Yes I can confirm it works with forge neo. I just downloaded the checkpoint file and placed it into my base model folder. It works for me using the same text encoder and vae I used for ZIT. Also I used fp8 quantization to fit it on my gpu.

@RisingV Ill try again, it gave me an error but we will see

Works fine. But I updated torch and sage first.

@makoshark1975 ok, I hope you can get it working

It loaded and ran after I did a clean install of forge neo, but only get black images now

@Volnovik Just giving me solid black images

@makoshark1975 check sampler and schedule, some give blacks. Just drag and drop my last images to check config

@Volnovik I fixed it, I uninstalled sage and it works fine now.

@makoshark1975 did you update torch? https://civitai.com/articles/25427/updating-torch-on-forge-neo-in-stability-matrix

@Volnovik Yes, but still didn't work until I got rid of sage, all good now

Training also. And the license will be updated at some point, yes.

License is apache-2.0 according to tongyi huggingface repo.
Also ai-toolkit already implemented training with z-image-base.

It's probably overtrained on chinese people. That's a common complaint with the turbo model. You should enter the nationality of the person that you want your subject to look like such as: Swiss, Nigerian, Mexican, etc.

Not havng this issue at all

I don't have this issue

z-image base solves this, way more variety in seeds and looks.

https://www.reddit.com/r/StableDiffusion/comments/1qozyms/a_quick_test_showing_the_image_variety_of_zimage/

@BlackBear31 Thanks

Yeah it needs it, the images remind me a bit of Pony V7, but quite a bit better.

It is not training Anime worth a damn.

that's not a Sage issue, that's a quant issue. Are you using qwen_3b or some albiterated vesion? Sage is just the catalyst. Also, put sage right up against your sampler

I am having the same issue, I am using forge neo. If i try to use the listed text encoder i get some kind of error about cant multiply mat1 and mat 2 or something and so i used the qwen3b and only get black images.

@makoshark1975 its because of text encoder, im using qwen3 4b without any problems on neo

@makoshark1975 are you using a conditioning zero out node or an empty clip text encoder?

Mat1 and mat2 errors are weight or model mismatches. Make sure that you have the dtype weight set to default in your diffusion model loader. Also, make sure you are using Lumina2 as the clip base. Also, make sure CFG is at 4.0 and you are doing at least 25 steps (not the issue or it'd be white, but check anyway)

@DevilLady I am not sure, I select lumina, set model to the new base model, use the same vae i was using and the qwe3 4b encoder, cfg at 4.5 and all i get are solid black images

@lonecatone23Im not sure what you mean by conditioning zero or antying about diffusion model loader or lumina2 as the clip base, i am using forge neo not comfyui 

@makoshark1975 ah.

I uninstalled sage and black images went away in forge neo

@lonecatone23 I'm using Qwen_3_4b.safetensors. I usually run my comfy with Sage running globally but I also tested KJ sage patch node while running comfy with pytorch attention and I still get black outputs.

@patrekt when was the last time you updated, or are you on (cringe) desktop?

@lonecatone23 Nah I'm on portable (v0.11.0-1). I updated comfy last night after the release

@patrekt odd. do you have it on auto, or set to a certain attention? I wouldn't be surprised if 2 or 3 are giving it issues (especially 3). Also, a lot of head scratching last night made me realize that you need an actual clip text encoder hooked up to the negative and not a conditioning zero out node.

@lonecatone23 Tried on auto and all the attentions, still getting the same black output. I'm running 3090 so Sage3 and FP8 variants doesn't work for me. and no I'm not using conditioning zero out. really have no clue why it works for some people and not for other ones like me.

@patrekt make a simple workflow, just diffusion model, clip, vae, clip text encodeer, model flow, latent, ksampler, decoder. , then plug it after the aura flow and before the sampler and see if you still have issues

@lonecatone23 still getting black output. maybe it has something to do with my torch or python version. I'm running Python 3.12.10 - Torch 2.8.0 - Cu128

Run it through Grok or Claude and see what it says. If you had Sage attention working before there is no reason why it shouldn't work now.

