gamma:
Broke the dih barrier for the most part, it's quite challenging to maintain the quality of beta and get accurate nsfw by merging LoRAs. It always seems to do well in one area and then botch something else. BUT, I have to give special mention to shootthesound and his custom LoRA loaders (https://github.com/shootthesound/comfyUI-Realtime-Lora), these have been a game changer for me with mixing models in with LoRAs I trained on AI Toolkit (thank you @ostris). Shoot's node allows you to dial in layers for z-image and even give you an analysis of what layers have more information and are stronger in influence. Check the nodes out and this youtube, give him some love! He's working on the second version of these very useful nodes and I know I for one and looking forward to them.
That being said, give gamma a shot and as always tell me what you love or hate about it. I trained several separate LoRAs focused mainly on male anatomy, then with male and female in the same image, but no NSFW activities and then some focused on NSFW concepts. It still isn't perfect, but it's a step closer and that's all that matters to me. Thank you all for the support and kind words. I do this as a hobby and still learning a lot as I go, your input is appreciated and valued. Cheers!
Recommended settings:
dpmpp_2s_ancestral + beta
5-8 steps
***the fp16 and fp8 versions are under the same model tab, just select the drop down to the right and select the version you'd like***
alpha:
Has a slightly different vibe from beta or at least a small step toward something different, hopefully better - try it out for yourselves, some gens will definitely need a 2nd pass/hi-res fix. Using odd resolutions seem to result it body deformities (extra hands, limbs, etc.). Still working on male anatomy, as well has some female features that are missing or lack in detail and accuracy. Look to your favorite LoRAs for Ds and Ps for now.
Recommended settings:
dpmpp_2s_ancestral + beta (detail daemon with split sigmas on the initial generation is very nice in the workflow)
5-8 steps
Hi-res fix:
1.3x latent upscale
heunpp2 + beta
4 steps
Upscale:
Ultimate SD Up or SeedVR2
USDU: ddim + ddim_uniform, 0.2 denoise, 4 steps
ps: some of the showcase images use prompts from recent community posts, I wasn't thinking about it at the time to bookmark ones I used. So, if you see a prompt that looks familiar, by all means, message me and I will link the original post on the image.
beta:
My impatience is growing waiting on a fine-tunable base model, so here is a beta I've been working on since the release of the amazing Z-Image Turbo. A huge thank you @ostris for his amazing work with the AI Toolkit and of course the Tongyi-MAI team for all the mind blowing work they do and the pace at which they do it. Kudos to you all!
Recommended settings:
The following all work great, some better with certain generations.
Euler + simple or SGM_Uniform
DDIM + DDIM_Uniform
dpmpp_2s_ancestral + linear_quadratic (my personal favorite)
All LoRAs merged with the model were trained by myself on AI Toolkit with my own personally curated dataset, most focusing on detail, anatomy, skin, lighting. I will release some lora's in the next week or so that are geared toward better anatomy for both male and female. Some merges didn't look great after they were baked in, so I left them out for now. Obviously it's a WIP and I appreciate any and all constructive feedback and how it can be improved. Cheers <3
Description
FAQ
Comments (11)
Impressive result! Any chance you could share the datasets sizes you used to train the LORAs? Thank you for your work
I appreciate that and no I don't mind sharing. On the first couple versions, I used ~500 images to train 10+ different LoRAs, some specifically for anatomy, some for tonal qualities (color, lighting styles, etc) and some for photography aspects (ISO levels, camera types, compositions). Of those 500 images, I tried not using anything less than 1024x1024 and no larger than 4096x4096. I have some that are smaller (512, 768) and lower quality/amateurish photo with grain and artifacts, but I haven't started injecting those yet. I plan on doing that for the next version
@DigitalAF Thank you! I tried full fine tune of Illustrious and noobai using Oneteainer with ~50k photo dataset I curated. Spent 8 months experimenting with no particular success. Now thinking if should change my approach and train Loras instead to merge them to models. Thanks!
@wind11 I would much rather be fine-tuning with Kohya_ss Dreambooth, but not sure when that's going to be possible without the base model. Even if they would support the De-Destilled model, would probably be better/more stable training that using LoRAs. I've used it with the Flux.1 on about 1,500 images and got decent results. It's definitely comes down to quality over quantity and making sure your training prompts are consistent between each other and across concepts/styles. My full dataset is around 35k images and I've narrowed it down to 1,500 that seem to make a good impact
fp8 please?
It's already up, there are two files you can download under the model info
@DigitalAF Thanks! I didn't notice...
Zero ability to generate pubic hair, least from my testing.
Not asking for it to be fixed, not saying its a bad model, this model is fantastic and I love it, but there was no pubes to be had.
No, that's great info and not surprising considering my dataset has a lot of shaved or very little hair on the woman. I appreciate that input a lot! Something I can add to the list for sure. Thank you!
Issue is lora trained with ostris doesn't work well with these models, the faces are very inaccurate.
use
https://github.com/kohya-ss/musubi-tuner.git and this model like a base for training with turbo adapterDetails
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.