News

Lyvia September update: Flux-Lyvia 4, model updates, unlimited queues and more

Lyvia

— 2 min read
Lyvia September update: Flux-Lyvia 4, model updates, unlimited queues and more

Hi all! There's many more features being added and improvements going on but half of them are not finished yet, so I'll just post what's already available on the app for now.

Model Updates

While the model names stays the same to avoid confusion, the models are consistently updated and trained with new data. Here they are:

Lyvia 4

Welcome Lyvia 4, our in-house model trained on Flux!

If Lyvia 4 didn't show on your models list, please make sure you have the latest version.

Somehow this model takes like 2-3 minutes to load for the first time. Currently tweaking the servers, see if I can make it load faster. This is only the first time load, after that it should be faster, like 8-10 seconds per image.

Lyvia 3

More detail, more prompt adherence, trained on more data.

Lyvia Pony

More pose & action understanding.

Lyvia X

Cross-training with Lyvia 3 and Pony, I'd say in the middle of both, but you gotta try it yourself.

Other Models

  • Lyvia Anime is kept back because people love using it and there is no alternative yet.
  • I have to retire legacy models due to lower image resolution, slower generation and large resource use. Please support the new models instead by using the models and giving feedbacks so I know how to improve them to your liking.

Now you can queue videoswap jobs!

Previously, you are limited to just one queued video faceswap job. Now you can queue as many as you want, as long as you have the credits to use.

Max duration for each video is 5 minutes/300 seconds. This limit will be increased for Pro users.

Better img2img

Now your initial image is being followed more using generation. Play with denoising strength, while describing the whole final image, not just what you want it to be. The higher the denoising strength is, the more the image will change. I'd say 0.4-0.6 is the best.

What I'm working on

  • Selective upscale: Since AI is so random that maybe only half of the generated images actually match your liking, it makes sense to only upscale the images you actually like. This reduces generation time so you can generate more images, then pick images you want to upscale.
  • Download 4x upscale: When you download an image, it will be upscaled to very high resolution (4096x4096).
  • Upscaler selector: You can also choose the upscaler. The UltraSharp is good for photo details, while R-ESRGAN is good and fast if you don't want grainy images.

I'm aiming these to be released before the end of this month.

Chat with me anytime

If you have questions, bug reports, suggestions, feedbacks or even just to say hi, feel free chat with me anytime at Reddit chat or Telegram. You can also email me at admin[at]cendra.co.

Thank you and happy generating!

Read more