Do you have issues with other models with Sage Attention?

@lonecatone23 It works completely fine with other models. I also did a post on reddit and seems like other people are also facing the same issue with sage.

@patrekt Yeah, It's an epidemic. Ther are a ton of requests over it

the same problem. I launched comfy without --fast - it helped me

We'll need to 're'train our ZIT dataset for this again? :(

@TribalDiffusion The Lora's are unstable. some work, some don't. every one trained on the site is crap. Plus, it does not merge with ANYTHING, even when forced.

Welcome Flux 2 Klein all over again

@lonecatone23 Sigh! I had created a ZIT Lora for faces of the people in native place, which seems to work. I'll test that with this Base mode. If it doesn't work, back it zit :-/

Yeah diversity is way better, less generic people and poses, e.g

https://www.reddit.com/r/StableDiffusion/comments/1qozyms/a_quick_test_showing_the_image_variety_of_zimage/

It is more for training than using directly. It isn't supposed to be more realistic than Z-Turbo today. You're going to need to wait for loras and fine tunes.

try to disable sage attention, when I had this problem I changed to pytorch and it worked.

@CRAZYAI4U That is what I did, just uninstalled sage from the venv and it seems to be working now. Crazy

@makoshark1975 which text encoder and vae are you using in neo? since the base is fp16, the provided text encoder is fp8 and the error message is that they cannot be mixed. I've tried using stuff from turbo version but the quality is worst than turbo.

IT WORKS WITH NEO BUT IT'S SLOW AND VERY INACCURATE

@Melodic_Possible_582589 I downloaded the VAE and Text encoder and model from https://huggingface.co/Comfy-Org/z_image/tree/main/split_files Seems ok so far

There are pros to waiting for me, actual pose/looks variety

https://www.reddit.com/r/StableDiffusion/comments/1qozyms/a_quick_test_showing_the_image_variety_of_zimage/

nope, variety and creativity of ZIB outweighs the speed of ZIT. i'm using 3070ti and can manage 1 image every 2 mins.

same blotches/ compression artifacts. It seems absolutely random. The safest way I've avoided them so far is sticking to 1024x1024. I went to turbos 1088x1920 and it just comes up every time, even at 40~70 steps itll do it and start to mutate.

Yeah I have similar thoughts, still gonna keep turbo for some random quick fun, and then use base when I want to have more specific images that listen to my specific prompts, with more creative images that have variety. Might start with low res to get ideas and then go full res when I'm happy.

read me comment next to yours, its the steps i think

Do you use sageattention? Try it without it.

I saw a comparison with sage attention doing that, and the one with it removed looked fine.

@HugMeIntoFace No sage attention for me here. such a mystery

@jj43797771 are you using 25-50 steps? and a CFG around 4?

@J1B Yep, using all the suggested settings. oddly lowing the CFG brings down the occurance of it happening.

you nee this TE, not fp8 https://huggingface.co/Comfy-Org/z_image/blob/main/split_files/text_encoders/qwen_3_4b.safetensors

I remember having a similar error while training a lora. It was due to incompatible training settings I think. This should have nothing to do with number precision (fp4, fp8 or fp16). If you are using a recent version of forge it is probably because of some wrong settings or wrong dependencies.

This is Z-Image-Base, the non-distilled version of Z-Image, which released on the 27th.

@theally It's weird. It weighs exactly the same as the turbor. 12Gb

The more important question is do z-image base loras work on ZIT and can you mix them successfully... I am just testing it now.

But yes ZIT lora do kind of work on base, but you have to set them on strength 3.0+

Guess it depends on what you mean with working. You can use them with base, but may not work as intended. I used one I trained with ZIT here, but at strength 0.1 it doesn't do much. Near full weight weight it looks pretty bad.

@RisingV I obviously mean
1) work as intended
2) look good

@Ses_AI ok, than no, at least for me
may be different for loras trained on realistic style images though

I unironically keep reloading the models page for this lmao!

