๐ Z-Image AIO Collection
โก Base & Turbo โข All-in-One โข Bilingual Text โข Qwen3-4B
โ ๏ธ IMPORTANT: Requires ComfyUI v0.11.0+
๐ฅ Download ComfyUI
โจ What is Z-Image AIO?
Z-Image AIO is an All-in-One repackage of Alibaba Tongyi Lab's 6B parameter image generation models.
Everything integrated:
โ VAE already built-in
โ Qwen3-4B Text Encoder integrated
โ Just download and generate!
๐ฏ Available Versions
๐ฅ Z-Image-Turbo-AIO (8 Steps โข CFG 1.0)
Ultra-fast generation for production & daily use
โซ NVFP4-AIO (7.8 GB) ๐
๐ฏ ONLY for NVIDIA Blackwell GPUs (RTX 50xx)!
โก Maximum speed optimized
๐พ Smallest file size
๐ FP4 precision - blazing fast
Perfect for: RTX 5070, 5080, 5090 owners who want maximum speed
๐ก FP8-AIO (10 GB) โญ RECOMMENDED
โ
Best balance of size & quality
โ
Works on 8GB VRAM
โ
Fast downloads
โ
Ideal for most users
Perfect for: Daily use, testing, RTX 3060/4060/4070
๐ต FP16-AIO (20 GB)
๐พ Same file size as BF16
๐ ComfyUI auto-casts to BF16 for compute
โ ๏ธ Does NOT enable FP16 compute mode
๐ฆ Alternative download option
Note: Z-Image does not support FP16 compute - activation values exceed FP16's max range, causing NaN/black images. Weights are cast to BF16 during inference regardless of file format.
Perfect for: Alternative to BF16 download (identical inference behavior)
๐ BF16-AIO (20 GB) โญ RECOMMENDED FOR FULL PRECISION
โ
BFloat16 full precision
โ
Absolute best quality
โ
Professional projects
โ
Also works on 8GB VRAM
Perfect for: Professional work, maximum quality
๐จ Z-Image-Base-AIO (28-50 Steps โข CFG 3-5)
Full creative control for pros & LoRA training
๐ก FP8-AIO (10 GB)
โ
Efficient for daily use
โ
Full CFG control
โ
Negative prompts supported
โ
8GB VRAM compatible
Perfect for: Daily work with full control
๐ต FP16-AIO (20 GB)
๐พ Same file size as BF16
๐ ComfyUI auto-casts to BF16 for compute
โ ๏ธ Does NOT enable FP16 compute mode
๐ฆ Alternative download option
Note: See technical explanation in FAQ below.
Perfect for: Alternative to BF16 download (identical inference behavior)
๐ BF16-AIO (20 GB) โญ RECOMMENDED FOR FULL PRECISION
โ
Maximum quality
โ
Ideal for LoRA training
โ
Professional projects
โ
Highest precision
Perfect for: LoRA training, professional work
๐ Turbo vs Base - When to Use?
โก Use TURBO when:
โก Speed is priority โ 8 steps = 3-10 seconds
๐ธ Production workflows โ Consistent high quality
๐พ Quick iterations โ Rapid prototyping
๐ฏ Simple prompts โ Less complex scenes
๐จ Use BASE when:
๐จ Creative exploration โ Higher diversity
๐ง LoRA/ControlNet dev โ Undistilled foundation
๐ Complex prompting โ Full CFG control
๐ซ Negative prompts needed โ Remove unwanted elements
โ๏ธ Recommended Settings
โก Turbo Settings (incl. NVFP4)
๐ Steps: 8
๐๏ธ CFG: 1.0 (don't change!)
๐ฒ Sampler: res_multistep OR euler_ancestral
๐ Scheduler: simple OR beta
๐ Resolution: 1920ร1088 (recommended)
๐ซ Negative Prompt: โ Not used!
๐จ Base Settings
๐ Steps: 28-50
๐๏ธ CFG: 3.0-5.0 (start with 4.0)
๐ฒ Sampler: euler โญ OR dpmpp_2m
๐ Scheduler: normal โญ OR karras
๐ Resolution: 512ร512 to 2048ร2048
๐ซ Negative Prompt: โ
Fully supported!
