v2.0 (2025/12/12)
Based on user feedback for v1.0:
· Race is easily overridden to Asian female
· Body type prompt words are easily overridden
· Image quality degrades when loading other LoRAs at the same time
For v2.0, the dataset has been carefully captioned/tagged and retrained. The above issues should have been improved to a certain extent (looking forward to everyone's further feedback).
Trigger words: Reversal Film Gravure, analog film photography,
Recommended strength: 0.5
Recommended sampler & scheduler: euler_sgm_uniform
ps.
If the prompt only specifies 'a beautiful young woman', this LoRA may tend to generate Asian women. If you need to generate a specific race/ethnicity, please be sure to explicitly specify it in the prompt, such as 'A Nordic woman' or 'a white woman'.
v1.0 (2025/12/01)
This LoRA was trained using a dataset of 150 carefully selected images with the ai-toolkit. Training steps: 24,000.
This LoRA is designed to generate portraits in the style of reversal film gravure / slide film photography. Compared to the base model, it features increased overall contrast, more refined image quality, reduced noise and film grain, and smoother skin.
Description
This LoRA was trained using a dataset of 150 carefully selected images with the ai-toolkit. Training steps: 24,000.
This LoRA is designed to generate portraits in the style of reversal film gravure / slide film photography. Compared to the base model, it features increased overall contrast, more refined image quality, reduced noise and film grain, and smoother skin.
Trigger words: Reversal Film Gravure, analog film photography,
Recommended strength: 0.5
Recommended sampler & scheduler: euler_sgm_uniform
FAQ
Comments (15)
not just a photo style lora, also has an asian female character in it too
No, the dataset is not composed of a single specific character. In fact, every image in the dataset comes from a different person. You can think of the new face obtained after loading the LoRA as another kind of "averaged" face. As for Asian female characters, yes, I do have a preference for Asian women.
@AIImageStudio it's not about the images you select, it's about how you caption them, as i said in my reply on other comment.
Update the description. Only makes Asian women.
If you don't want Asian women to appear by default, please explicitly add something like 'a beautiful white young woman', 'white woman' to your prompt.
I like the style of composition from this LoRA, sadly it imposes a body style that isn't what I'm trying to achieve. Please focus on adding and tagging better training data if you make update! So far it thinks the bodies are part of the style which is an issue.
Thank you for your comment. LoRAs more or less carry the creator's personal subjective aesthetic preferences. The current version aligns with my own aesthetic taste, so I apologize that it may not meet your expectations. Considering the feedback received so far and the diversity and complexity of aesthetics, I do not plan to release any further updated versions.
@AIImageStudio all that would be needed would be to describe the body types accurately using 4-5 danbooru tags per training image caption and then the lora would learn that those body types depend on the tags and can be overridden while keeping the same aesthetic style.
ie: medium breasts, 28yo woman, asian
those 3 tags would go a long way.
I do really like the lora style, it's just not really useful as a style lora because it overrides all my other loras even on 0.1 strength. bit of a one trick pony.
@MetaGen OK. Thank you for your suggestion. I'll reorganize the captions and try training again.
@MetaGen After adding captions, the required VRAM has increased significantly, and my current GPU can no longer support continued training. So I'm very sorry, but for now, it's not possible to retrain the model in this way.
@AIImageStudio that's okay, thanks anyway!
@MetaGen a new version 2.0 added. You may try it, it should be better to solve your issue.
@AIImageStudio Hey thank for the update I'm coming back to you after testing it overnight, it is more permissive than it used to be when mixing with other loras and ethnicities, at weight ~0.25-0.30. It does tent to oversaturate the image a bit too much, but that can be put as artistic spin so it's not a big deal (although if you caption for it is should become controllable?), however one issue I noticed regularly is that it tends to introduce wrong number of toes on feet (almost every gen with bare feet has 6 toes, ranging from 4 to 7 toes), and sometimes wrong number of fingers on hands (less common), did you have ai training data or cropped feet in the mix perhaps? The Asian ethnicity still bleed a little but not every gen at least.
@MetaGen Thank you for your attention.
Regarding mixing with other LoRAs, I cannot predict the various combinations that different users might try. This can only be explored by the users themselves to find the parameter combinations that work in specific cases.
As for saturation, this is by design—an inherent feature of this LoRA, which is styled after reversal film stock. It is not a bug, but rather the intended characteristic I aimed for.
The toe issue rarely appears on my end.
In the end, what I want to say is that since this is a style LoRA, broad compatibility and universal applicability are not my goals, and I do not consider full compatibility with other LoRAs.
@AIImageStudio bad mindset tbh but ill leave it at that. over saturation is not the same as b&w curves, and mixing with other loras is irrelevant in my latest feedback other than i was preaching it worked better than v1. but i do think the training data has bad feet examples, and also i suspect the training data has too many images with blurred background, furthermore the "intended contrast" darkens even sunny days, making the environment look unrealistic and blurry. i'm not trying to attack you with this feedback only be helpful in case you consider making an update at some point, which you are 100% free not to do. that being said, since you posted the lora in a public space, I feel compelled to give my feedback, even if you don't update the lora you might learn some insights from it, and even if you don't, someone else might come around and decide to make a better lora of the same concept, which then my feedback will be helpful to them. please understand i'm not trying to criticize to be mean, i really enjoy gravure photography and just want to see the concept reach its best capabilities for ZIT. you don't have to feel compelled to explain your motivations to me, it's fine either way :)