: https://github.com/Nerogar/OneTrainer/wiki/Quantization#artifacts
• you might want to experiment with timestep shifting, because Z-Image non-Turbo uses a higher shift during inference than Z-Image Turbo
• some people have commented that Z-Image non-Turbo requires a higher learning rate; but this might also be related to timestep shifting.

[niceguy4]

edit: this was confirmed by the Comfyui team as of yesterday. Comfyui is unable to generate images using finetuned models from the base model.

full-finetunes? or including LoRA? please link the issue

Yes, full finetunes on the Z-Image "base" model. I posted an image of ComfyUI's response. I can tag you in discord.

[kentrob-byte]

My character lora results improved significantly switching from adamw .0003 (the setting I'd been using with ZiT) to prodigy 1. I wasn't getting much likeness when using the lora with ZiT even if I trained out to 200 epochs, but now I'm getting a good likeness at 150.

Edit: Using adam_adv with a higher LR (0.001) was also viable, as I continue to experiment.

[qnxsgwy]

Please add the training preset for Z-Image non-Turbo as soon as possible. Thank you very much, I’m already eager to test it

[niceguy4]

Please add the training preset for Z-Image non-Turbo as soon as possible. Thank you very much, I’m already eager to test it

Just a FYI. You don't need to the preset to do a run. You can use the Turbo preset and modify it using dxqb notes above.

[ansinerd]

i used turbo preset and changed turbo model to base model but it fails to load the z image base.

[FennelFetish]

@ansinerd
I first downloaded the Z-Image repo manually like I did with Chroma, bit it couldn't load the model when I set a local path (something about "can't load single file format - try diffusers").
I had to set the base model to Tongyi-MAI/Z-Image and let OneTrainer download and cache it.

Had very good results with a character LoRA rank 16, alpha 1, full layer preset.
LR 1e-4, constant schedule, Adafactor, bf16, no quantization.
400 images trained for 3100 steps at BS 2, original 1k - 1.5k mixed resolution.
VRAM usage was around 20 GB but couldn't fit BS 3.
Offset noise 0.02. Perturbation noise 0.01. Masks with 0.6-0.8 for the background + normalized area loss.
Timestep shift 1.33.

I used 20 steps with DPM++ Karras for the samples. They're noisy but good enough to see when the likeness comes through (was around step 2000).

It looks great with Z-Image Turbo in ComfyUI at LoRA strength 1.75.

I tried a more complex rank 64, alpha 4 LoRA with timestep shift 2.5 - 3.0 and it went nowhere.
LR 2e-4, Adafactor, Cosine schedule, full layer preset, 4500 steps at BS 4, 1k resolution, bf16. Filled 24 GB VRAM to the brim (+900 MB from OS).
4k images captioned with only automated tagging and processing. I think it doesn't like that.

[carlo0000]

I tried a more complex rank 64, alpha 4 LoRA

last time i read about that, i remember it said need to put in alpha the same amount or half of rank , so can you try 32 or 64 ?
and keep us update if new tests and setting, i can't try my self my gpu is very slow, so i always hope it's good enough at the first start

[blissiraDC]

@ansinerd I first downloaded the Z-Image repo manually like I did with Chroma, bit it couldn't load the model when I set a local path (something about "can't load single file format - try diffusers"). I had to set the base model to Tongyi-MAI/Z-Image and let OneTrainer download and cache it.

Had very good results with a character LoRA rank 16, alpha 1, full layer preset. LR 1e-4, constant schedule, Adafactor, bf16, no quantization. 400 images trained for 3100 steps at BS 2, original 1k - 1.5k mixed resolution. VRAM usage was around 20 GB but couldn't fit BS 3. Offset noise 0.02. Perturbation noise 0.01. Masks with 0.6-0.8 for the background + normalized area loss. Timestep shift 1.33.

I used 20 steps with DPM++ Karras for the samples. They're noisy but good enough to see when the likeness comes through (was around step 2000).

It looks great with Z-Image Turbo in ComfyUI at LoRA strength 1.75.

I tried a more complex rank 64, alpha 4 LoRA with timestep shift 2.5 - 3.0 and it went nowhere. LR 2e-4, Adafactor, Cosine schedule, full layer preset, 4500 steps at BS 4, 1k resolution, bf16. Filled 24 GB VRAM to the brim (+900 MB from OS). 4k images captioned with only automated tagging and processing. I think it doesn't like that.

---

THE most useful comment of 2026! Thanks for taking the time

" LoRA rank 16, alpha 1, full layer preset. LR 1e-4, constant schedule, Adafactor, bf16, no quantization" - Came out mint!

[dxqb]

presets here: #1280

• using float8 (W8) because z-image is hard to quantize. you might want to switch to float W8A8 if you have an RTX 40xx, or to int W8A8 with a high svd rank if you want to wait for SVD
• left timestep shift at 1.0 because we don't have a definitive best value. 1.0 works - higher values might work better
• best practices and tricks how to train a non-Turbo LoRA for Turbo like [Feat]: Z-Image non-Turbo: workflow and preset #1269 (comment) aren't meant to be covered by a preset. the preset is only a starting point
• I'm going to convert this issue to a discussion, so you can continue discussing these things there

[knoic]

I performed a full fine-tuning on Z-Image using float8 (W8) precision, 512 resolution, the PRODIGY_PLUS_SCHEDULE_FREE optimizer, on 260 images with a timestep shift of 1.33. The results were good, and it sampled normally in OneTrainer. However, in ComfyUI, it produces a snowflake pattern. I tried the conversion script provided by ComfyUI (https://huggingface.co/Comfy-Org/z_image_turbo/blob/main/z_image_convert_original_to_comfy.py) and the script mentioned in another issue (Comfy-Org/ComfyUI#12165), but neither worked. Has anyone encountered or solved this problem?