๐ Quick Overview
Turbo Versions
โซ NVFP4 โ 7.8 GB โ RTX 50xx only โ Max Speed ๐
๐ก FP8 โ 10 GB โ 8GB VRAM โ Recommended โญ
๐ต FP16 โ 20 GB โ โ BF16 compute โ See FAQ โ ๏ธ
๐ BF16 โ 20 GB โ 8GB VRAM โ Max Quality โญ
Base Versions
๐ก FP8 โ 10 GB โ 8GB VRAM โ Efficient
๐ต FP16 โ 20 GB โ โ BF16 compute โ See FAQ โ ๏ธ
๐ BF16 โ 20 GB โ 8GB VRAM โ LoRA Training โญ
๐ก Prompting Guide
โ Good Example:
Professional food photography of artisan breakfast plate.
Golden poached eggs on sourdough toast, crispy bacon, fresh
avocado slices. Morning sunlight creating warm glow. Shallow
depth of field, magazine-quality presentation.
โ Bad Example:
breakfast, eggs, bacon, toast, food, morning, plate
๐ Tips
DO:
โ Use natural language
โ Be detailed (100-300 words)
โ Describe lighting & mood
โ Specify camera angle
โ English OR Chinese (or both!)
DON'T:
โ Tag-style prompts (tag1, tag2, tag3)
โ Very short prompts (under 50 words)
โ Negative prompts with Turbo
๐ Bilingual Text Rendering
English:
Neon sign reading "OPEN 24/7" in bright blue letters
above entrance. Modern sans-serif font, glowing effect.
ไธญๆ:
Traditional tea house entrance with sign reading
"ๅค้ต่ถๅ" in elegant gold Chinese calligraphy.
Both:
Modern cafe with bilingual sign. "Morning Brew" in
white script above, "ๆจๆฆๅๅก" in Chinese below.
๐ฅ Installation
Step 1: Download
Choose your version based on:
GPU: RTX 50xx โ NVFP4 possible
VRAM: 8GB โ FP8 recommended
Purpose: LoRA Training โ Base BF16
Step 2: Place File
ComfyUI/models/checkpoints/
โโโ Z-Image-Turbo-FP8-AIO.safetensors
Step 3: Load & Generate
Open ComfyUI (v0.11.0+!)
Use "Load Checkpoint" node
Select your AIO version
Generate!
No separate VAE or Text Encoder needed!
๐ Credits
Original Model
๐จโ๐ป Developer: Tongyi Lab (Alibaba Group)
๐๏ธ Architecture: Single-Stream DiT (6B parameters)
๐ License: Apache 2.0
Links
๐ Z-Image Base: https://huggingface.co/Tongyi-MAI/Z-Image
๐ Z-Image Turbo: https://huggingface.co/Tongyi-MAI/Z-Image-Turbo
๐ง Text Encoder: https://huggingface.co/Qwen/Qwen3-4B
๐ Version History
v2.2 - FP16 Clarification
๐ Updated FP16 descriptions for technical accuracy
โ ๏ธ Clarified: FP16 weights โ FP16 compute
๐ FP16 files are cast to BF16 during inference
v2.1 - NVFP4 Release ๐
โ Z-Image-Turbo-NVFP4-AIO (7.8GB)
โก Optimized for NVIDIA Blackwell (RTX 50xx)
๐ Maximum speed generation
v2.0 - Base AIO Release
โ Z-Image-Base-BF16-AIO
โ Z-Image-Base-FP16-AIO
โ Z-Image-Base-FP8-AIO
๐ ComfyUI v0.11.0+ support
๐ Qwen3-4B Text Encoder
v1.1 - FP16 Added
โ Z-Image-Turbo-FP16-AIO
๐ง Wider GPU compatibility
v1.0 - Initial Release
โ
Z-Image-Turbo-FP8-AIO
โ
Z-Image-Turbo-BF16-AIO
โ
Integrated VAE + Text Encoder
โ FAQ
Q: Which version should I choose?
RTX 50xx + Speed โ NVFP4 ๐
Most users โ Turbo FP8 โญ
Full precision โ BF16 โญ
LoRA Training โ Base BF16
Q: Turbo or Base?
Fast & simple โ Turbo โก
Full control โ Base ๐จ
Q: Will NVFP4 work on my RTX 4090?
โ No! NVFP4 is only for RTX 50xx (Blackwell architecture).