Will hopefully go in today!

https://www.reddit.com/r/ZImageAI/comments/1qpalp0/zimage_base_finetune_process_experimentation/

Yo, I tried it and the tag "flat chest" does not seem to work, "small breasts" seems to somewhat work, but the best you can do is using "large breasts" etc. in the negative prompt.
I scheduled an image post with different prompts for this (PG and PG-13 only images):
https://civitai.com/posts/26216103

@RisingV thank you so much!
dammit, this one thing prevents me from training my art using z-image.
it just won't adhere or recognize flat chest and different chest sizes, regardless of the tags or captions. 😒

people laugh at me when I ask this, but I have flat-chested grown up women characters who z-image INSISTS on giving large breasts to, regardless of how I set up the training.

No problem. I guess you have to wait for a finetuned checkpoint then. Or you can maybe try to do text encoder training? Don't know if it will work with civitai trainer though.
Also training with danbooru style captioning seems to work with zimage, but maybe it is better to use natural language? Have not tried that yet.
Well, no love for flatties out there ;)

Something like this will almost certainly be solved by finetunes, but those will take some weeks/months to shake out.

Hang on for a little while longer. Now that Base is out, Z-Image stuff is definitely going to kick into higher gear.

sometimes a nipple-based LoRA can fix this automatically since it's giving more attention to a specific detail than making sure there is always big booba

Are you using ComfyUI with Sage Attention by any chance? If you are, don't use Sage Attention. It's a well know problem that Z-Image base introduces image artifacts with Sage Attention active, and sometimes it doesn't even work at all (gives black image).

@mmdd2543 Thank you for telling me that. I'm heading off to try again!

I'm having this problem, even though i never installed Sage Attention. I even reinstalled the latest Comfyui portable and i still have this problem

Yep. Took me forever to figure it out, almost gave up. Felt like a dumbass when I realized how simple the solution was.

you mean zimage turbo?

Ostris is working on it

If you want a distilled low-step model, that’s literally what Z-Image Turbo is.
This release is specifically for everything that Turbo cannot do — LoRA, diversity, stylistic control.
Asking for a DMD version of the base model is like asking the devs to remove 80% of its capabilities again 😄.

The turbo version is based on DMD. So, hopefully a DMD LoRA, which can be stacked on any finetuned non-distilled model.

https://x.com/ostrisai/status/2016561737055805914

Turbo is an 8 step model. It's going to result in a more realistic and higher quality image that is far less malleable.

@aising23 The student model parameters of DMD trainning will have catastrophic forgetting of the teacher model parameters。So, unless you do brand new CFG & steps distillation training or undergo a Full Rank LoRA to refine Zimage(non turbo) revert back to ZiT ; (

@whateverr Yes, a Full Rank LoRA can change Zimage (non turbo) back to ZiT, but...

It is not possible to make this LoRA part effective by modifying the weights. Currently, the only acceleration effect is Z-Image-Turbo-DistillPatch, but the acceleration effect is limited due to the inability to fully align the ZI "Base". We are trying to communicate with the Lightx2v team and HyperSD team, and perhaps we can try2 training a distillation adapter？

Did you use the same lora weight for generation?
I observed that I have to double the weight with loras trained on Base and using it with ZIT, compared to a lora trained on ZIT, to get the same impact.

I tried training with this but got an error

Why do not you just use lora you have trained on Base, with the Base itself?

@RisingV No, a LoRA trained with the base model doesn’t need any weight adjustments when using ZIT. I discovered that training beyond 2,399 steps doesn’t require any weight change when using the turbo model. Fewer than 2000 steps will require weight change.

@stewi0001 What app are you using? I use onetrainer. 

@Wingo Its slower, and I don't get good results from the base model. Its just a good model to train loras on.

@Mayer2003 how do you get good result from baseZimg Lora on Zturbo.
I trained on Aitoolkit with 3k step Linear LR0.0002 (Zbase).
when i try on Zturbo look like my Lora not working. face is not look like dataset.

