About this LoRA

I would say this LoRA is more like a quality LoRA, something like (masterpiece, very aesthetic, etc). By default, it generates anime screenscap style.

It's trained on MOSS SFT using my personal curated dataset of anime screenshots at multires. The captions use a mix of (NL, tags, and BBox), along with custom dropout.

This was also my seventh attempt. I almost gave up and went back to Anima, lol. The LoRA is pretty stable now. You might even think I'm using an anime fine-tune of MOSS Image x)

Prompting Tips

A quick tip: Try mastering the prompting. MOSS Image is pretty strict about what you input. The LoRA mainly provides the style, so don't expect it to easily generate things like 1girl, standing, good pose on its own.

I provide metadata in some previews so you can see the prompts and settings I used.

Camera angles and shot types can make a big difference. Try prompts like medium shot, long shot, wide angle, dutch angle, low angle, and high angle.

You can also experiment with filters and lighting effects, such as chromatic aberration, film grain, cinematic, backlighting, rim lighting, soft light, particle lighting, and any color grading.

Unlike Anima, MOSS SFT tends to generate less lively anime images aka boring. You'll need to work harder on your prompts to get good results.

If a pose is too difficult to describe with prompts alone, I recommend trying BBox T2I. You can also try my simple workflow.

Recommended Settings

I recommend finding your own sweet spot for steps, CFG, and LoRA weight.

Using a Turbo LoRA might affect image quality, but not using one can lead to issues with hands. If you decide to use a Turbo LoRA, you'll probably need to experiment with different seeds quite a bit.

MOSS SFT doesn't have a strong aesthetic of its own and tends to generate boring poses more often than not, so finding the right combination of settings and seeds can take some trial and error.

Character Accuracy

For characters, I've noticed that you might also need to include the title of the series they're from to improve accuracy.

For example:

Aqua (Konosuba), Kono Subarashii Sekai ni Shukufuku wo!

This seems to help quite a bit.

I've only tested this with anime characters, though, so I'm not sure how well it works with characters from gacha games.

Using This LoRA with Other LoRAs

Stacking this LoRA with other LoRAs might help you get better results!

In my previews, some images use a stack of LoRAs, such as Chainsaw Man LoRA and Official Art Games LoRA.

For example, using weights like (1.0) + (Chainsaw Man: 0.3) might help you achieve a more cinematic look and better anatomy with the help of the Chainsaw Man LoRA.

Of course, the results will depend on the LoRAs you use and their weights, so feel free to experiment.

A Note on Usage

Please don't merge this LoRA with other LoRAs or checkpoints and publish the result as your own.

I might allow training with this LoRA as context rather than merging it, though.