[dxqb]

However, in ComfyUI, it produces a snowflake pattern. I tried the conversion script provided by ComfyUI (https://huggingface.co/Comfy-Org/z_image_turbo/blob/main/z_image_convert_original_to_comfy.py) and the script mentioned in another issue (Comfy-Org/ComfyUI#12165), but neither worked. Has anyone encountered or solved this problem?

Comfy created their own format (for full finetuning checkpoints, not for LoRA), that Z-Image themselves did not publish. Maybe we'll support it directly. I'll have to have a closer look if it's at least consistent with the original inference code, and not just some made-up format.

I've seen comments that the convert scripts worked for other people though

[dxqb]

https://github.com/Nerogar/OneTrainer/pull/1293/changes

[knoic]

nice

[DarkAlchy]

full layer preset? I take it you meant this?

[carlo0000]

hi, so what files are you using?

i m trying again, with the not turbo mode but so i have downloaded the original z_image_bf16.safetensors
and the z-image-Q8_0.gguf (i put in override transformers right?) i'm only familiar with the sdxl setting when there is only 1 file to load

but got error
NotImplementedError: Loading of single file Z-Image models not supported. Use the diffusers model instead. Optionally, transformer-only safetensor files can be loaded by overriding the transformer.
Exception: could not load model: C:/Users/carlo/AppData/Roaming/StabilityMatrix/Models/DiffusionModels/z_image_bf16.safetensors

[FennelFetish]

@DarkAlchy
Yes, but attn-mlp might be accurate enough with smaller risk for degradations.

@carlo0000
It also needs the text encoder, tokenizer, VAE, etc. which are not stored inside the checkpoint file anymore. Set the base model to "Tongyi-MAI/Z-Image" and it will download everything from the huggingface repository in the required diffusers format. It will be placed somewhere in your user's home folder, not sure where exactly on Windows.

[carlo0000]

@carlo0000 It also needs the text encoder, tokenizer, VAE, etc. which are not stored inside the checkpoint file anymore. Set the base model to "Tongyi-MAI/Z-Image" and it will download everything from the huggingface repository in the required diffusers format. It will be placed somewhere in your user's home folder, not sure where exactly on Windows.

hi i did that with the turbo but there is a problem with that, when i tested again a few days later it downloaded all again

if we add the q8 or the q4 in the override transformer, it should use less vram right? or that doesn't work in training ?

the folder is the /username/cache maybe should i copy the file to the models folder

ok so i got it to work, i use the Q8 to override transformer and confirm it reduce Vram 9.4 G with and 14 G without , and the speed on my gpu it's 8x faster with the Q8 , as it use share memory (gpu with 11G of Vram) at resolution 512 gona try higher resolution

[carlo0000]

hi i'm wondering with the curent setting we can only train from the original model right?
but i guess later we can train from an other fine tuned model, same as in sdxl ?

or we load the original base model, same as now and any other trained model we want to use will be put in override transformer?

i m wondering if in the future we can get lower vram like with sdxl
because with the curent setting i'm stuck at training below 1024 pixel
only setting what works is training is bf16 with loading of q8 gguf but increasing a bit of resolution use big amount of memory and speed x 12
even technically it's below 22giga of vram but at the start it use way more and get out of memory even there is some share memory still avalable

as on sdxl i could use fp8 and going from 1024x to 2048x pixel for exemple used less than 1 giga of vram more and even the speed did not double

ideal would be training at low resolution then the 1-2 last epoch at high resolution to have some more details even it's super slow

i did not try yet the gguf q4 to see if it decrease memory , ok i tried and it reduce the Vram a lot so i could increase the resolution, i guess for making a lora it should not make any difference

but for now , i'm wondering of my lora is working
with my low vram gpu, i cannot sampling in one trainer, (it will do a sample in 2-3 hours and use 24gig of memory) because it will sample on the original model and not the gguf Q8

so i'm at 4500+ steps lora rank 32 + alpha 16 with prodigy adv + cosine i put a higher learning rate 0.0008 time shift 1.33 , i did a test with q8 turbo, and there was 0 difference with or without the lora, and with the q8 not turbo, the picture was different but not even the hair color was right
4500 step but still nothing strange, i tried with the lora strength at 0.8 and 1.8

7500 step now and looks likes it's half way, prodigy the learning rate is supposed to be 1 but that is not working above 0.01 is to high and get only noise at output and also switched to constant , google IA suggested to use constant and keep learning rate at 1 and reduce lora alpha to 1 and reduce d coefficient to 0.5 or even 0.1 but i put d coef back to 1 and reduce learning rate rate, it s the same anyway

[carlo0000]

hey guy so how is your training going?

because mine has problem , my lora same's to work good on z turbo Q8 the character face match +90% of the time but when i test it on the no turbo version (Q8) then 95% of the time the face is not matching the subject, even worse than on sdxl (where you also had to try a lot of time to have the exact match)
i used time schifting 1.33 as someone posted here