🔄 Updates
12.07.26 - There are better Krea2 models on this site , i made this one for me . The format is INT8 ConvRot , based on TURBO , so CFG 1 , 8-11 steps , euler_a - bong tangent ( at 1024x1440 i get ~4s/it on a RTX3060 12GB , FP8 i get ~6.5s/it ) . It very much likes simple workflows ( all my images have them embedded) , for sampler and scheduler see workflow or image details .
10.07.26 - V14 best settings in my opinion : Resolution 832x1216 , euler_a with Bong_Tangent , 11 steps , cfg 1 , clip skip -2 , simple workflow .
07.06.26 - V8 is in lora form ( bastardlora2)
06.06.26 - managed to get V7 to generate on site but it does not work , don't waste buzz with model , just use the LoRAs
14.05.23 : Remade the 7th model , if you get black images with F16 model use the repaired BF16 model .
02.05.26 : Remade the model, should work now
01.05.26 https://pastebin.com/sKxCUSb6 script i used , replace base model , variants , weights , density and lambda with what you want . takes ~10 minutes on 32GB ram . If you encounter problems with the model please leave a comment or an image with prompt.
22.04.26 : Model (5.5) still not showing in model page or my profile :)) after almost 24H , I sure picked the best time to upload lol . Works with lora , tested some character lora as well , i have no idea why some work and others are not working ... almost deleted this model with all other model tests.
Works best with euler simple , CFG 1 and 9 steps . Try CFG 1.1 or max 1.5 with 9 to 11 max steps . Use a clip set last layer node and set it at -5 ;) .
15.04.26 : Made an inbreed model , it manages better in my opinion . Still not NSFW enough :( no breasts, nipples or puss control . LoRAs not tested with it . leave a comment or a image even with errors to see what i can do about it .
30.03.2026 : For fast conversion use this nice TOOL , made a Q8_0.gguf in " info: Completed in 29.14 seconds " .... yeah
25.03.2026 : Managed to modify the convert.py script from city96 to make the model directly in Q8_GGUF . Gemini for the win . 
23.03.26 - Modified this nodes so it can accept negative weights and remade 3.5 with original ZIT for better layer stability ...i think . Merged with 2 random models from tensor hub for aesthetics , poses , clothes . I have no idea what it does better than v3.5 but i think it's different .
Usage : same as before - euler with simple , CFG 1 with 9 steps (best working and less horror, original text encoder (it does everything you want as long the model knows it).
Same as before i can't convert to GGUF , if someone knows and manages to do it he can post the GGUF model as his own (with a little mention of me :D ) ...
(Sample images are 1152x1536 , euler simple, 9 steps , no aura flow , no clip-skip , no LoRA , first run...the first run is bad for me after comfyui 0.17 update so the 2nd run without changing the prompt is better)
05.03.26 - File weights are broken because of lora / model merger node and can't convert to GGUF // stuck with this ...
04.03.26 - V3.5
Sample images are 1024X1536 , 15 steps , CFG 1.5 ( on a RTX3060 12GB i get ~5-6s/it , Prompt executed in ~100 seconds - depends on how long the prompt is)
About this version
>>>Remake of V3 with new custom nodes from capitan01R (leave a star for him and try the new lora loader it's awesome) + some more LoRAs<<<
Just like V3 most samplers/schedulers work and it has many changes from one to another , test them , mix them , some have body merging some don't , some work for fantasy some work for "erotic" stuff . Try them , try them with CFG >1 , try them with 10+ steps etc. Will update when i get more interesting LoRAs .
Sampler : Euler_A / Euler
Scheduler : sgm_uniform / simple
Steps : 8 - 15
CFG 1 - 1.5
No Aura Flow Shift (better text render and no skin correction or whatever that shift does ) ,
First generation with these settings is sometimes bad without a good prompt but the 2nd or 3rd without changing the prompt is good .
It should work with ZIB and some ZIT LoRAs at low strength (did not test) .
Best settings for me V3:
Sampler : er_sde / Euler
Scheduler : Simple
Steps : 8-9
CFG 1 (sometimes 1.5)
No Aura Flow Shift (better text render and no skin correction or whatever that shift does ) ,
Added a CLIP set last layer node and i like to use -3 . Every change in clip skip is wild try from -1 to -10 (I'm too lazy to go past that) .
First generation with these settings is sometimes bad without a good prompt but the 2nd or 3rd without changing the prompt is good .
26.02.2026 - changed file
Z-Bastard V3
V3 = V2 (identical model).
One version was removed because merging and re-merging produced the same result.
Recommended settings (ComfyUI):
Sampler: Euler , Euler_A , DDIM , res_multistep etc .
Scheduler: Simple , DDIM_uniform , SGM_Uniform , Beta etc.
Try Aura Flow Shift: 2.5–3 max
Also try without Aura Flow Shift — it behaves like a different model and the text is better.
I use different sampler and scheduler almost every time, just load the image you like in comfyUI to see and compare with others .
Higher Flow shift tends to make the skin more SDXL like .
🧠 How to Use Z-Bastard V1
Use it like any Z-Image Turbo style model.
ZIB-based LoRAs work well.
🧠 How to Use UnstableBastard Illustrious V1 and V2
General Generation
Sampler: euler_A
CFG: 3–5
Steps: ~35
Test different samplers — results can vary nicely.
