The latest version workflow and nodes are here
https://github.com/peterkickasspeter-civit/ComfyUI-ZImageTurboProgressiveLockedUpscale/tree/experiment
=====
Custom node: https://github.com/peterkickasspeter-civit/ComfyUI-ZImageTurboProgressiveLockedUpscale
Workflow is in the showcase images. More details here - https://www.reddit.com/r/StableDiffusion/comments/1r1qyj2/zimageturboprogressivelockedupscale_works_with_z/
Distill lora for base - https://huggingface.co/alibaba-pai/Z-Image-Fun-Lora-Distill/tree/main
*use either Z-Image-Fun-Lora-Distill-8-Steps.safetensors or Z-Image-Fun-Lora-Distill-4-Steps-2602-ComfyUI.safetensors
Description
test version. i am experimenting with a new node that i have been working on these last few days. the lora has huge dark skin bias unless prompted for something else. i also signed up for this civitai creators program so i set the version to early access to support my runpod costs. i do need some donations. only download if you want to donate. thanks
if you download it, you can just use it with the previous version workflow or basically any workflow. results sometimes are better than with the node i am working on and the showcased images (the node still has bugs and i am testing so i cant post it yet)
you can use "raw photo candid" in your prompt but i don't think it helps so much. worth a try
FAQ
Comments (13)
You're like a Deedeemegadoodo wise wizard of Tongyi-MAI/Z-Image-Turbo/Base.
Have you ever thought of working on workflows, nodes and LORAs for a really great anime model called circlestone-labs/Anima It's also a model that's made in collaboration with Comfy Org. It's still in preview but it's a very very good model with Incredible prompt adherence. And uses a very small but capable text encoder called Qwen3 0.6b.
😆 anime is not my area of expertise. I can't tell if my lora is generating good anime images or not. I don't even know from where I can find good anime training data
The LoRA is quite nice, but it has a major bias towards ... prominent (putting it nicely) noses. Can be somewhat prompted away, but, like its tendency to make more realistic bodytypes, it doesn't really disappear. Not a plus or minus, just a thing. The real con, is that it literally cannot do freckles, or any sort of skin detail beyond a (fairly nice) kind of microtexture.
this lora also makes people fat, and tits big and face fat
freckles look fake ai to me so i dont have much in my training dataset. i have nightmares about it after looking at flux krea. noses i dont know, they look alright to me. its just a massive dataset so maybe i need to clean it up a bit
@rahulsharma133579137 are you the one who is commenting on my reddit posts too? just use another lora if you dont like it. i dont want to generate flat chests
@peterkickasspeter Flux's freckles were trauma inducing, to be sure. ZIT and Klein do fairly well. Niche audience either way I suppose. Personally this is the only ZIT LoRA I've used that doesn't actively ruin image quality, so I'm happy for that.
@peterkickasspeter hey peter, yes i was the one in reddit too, anyway, I really want that realism without big tits. so I want to train my own LoRA can you give me any suggestions on how to do it. your orthogonal subspace projection node and python I will use. I just want to train on flat chest images. please help
hey bro, i see this same lora is working for both ZIB and ZIT, i trained a lora for ZIB using ostris AI toolkit and that lora is not working for ZIT. how to create lora that works for ZIB and ZIT...
I don't know if you have any experience training Flux Klein 9b but if you have time please could you make a saggy breasts LORA as it's a very decent model. Maybe even even using the very good [KLEIN 9b] Unchained XXX V2 LORA by the great alcaitiff, as a base as it's a great NSFW LORA and improves the anatomy too. Great work as always.
The model is awesome however I am having problem to add a lora from an specific character, I trained a lora model with zturbo but it seems its not working. do I need to train a model for zbase and zturbo or what is the correct process? thanks
I stumbled upon your message by accident. Aside from the fact that you're using this lora and the question is general about character lora, I'm actually training my character's lora and using it with ZimageBase and turbo at the same time. I trained it on base version and can use on both models, but trained on turbo version cause a lot of noise and dont work as i want if i generate in bbase, maybe u can achieve normal quality using different samplers and denoisers. I`m using mixed workflow generating first half on base and second half on turbo model and using same base trained character lora and it works perfectly. Using character lora trained on base model in turbo im my case generates more "asian" appearance. Generating fully on base works good aswell as mixed, but takes more time. I also use UltimateSDUpscale to add more quality and realism
train for only zbase, then apply on turbo (distillation = loras bad 99% of the time)