Check out Big Love ZT3 at https://civarchive.com/models/897413?modelVersionId=2695250
Big Love Z is a fine-tune of Z-Image-Turbo on a subset of the Big Love dataset. It produces a more photorealistic look, more diverse faces and better skin (with the right prompt and seed). You can run it locally with ComfyUI and Forge Neo as well as online on Tensor Art and Civitai (if you bid for it). The fp8 version produces more noise thus may sometimes look more realistic, but details are not as good. Most of the showcase images have a ComfyUI txt2img workflow embedded. The rest offers Forge Neo metadata.
Settings:
8 steps, cfg1, Euler sampler, Normal or Simple scheduler, 832x1216, 1024x1536 or 1280x1920
Download (on the right):
With 12+ GB VRAM: bf16 version
With 8 GB VRAM: fp8 Version
Place it in the Diffusion Models (ComfyUI) or Stable Diffusion (Forge Neo) sub folder.
Additionally you need:
With 12+ GB VRAM: qwen3_4b.safetensors
With 8 GB VRAM: qwen3-4B-fp8.safetensors
With any VRAM: ae.saftendors
Place them in the text_encoder and VAE sub folders.
Description
Big Love ZT1 is a fine-tune of Z-Image-Turbo on a subset of the Big Love dataset. It produces a more photorealistic look, more diverse faces and better skin (with the right prompt and seed). Still experimental.
FAQ
Comments (40)
Looks good
cant use it with my character lora :(
Still early days. Still experimental.
dont work with loras man
Working well! :)
Does not work with any Lora, using them with this causes the image to degrade.
Still early days. Still experimental.
This is true of any checkpoint that undergoes a lot of training.
Not worth to finetune the turbo. Every model is worse, than the base. Wait for full model.
Yep. Current methods are limited.
Partially you are right.
Oh nice seeing BigLove dipping into Z-Image! Hopefully -Base comes out soon so you can properly finetune! c:
Thank you! It's working fine with 12GB dedicated vram only for the image generation and 32GB system ram with the bf16 model.
fp8 should work even faster, but quality is lower.
Ok, Jib recommends bf16 for 12 GB VRAM. So I changed it in the description too.
90% of the generations limbs are messed-up
You must be doing something wrong. Never got that with my tests.
@SubtleShader sorry, will try more
Fix your workflow
The fastest model
Should not be faster than Z-Image-Turbo
solid start on it's own but LoRA compatibility is wrecked.
Will come
I've had pretty good success with LoRa with this, but you have to turn them down about 1/4 - 1/3 the recommended value... ie: 0.10 - 0.25 range. At that range, they act like they would at 0.75+ in other models. Remember, this generation is extremely different from other models -- steps, cfg, etc, so LoRa values are touchy, too.
@herewegoagainfucker oh yeah for sure, but my character LoRAs work perfectly with the base ZT model at 1. Can't stack other loras with them, but that might be due to LoRA training being drastically different as well and that not being common knowledge yet and everyone creating LoRAs how they always have previously.
Character LoRAs from my testing work far better with only one token as a caption (like '1girl' instead of detailed captions), so I wonder if that works with concept LoRAs as well somehow. But super detailed captions that describe everything in the images tend to cause them to be way overcooked with ZT, hence having to drop the strength way down.
So I'm not sure how people are creating new base models at the moment for ZT, I havent found a single checkpoint yet that is truly an improvement over the base model yet and almost all of them completely ruin character LoRA compatibility. So I think it's a matter of waiting on someone to figure out model training differences the same way LoRA training is completely different now as well.
@TiffanyHeat I get much better results from this model than the zit base model.
Improved Big Love Z Output
I'm working on a new sampler (for Forge & ComfyUI) with grain control that will allow you to add a much texture as you like during image generation. Skin looks much better with it.
I look forward to seeing the result!
@thaddeusk Thanks for reminding me!
Excellent model and an instant favorite. I'm running this on ComfyUI and I use all the recommended resources (qwen3 in CLIP and the included VAE). I have had slightly better results with Euler A than regular Euler (fewer problems with limbs and hands). Also, I have had slightly better results with the Simple scheduler than the Normal scheduler, Other than that, the recommended settings yield amazing results. Great work!
PS: Also, I am using the full bf16 version and the Qwen3-4b un-quantized, but using the two lower res work a bit better. I only have 8GB VRAM, but 32GB system RAM and a decent proc, so the offloading means I wait about 30 seconds for each image, but it's significantly better output.
For some reason, character LORA doesn't work with Zimage custom checkpoints.
Yep. Will improve once Z base is here.
@SubtleShader Then it would be cool if you released detailed instructions for learning LORA on ZiB, because I've tried 1000 times already and it turns out crap
@zagzagem685 I did not train anything on base yet, but this article might be helpful: https://civitai.com/articles/26358/z-image-base-training-is-solved-mostly-here-is-the-solution
Really excited to see the evolution of this. Obviously we're limited by the distilled nature of Turbo, but when those base weights are released I can see this becoming the #1 Z-Image fine-tune.
Amazing
Just like all the other fine tunes, despite your claim, they including this all lose the realism and skin texture of the original model unfortunately
Check out ZT2 at https://civitai.com/models/897413?modelVersionId=2580484 It is more advanced.
I actually do not like the original Z look that much. Skin texture might look more clean than ZT in various cases but it is blotchy. No comparison to Big Love Photo and Ultra. Yes, ZT has softer skin in various cases, but I still prefer the look and the faces. Besides, genitals in the original Z are really bad compared to ZT2.
So far, my work on ZT has rather been low effort compared to what I did for Photo & Ultra. I did not even get started. Was just testing the waters. ZT3 will be a lot better.
@SubtleShader But why does every single fine tune of yours have some weird sharpeness filter it looks very unrealistic
@MonkeyForever No sharpness added to ZT1 & ZT2. It really depends on what settings and tool you use if it looks more or less realistic. With my settings in Forge it looks better.