@Mayer2003 I was trying a workflow i have I comfyui. It was saying that the model was missing parts needed for training. I might have been using the wrong node. I will give onetrainer a try. Thank you

@pathAi Try LR 0.0003 or 0.0005. Also try a higher lora rank like 64/64, 64/128, 128/128.

@Mayer2003 tried onetrainer and still getting errors. Any chance you can share your preset?

@stewi0001 

@Mayer2003 ok, thanks
may be because I trained on drawings or different settings

Did not know that onetrainer supports it. Does onetrainer support lowvram settings (like quantization and layer offloading)?
EDIT: seems like it does from what I can see from github/deepwiki page.

thats usually how things go with base models. wait for finetunes.

how is that? xD

@RPxGoon I meant it's not plug and play. You need to understand how to train loras/lokrs. You also need to put effort in prompting to lock it in a specific look. Negative prompting can no longer be ignored like before etc.

@cobaltpixiv520 在base得不到的，在ZIT也不会得到，反而更糟。相比之下Tubro最大的优势就是速度快和资源占用少

@cobaltpixiv520 ,let me describe my own impressions.

ZIT prompting is like FLUX one: what I prompt is what I get with good results. And personally I expected from ZIB the same quality as ZIT but with some fixes,something like ZIT de-deTurbo. ZIT is a powerful thing with loyal requirements to hardware and output quality comparable with SDXL despite all problems.

ZIB prompting is closer to SDXL one with grown system requirements due usage LLM as a text encoder and 30-50 steps to get good quailty (for vanilla SDXL often 25-30 was enough) and some things availible in ZIT out-of-box are hidden in ZIB beneath more complicated prompting. For people with 8gigs or less it's rather sensitive.

There shouldn't be a difference between ZIT and ZIB more than one between Flux Schnell and regular Flux, but it is. Even loras between ZIB and ZIT are not completely interchangeble.

That's the worst argument I've ever seen about a model. What you're saying is actually bad for the model because if you need to be an expert to set up a model and train images or prompts, then this model is just the worst thing ever.

@Fadoo2077 My point is simple. You want good generation? go with ZIT. You want better generation? put effort in learning ZIB, and use both.

It is a lot harder to use than ZIT, you need a good solid negative prompt and more careful prompting, I have struggle for days before reaching out for help from the Reddit /R/StableDiffusion community, some good tips here: https://www.reddit.com/r/StableDiffusion/comments/1qr60ja/how_are_people_getting_good_photorealism_out_of/

Agreed. Don't be lazy with the prompting and your results will improve. Lora use tho.. eeeh. Stick with prompting if you can.

@TheP3NGU1N  Maybe the best negative prompt for image quality I have tested so far is:

"mutated, mutation, deformed, elongated, low quality, malformed, alien, patch, dwarf, midget, patch, logo, print, stretched, skewed, painting, illustration, drawing, cartoon, anime, 2d, 3d, video game, deviantart, fanart"

I have also tested a longer on but I am not sure if it is better

"mutated, mutation, deformed, elongated, low quality, malformed, alien, patch, dwarf, midget, patch, logo, print, stretched, skewed, painting, illustration, drawing, cartoon, anime, 2d, 3d, video game, deviantart, fanart,noisy, blurry, soft, deformed, ugly, drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly, bokeh, Deviantart, jpeg , worst quality, low quality, normal quality, lowres, low details, oversaturated, undersaturated, overexposed, underexposed, grayscale, bw, bad photo, bad photography, bad art, watermark, signature, text font, username, error, logo, words, letters, digits, autograph, trademark, name, blur, blurry, grainy, morbid, ugly, asymmetrical, mutated malformed, mutilated, poorly lit, bad shadow, draft, cropped, out of frame, cut off, censored, jpeg artifacts, out of focus, glitch, duplicate, airbrushed, cartoon, anime, semi-realistic, cgi, render, blender, digital art, manga, 3D ,3D Game, 3D Game Scene, 3D Character, bad hands, bad anatomy, bad body, bad face, bad teeth, bad arms, bad legs, deformities, bokeh Deviantart, bokeh, Deviantart, jpeg , worst quality, low quality, normal quality, lowres, low details, oversaturated, undersaturated, overexposed, underexposed, grayscale, bw, bad photo, bad photography, bad art, watermark, signature, text font, username, error, logo, words, letters, digits, autograph, trademark, name, blur, blurry, grainy, morbid, ugly, asymmetrical, mutated malformed, mutilated, poorly lit, bad shadow, draft, cropped, out of frame, cut off, censored, jpeg artifacts, out of focus, glitch, duplicate, airbrushed, cartoon, anime, semi-realistic, cgi, render, blender, digital art, manga, 3D ,3D Game, 3D Game Scene, 3D Character, bad hands, bad anatomy, bad body, bad face, bad teeth, bad arms, bad legs, deformities, bokeh , Deviantart"