Use FP8 instead for RTX 40xx and older.
Q: Do I need separate VAE/Text Encoder?
โ No! Everything is already integrated.
Just Load Checkpoint and go!
Q: Works on 8GB VRAM?
โ Yes! All versions work on 8GB VRAM.
(NVFP4 requires RTX 50xx regardless of VRAM)
โ ๏ธ Q: What about FP16 for older GPUs (RTX 2000/3000)?
Important technical clarification:
Z-Image does NOT support FP16 compute type. Here's why:
๐ Technical reason:
- FP16 max value: ~65,504
- BF16 max value: ~3.39e+38 (same as FP32)
- Z-Image's activation values exceed FP16's range
- Result: Overflow โ NaN โ Black images
What actually happens:
ComfyUI automatically casts weights to BF16 for computation
You can see this in logs: "model weight dtype X, manual cast: torch.bfloat16"
"Weight dtype" (file format) โ "Compute dtype" (actual calculation)
For RTX 20xx users (no native BF16):
BF16 is emulated via FP32 = slower but works
There is no way to run Z-Image in true FP16 compute
FP8 with CPU offload may be a better option for limited VRAM
TL;DR: FP16 and BF16 files behave identically during inference. Choose based on download preference, not GPU compatibility.
๐ Get Started Now!
Download โ Load Checkpoint โ Generate!
Recommended versions:
๐ก FP8 for most users (best size/quality balance)
๐ BF16 for maximum quality
โซ NVFP4 for RTX 50xx speed
All versions work on 8GB VRAM
Happy generating! ๐จ
Description
! Update ! FP16 version released
Why an FP16 version?
FP16 is natively supported by most older GPUs โ especially NVIDIA GPUs below the 4000 series. This makes the FP16 version the best and most compatible choice for users running older hardware.
NVIDIA GTX 1000 / RTX 2000 / RTX 3000 / RTX 4000
The other two versions, BF16 and FP8, are recommended if you are using an NVIDIA 4000 series GPU or newer, as these architectures are optimized for those formats and can take better advantage of them.
In short:
FP16 โ best compatibility for older GPUs
BF16 / FP8 โ best choice for NVIDIA 4000 series and newer
This update simply gives everyone the option to use the version that fits their hardware best.
FAQ
Comments (22)
! Update ! FP16 version released
FP16 offers the widest compatibility and works on virtually all GPUs.
BF16 is supported starting with NVIDIA RTX 3000 (Ampere) and newer.
FP8 is primarily optimized for NVIDIA RTX 4000 series and newer, where it can be used most efficiently.
Choose the version that best matches your hardware for the best experience.
Thank you! Did you use the regular VAE or UltraFlux?
@Aieditorย the regular VAE ๐
Me and my Radeon thank you! o7
I have an RTX 3060 with 12GB Ram, trying out the fp8, seems to work fine (around 25 seconds for a 832x1216 image)
Can I make the FP16 version work? I have 32 GB Ram. But I don't know if that'll make it much slower.
Not to sound like a dick - because this IS my go to checkpoint for Stability Matrix, so really, thanks for the work you did - but using different seeds and samplers even if the prompts are the same make the differences between the checkpoints look a lot bigger than they are. Take for example the dragon image. You're using 3 different seeds and 2 different samplers. Even if everything else is the same, that's still enough to give us different images. Is FP16 softer than FP8 and BF16? or is it just euler + Beta that gives it that look? Are those really the differences between FP8 and BF16 or is it just to the seed.
I've tested both the FP8 and BF16 earlier in the month and there are differences, but for the same seeds the composition is mostly the same, FP8 being just more prone to error than BF16 (and extra hand here, and extra leg there). I unironically preferred the FP8 images in most cases I tested, even if they were more likely to have error. But based on the images you're showcasing, the differences seem a LOT bigger because of different seeds. (Of note though, ComfyUI seems to have had an update with their KSampler and it seems to no longer have a fixed seed option in it and I haven't had much luck with forxing a fixed seed with other nodes, so in case you though you were using the same seeds, that might be why your'e not)
BTW, the FP8, did you do any calibration on those or just weight quantization without calibration?
Thanks for the detailed comment โ and no worries at all, I didnโt take it the wrong way ๐
I really appreciate you taking the time to test things thoroughly, especially since youโre actively using this checkpoint in Stability Matrix.