🔍 High-Res Fix script (ComfyUI)
Using Efficiency Nodes Highres-Fix Script:
Upscale: 1.25
Steps: 12
Iterations: 1–2 (2 is my favorite , refined from 768x1152 to 1200x1896 in ~120s with my 3060 12GB)
Denoise: 0.26-0.36
👤 Face Detailer (ComfyUI – Impact Pack)
Recommended settings:
Steps: 8
Denoise: 0.26-0.36
Lower chance of male faces becoming feminine
Works with upside-down faces
UltralyticsDetector Face model
If using ComfyUI, just load one of my images (workflow is embedded) and run it.
🎨 Realistic Images
For realistic outputs from ill with v1 and v2 i use :
CyberRealistic / CyberIllustrious
Highres-Fix Script enabled with one of the checkpoints above
You can use any checkpoint with this setup.
🎭 Artist Prompt Examples
You can improve style by adding artist tags like:
(artist:wamudraws:1)
(artist:slugbox:1)
(artist:zankuro:1)
(artist:mazjojo:1)
(artist:kotteri:1)
(artist:pigeon666:1)
(artist:makoto-shinkai:1)
(artist:zawar379:1)
(artist:yoneyama mai:1)⚠️ Weaknesses
Not great with complex backgrounds
Struggles with fine detail-heavy scenes
🔗 LoRA Strength Guide
1–2 LoRAs → 0.5–1.0 strength
Multiple LoRAs → 0.2–0.5 strength
(Adjust as needed.)
🧾 Prompts
Positive Prompt (My Go-To)
embedding:lazypos,
(masterpiece, best quality, amazing quality, ultra-HD, high detail, very awa,
newest, very aesthetic, ultra-detailed, absurdres, 8k)(Uses 1 embedding.)
Negative Prompt
embedding:lazyhand,
embedding:lazyreal,
Bad Quality, worst Quality, low Quality, photorealisticNote:
You can use
embedding:lazyrealin the positive prompt if you want a more realistic feel.
Description
Inbreed version of 3.5 , 4.5 and a failed version that i didn't post . Merged with ModelMergeBlocks node between all 3 . Not tested with LoRAs . It has improved somehow
FAQ
Comments (22)
FP8 Safetensors please
Is q_8.gguf not good ?
@BastardOG I cant use gguf
@Unicom oh , i'll upload a fp8 version of 5.5 , what version do you want e4m3 or e5m2?
@BastardOG Last version. thx!
@Unicom uploaded e4m3
@BastardOG GGUF often works slower in my personal case than safetensors (it spends additional time for ungguffing process). In case of FLUX I generated on GGUF Q4 with comparable time with FP8. So personally I try to avoid GGUFs.
@mphobbit for me (on rtx 3060 12GB) the s/it are the same for fp8 , fp16 , bf16 , and Q8...on my RX580 8GB that's where the s/it varies but not much . I got lucky and got a 3060 in 4 equal instalments without interest for about 350$ now that same card is ~400
ggufy https://github.com/qskousen/ggufy can make safetensors FP8 :) and also has an optional GUI now
@BastardOG it's not about it/s. Unggufing is a separate process before generation. It takes in my case of 3050 around 100-200 seconds and I prefer to spend them to generation itself.
@mphobbit oh , did not know tha , i'm just a beginner here who likes to test things and make them to my liking :D
@mphobbit i'm curious what you are using to generate images that requires you to de-quantize from gguf? the programs i'm familiar with allow you to use the gguf in inference directly without dequantizing
@ferretduck I'm not a technical specialist but the general idea that ggufs uses integer type instead of float. GPUs "think" in terms of floats. So with ggufs the GPU converts ints to floats. Here the brief explanation: https://huggingface.co/city96/Qwen-Image-gguf/discussions/13
@mphobbit you're not wring that gguf is not stored in a format that the GPU can use natively. however, i'm mostly curious about your comment that it spends 100-200 seconds before generation just to unquantize. or did i misunderstand, and you were saying that total generation time using gguf is 100-200 seconds? i beleive gguf de-quantizes in place for each layer as it goes through, keeping the vram usage low but potentially generating a little slower. on my 3090, there's no noticeable difference in speed in comfyui with safetensors vs gguf.
@mphobbit RTX20x0 and RTX30x0 can use FP16 and INT8, but not FP8, when those gpu see FP8 istructions they have to convert to FP16 wasting 2 cycles for EVERY fp8 instruction
@Meandmeitsperfect I'm not saying you're wrong, because there's a lot of this stuff I don't fully understand yet. But I just tested with a safetensors Klein model in bf16, f16, and f8 (same model, converted) on my 3090, and they were all very close to the same speed. The bf16 was 4.07s/it, f16 was 4.16s/it, and f8 was 4.05s/it
@Meandmeitsperfect I have no problems with any fp8 checkpoint or TE (e4m3 and e5m2 both) on my 3050. You misrecognized it with FP4. which, yeah, designed for 50xx series.
@mphobbit you can talk as much as you want, but RTX 3050 hardware DOES NOT SUPPORT FP8. comma.
Well, probably bunch of my generated pictures on FP8 checkpoints using my 3050 are just materialized from the Matrix XD
@mphobbit maybe your IQ never materialized, as I just said before (read) your gpu uses an FP 16 instruction every FP8 request. it works, but very badly. same time of FP16 but with half precision, do you get the point, yes or not?
Fantastic Z-image model. Thank you for this!
I'm glad you like it :D thanks for the buzz .. the site did not show it lol
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.