These things take so long to test in Z-image!

The standard negative, provided by Tongyi, is

"泛黄，发绿，模糊，低分辨率，低质量图像，扭曲的肢体，诡异的外观，丑陋，AI感，噪点，网格感，JPEG压缩条纹，异常的肢体，水印，乱码，意义不明的字符"

You can translate it to english, but it's worse compared to the chinese results.

Tongyi's defaults are euler + simple + 30 steps, whereas comfy is res_multistep.

Did people struggle? I tried running prompts that ZIT couldn't quite get right, and Base was nailing and looked better. Maybe users prefer the more efficient, snappier tag format.

@TheP3NGU1N  What's the issue with Loras? Does Z-Image just not inherently support it like SD/Lumina?

Z-architecture requires natural language, not tag-based like 1-3 words, then comma and repeat, or how prompting works for danbooru-based data.

There are extensions that can do this for you

https://github.com/KohakuBlueleaf/z-tipo-extension

Will take you a while to get used to, but then it's time to jam.

@Fatbuns Ah thanks, that explains why my images were coming out with a yellow tint before as the first part of that default Negative translates as "Yellowed/discoloured, greenish tint,"
now it seems mostly fixed , great find!

@J1B If you're on Forge, try to reduce "Shift" parameter (a tip from somewhere from Forge-Neo github). Forge's default 6 , I reduced to 4 and it started working.

Definetely work: euler, euler_a, DPM2, res_multistep; scheduler: simple.

Here is my post on FP8 with shift=4.

My settings are Res_multistep + beta, shift 4, 30 steps, cfg 4

@mphobbit I use ComfyUI, but perhaps the ModelSamplingAuraFlow node in ComfyUI is the equivalent shift value, I will have to test that out and see if it is compatible with Z-Image. It never really changed much in Qwen-Image models for me, so I stopped using it.

Perhaps the ModelSamplingAuraFlow node in ComfyUI is the equivalent shift value

Just checked Comfy. Yes, you're right, that's a shift value. It's 3.0 by default, has to work.

> It never really changed much in Qwen-Image models for me, so I stopped using it.
Qwen uses its own architecture, Z-image based on Lumina according to the model card.

It is great to work as a base for a 2nd ZIT refiner pass as it gives so much more varited and interesting faces and poses.

I have a ComfyUI workflow posted here: https://civitai.com/models/2231351?modelVersionId=2644538

@J1B +1👍

I also recommend using a combination with UltraFlux VAE.

Seems slightly worse image quality in my testing: https://civitai.com/posts/26326998
about the same quality as the fp8 version, but that is just 1 image, but that took over 20 mins to make on my 3090.

@J1B If you use your mega negative, try something shorter. In my tests, the result is an order of magnitude higher than the BF 16 version. Also, the pure model gives better results than using the LORas that are currently available. I also use a standard loader. In terms of speed, it doesn't take much longer to generate than BF 16.