Youโre absolutely right that different seeds, samplers, schedulers, and step counts can significantly change the output, even when the prompt itself stays the same. In my showcases, I intentionally keep the prompts mostly identical but vary settings like seed, sampler, scheduler, and steps, because all of these factors play a major role in the final image. The resulting differences are therefore expected and, to some extent, intentional.
Regarding FP8 vs. BF16:
The main difference I observe is that FP8 introduces more noise. FP8 images often appear dirtier or grainier, while BF16 tends to be cleaner and more stable. With identical settings, composition between FP8 and BF16 usually remains very similar, but FP8 is more prone to structural errors.
So FP16 or BF16 themselves arenโt inherently โsofterโ โ what often gets interpreted as softness is usually a result of noise characteristics combined with sampler and scheduler behavior, rather than the numeric format alone.
On the topic of fixed seeds in ComfyUI: I just tested this again using two checkpoint loaders (FP8 and BF16) with the same seed, and in my setup the results remain fully deterministic. I wasnโt able to reproduce the issue you mentioned, so at least on my end, fixed seeds are still behaving as expected.
As for FP8 specifically: these builds are weight-quantized without an additional calibration pass. That choice was mainly about maintaining iteration speed and compatibility, but calibrated FP8 variants are definitely something I may explore further.
Thanks again for the thoughtful feedback โ discussions like this are genuinely helpful and push the project forward.
@SeeSeeLPย It's the Node 2.0 thing comfyui is testing. Turn that on and you lose access to the ksampler's control after generation option for seeds. The new node 2.0 UI hides A LOT of stuff and it's usually for the worse.
As for my point about seeds, what I meant was that since the only difference between these checkpoints is precision (as opposed to your anime checkpoint), people would be better served were they able to see the same set of images generated under the same conditions bar the checkpoint and judge the precision trade-off without having to download and test each checkpoint individually.
Thank you for the update! <3
One of the only AIOs I've tried - I am quite impressed. I will mess around with the FP8 for a bit.
Thank you so much for the feedback :-)
@SeeSeeLP
the best AIO Pack PERIOD!!!!
much love and keep it going ๐โค๏ธ๐
Thank you so much, cynic2010! ๐ฅฐ
Your kind words really made my day. Iโm super happy youโre enjoying the AIO pack โ much love back to you! ๐โค๏ธ
ๅจC็ซไธญไฝ ็fp8ๆฏๆ่ฎคไธบ็ฐๅจๆๅฅฝ็๏ผ็ฎๆด็้ข๏ผไบบ็ฉ่ขไฝ่ฏๅฅฝใ่ฐข่ฐขไฝ ๏ผ
่ฐข่ฐขไฝ ๏ผๅพ้ซๅ ดไฝ ๅๆฌข FP8 ๆจกๅ๏ผๅธๆไฝ ็ฉๅพๅผๅฟ๏ฝ ๐โจ
Anyone else getting a BSOD when trying to run any of these? Using a 4060
Hi @TeKniKo Iโm actually using an RTX 4060 myself and havenโt run into any BSOD issues. Itโs likely something with your settings. You might want to check temperatures, your power supply, or try running a fresh ComfyUI installation in a separate folder to see if that helps.
@SeeSeeLPย I figured out the issue. I was using base ForgeUI which doesnt support ZIT. I had to use a different branch. Now Im addicted to the quality lol.
@TeKniKoย Nice, glad you figured it out! ๐
Yeah, base ForgeUI can be a bit tricky with ZIT support. Happy to hear itโs working now โ and welcome to the addiction, the quality really hits hard ๐๐ฅ
Have fun generating!
Does the Qwen 8B text encoder work with z-image, or only the smaller one?
Hi @rivdemon1221554 , I havenโt tested this myself, but I donโt think it will work. Z-Image uses Qwen3-4B as the text encoder, not the Qwen3-8B version, so the 8B encoder is likely not supported there. That said, this is just my assumption since I havenโt tested it directly.
@SeeSeeLPย Yeah, doesn't for me. It's a shame, Z-image is fast but it just doesn't understand the level of detail that FLUX Klein 9b and especially Qwen 2512. I get people hate slow stuff but I've tested multiple character images with all 3 and I see now, there's definitely a reason Z-image is faster and cheaper.