Created by GeomanticArts and please consider supporting the trainer at geoloras.com
Be guided to the best generations, thanks Tingyun!
Since I don't have Tingyun, I'll just have to make a LoRA for her. :_;
Example images all made with AOM2-7thV3_C_40 at weight 1
'Tingyun [Honkai Star Rail], fox tail, chinese clothes' - Example prompt for the normal outfit. It will regularly get the tail without needing to prompt for it, but it's definitely less reliable than you'd expect (since it's in 100% of the images..?)
(collar:1.1) in the negatives if you're getting the sticky ribbon collar
LoRA metadata has all the tags used.
The trainer's twitter can be found at https://twitter.com/GeomanticArts
Trained on NAI.
Description
FAQ
Comments (8)
I've used no less than 6 Tingyun LoRAs and this is the most flexible one yet. You can actually prompt different clothes.
Probably doing Loopback training to achieve this (making some nude/alt outfits with SD and then feeding those images back into another LoRA ver), as there really aren't many images of Tingyun that are nude or different clothes to train with.
Another solid LoRA by Geomantic.
Usually that kind of loopback training is avoided at all costs. When I've personally tried it, it tends to get some weird artefacting and extremely bizarre lighting/shading on the outputs. The earliest versions of Serval were done this way in an attempt to fix that very rigid clothing (because of her extreme lack of other outfits in available art), and honestly those trainings were terrible. Eventually the whole dataset was trashed and started fresh, which is the current version that's up right now.
Edit to add: There is 1 generated image in this dataset of 62 images
I've had great results with loopback training the time I did it as an experiment, but I took my time (fixing errors in the generations, doing a wide variety of checkpoints/art styles/angles/lighting/etc) - I think if you take the time and do it right, it's very useful for characters or concepts with a limited dataset
Can you provide the prompts and information so that we can copy and paste your generations? I'm trying to replicate it but to no avail
Well that's not great. I would think the LoRA works without needing to copy and paste. The LoRAs should all be easily usable without just copying, I already know my prompts work after all, not much is gained by just copying what I already did. It's making sure they work for everyone else which is the difficulty..
Which image are you trying to recreate? The first image prompt is
<lora:TingyunHonkaiStarRail4:1> highres, detailed, soft lighting, outdoors, forest, chinese clothes, tingyun [honkai star rail], 1girl, hand on hip
neg:
(simple background:1.2), (monochrome), multiple views, multiple angles, (letterbox:1.1), (lowres:1.2), (out of frame), (blurry, bokeh:1.4), (text), (worst quality), (low quality), (normal quality), jpeg artifacts, (signature), (watermark:1.2), (username), cropped
Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 3647769207, Size: 448x704, Model hash: 3e4fb056e3, Model: AOM2-7thV3_C_40-Pruned, Denoising strength: 0.6, Clip skip: 2, ENSD: 1234, Version: v1.2.0, Hires upscale: 1.5, Hires upscaler: Latent (antialiased)
@RockRockRock Thank you, actually I've managed to make it work but I didn't see your comment until now. Thanks!!
This model is great!
But im getting a random shirt on her arms and detached handsleeves in the final image. Anything I can do to avoid these?
Yeah part of the problem with virtually all training images being the same outfit, you can get various sticky clothing pieces. 'Sleeveless' should work 'bare arms' also works. 'Sleeves' in the negative should work too.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.