@lizirgin2012227 idk about faster. It depends on your overhead. I was doing 1280x1600 and with the bigger model my overhead is way down shrank and those can take 2-4 minutes, for an image lol. In terms of quality though it's insane. All those weird hands, weird stuff in the background, random small errors that would happen all go away with the full fp32 model.

@tenstrip I don't know I have been using that fp32 version most of the day and still seeing lots of broken hands, the only thing that fixes them is a 2nd pass with a ZIT model, it is just so much better at that.

So, I have tested the FP32 next to BF16.
On my peasant 4070ti 12gb vram (2.3gb ram used, 11.3/12.0 gb vram used), the speed is the exact same on the latest comfy according to the console (2.45s on new prompt, 2.3s on reused prompt).

In terms of quality, it may be slightly better on creative content, but not on realistic content.

HOWEVER.

NSFW is still botched. I get drastically more limbs than I used to, with the same prompts.
It does also seem to be closer to the same seeds like turbo is, which I find suspicious and strange.

I'm not using the CacheDiT extension to speed up my creations.

@Fatbuns if you're getting same rendering speed; that tells me everything I need to know about your test lol.

And you just wouldn't care to elaborate, nope. Waste of your time I bet.

120% AI generated bs, and from a terrible AI too

Would you care to do elaborate with side-by-side seeds with the same settings & providing the metadata to them?

It is not any slower in my testing. apart from the first load.

@J1B thats because you're not running it in fp32 which defeats the whole purpose of using it. But I still find QWEN to be better and way faster in fp8

@delta45424155 No, I always run it default weights, it is probably because I have 24GB of Vram (or 32GB if I using it on Runpod) so it doesn't have to offload to RAM.

I never use Qwen-Image in fp-8 the lora compatibility is way worse.

@J1B lol you have to use runpod to achieve 32gb VRAM.

@robotrobotbeepboop386 "As of early 2026, approximately 0.37% of Steam users own an NVIDIA GeForce RTX 5090."

@J1B if the fp32 renders at same speed as whichever one you're testing; it is not using fp32 weight... Even on my new 5090 the fp32 version takes roughly 3x longer vs this model. The only advantage I have found tho, with the fp32, is the detail you can go into with your prompt and not have the output messed up or not follow the prompting.

@delta45424155 I just tested it again, and FP32 is apparently running at the same speed as the FP16 (doesn't really make sense). I'm not sure why, I am not selecting quantisation on the model loader and it is the FP32 model as it is 22.9GB.
I know how to load models at different precisions , I have published over 50 models.
could be a bug in ComfyUI or my workflow I guess, but the outputs look differnet so I doubt it is just running it at fp16: https://civitai.com/images/119697733
Could be a fake of broken model as it was never officially releases just "leaked", although I don't really buy that story, something dodge is going on with the people that have been posting it here with brand new accounts with zero posts.

@J1B when you launch comfy; use the flag that forces fp32 when testing. But you'll also want to build your text encoder yourself from shards. I do not recommend it unless you have 64gigs+ of ram and at least a 5090. You'll notice the results with a heavy prompt, but imo it isn't worth it. Especially since loras will be trained and tested on what works for way lesser hardware.

Relax with the conspiracy theories. :) Not everyone treats Civitai as a social network. Some of us just use the tools.

​I tested the FP32 version on high-end hardware because I need the best possible source for training, not for 'hype'. If providing a technical breakdown of why 32-bit float retains more detail than 16-bit quantized makes me 'suspicious', then so be it. Use whatever works for your workflow.

@lizirgin2012227 Why not share your testing then? I shared my testing and it looks worse than bf16: https://civitai.com/images/119697733

@J1B To be honest, I’m already quite busy managing my current social channels for my fanbase, so I can barely keep up as it is. If I find the time and motivation, I might make a comparison post later.

​I admit I’m a bit surprised by the divide regarding the FP32 version—it’s strange that some see a difference while others don’t. Like everyone else, I’m confused about this leak, but for me personally, the difference is obvious.

"To be honest, I’m already quite busy managing my current social channels for my fanbase, so I can barely keep up as it is. If I find the time and motivation, I might make a comparison post later.

​I admit I’m a bit surprised by the divide regarding the FP32 version—it’s strange that some see a difference while others don’t. Like everyone else, I’m confused about this leak, but for me personally, the difference is obvious."

Another em-dash..

@J1B What startup flags did you use; if using comfyui?

@delta45424155 As expected, you don't care about elaborating topkek. You are asking for the evidence that you don't provide yourself. Hypocrite.

@delta45424155  I have one .bat file with --UseSageAttention and one without, thats about it. But I have 24GB of Vram.

those are just bs, afaict, the so called fp32, saved by diffusers v0.36, is just a older version of final released model which is saved by newer mainline diffusers (not released so tagged as v0.37.dev). Seems z-image team decided to postpone the release at the final moment for a fine-tuning.

And, if those bs accounts can prove fp32 is different drastically than bf16. don't be shy, share the evidence, it gonna shake the foundation of the entire AI world.

those accounts just want people to run an old z-image model in fp32 which requires 24gb vram and is extremely slow, for no reason. Even with AI generated bs. sus

👍😁

The main problem with this 8 step lora is it is killing the image variation in Z-image which was the main reason I have been using it. I am trying to make a ComfyUI workflow that brings it back.

@J1B +1 but good for photorealism.

Try Euler - simple.

Are you using a good Negative prompt?

This model needs a negative even more that a positive!

remove the --enable--sage--attention line from cli, also don't use sage attention node

Simple/beta/sgm_uniform schedulers typically work.

I'll make the adjustments and try using it again. Thanks to everyone for the tips!

nice story,

first, sage attention + z-image is unusable, no body is using it.

Second, z-image is a bf16 model. fp16_accumulation has nothing to do with it.

I'm really curious, why did you do this? misleading people that devs is hiding a "better version" and the model requires 24gb vram to run and is extremely slow?

You have to remember that it's a Base Model. It's primary function is for training, fine-tuning, and LoRA development, not inference. If you look back at "Base Models from History", most of them have been pretty rocky at launch, and only through community effort have become fantastic derivative resources.

@theally I hope this is true.

I only wonder how good ZIT is compared to it's base. Okay we don't know what the makers actually used to make ZIT from but I wonder if it really was this base model or sth else.

Some of the quality loss comes down to image resolution as well, With the on site generator Civitai are pretty mean with their max resolution choices (Apart from PonyV7 that gets special treatment)
Here is an image made locally at 1280x1568:https://civitai.com/images/120083106
and the same prompt on the generator at 832x1216:https://civitai.com/images/120070564
It is not a huge difference in this example but the image does feel less pixelated and more fine detailed at the higher resolution.

"Proper" prompting does help. With that said, you will need the help of an LLM for that.

For instance, you can try and throw a random image into one and have it describe the image in a detailed natural language format. Depending on its answer, you will get 6 to 16 lines. You can start out with the full prompt, then shrink it down depending on what you want out of it - if it's a good prompt :P

@Fatbuns I will try that, thanks

Check your settings on local machine .

If you're forge user, first, reduce Lumina shift parameter. I got into trouble when kept default value of 6. I reduced to 4 and ZIB started working. In comfy ModelSamplingAuraFlow node is responsible for shift. You have to unlock it if you use default ZIT workflow and do nothing (it's already set to value of 3).

Try:

Sampler: Euler/Euler_a/DPM2/res_multistep

Scheduler: beta/simple/sgm_uniform

Cfg: 3.5-4.5

Negatives: very long list, use your typical load from SDXL.

Here are some my outputs with FP8 to give you ideas.

Online ZIB is rather poor (I tried on TA)

On CivAI I find a cfg one 1 is pretty much mandatory if that helps.

Sorry, haven't heard anything new about Z-Image-Edit or Z-Image-Omni-Base!

No rumors on china-net either.

Use encoder and VAE from same page (other tabs on top).

Personally I use this encoder with Forge Neo.

@msux oh damn im an idiot. thanks so much!!!