Helps create HD images of Melissa Benoist
Description
Improved consistency with lesser weightage.
FAQ
Comments (20)
No likeness whatsoever. This is the third LoRA in the last week that has failed to work at all. I have 2,000 Flux LoRAs, and I test every one, and yours simply does not work. Included in the screenshots are complete generation data, so you can see my settings, prompt, and model.
https://imgur.com/a/f5iHjyx
Thank you for your feedback. Let me look into this. I did get good likeness while generating images. Anyways i will test it out and re-train the data. Thank you
Which are the other LoRas that were not working for you?
@AmateurAiArtist - in my discussion with another user about their Audrey Tatou model here:
https://civitai.com/models/978280/audrey-tautou-flux-actress
This one also does not work:
https://civitai.com/models/1030933/erika-eleniak-flux-actress
I will quote you some of the discussion, which I didn't fully understand, because it's about fp8 versus fp16, and NF4? I have never used anything but base flux1-dev and my clip, encoder, and VAE are all standard and work fine. I have been using Forge with Flux since it came out and only very recently - like the last two weeks - have I encountered Flux LoRAs that I couldn't use with normal settings. In the case with Audrey, I tried several different base models and a wide variety of settings for low bits and such, all to no avail, and the downloads alone took me many many hours on my throttled hotspot.
Here is some of what they said:
> The only quantization scheme available is NF4 for OneTrainer, that's the problem. A lot of people is waiting for the update to a FP8, so this problem won't happen. For now, using this config should work out (you prob need to change the CLIP as well): https://imgur.com/a/via3cwH (this link appears to be dead now, sorry)
> the think is, at least of what I've read and seek out the discord
> weight matters
> bfloat16, float16, float8
> this is precision for LoRA
> my bet it's on this
> yeap, it's probably because of the optimizer that i'm using
there's more but there's some stuff to think about.
If you want to, I mean.
I'm just sad I can't use your LoRA :(
@leisure_suit_larry thanks for explaining this. I have retrained my LoRa. I will uploading the updated version soon and hopefully this can yield a better result for you. I tried using the same prompt as you did in the examples and I do see better results now. Do let me know your feedback once I upload it.
It's updated. Also try using a prompt which is more specific and with a natural language. Just a suggestion.
@AmateurAiArtist - I will have to wait to try it tomorrow, as I've just begun a training and am headed to bed. Thanks for your efforts, much appreciated.
@leisure_suit_larry sure thing.
@AmateurAiArtist - V2 works. What did you do differently?
@leisure_suit_larry changed the dataset and training parameters :)
@AmateurAiArtist - cool. Thanks a lot. I would really like to know what settings are causing the incompatibility, though, as there are other models that are problematic. My loras made with fluxgym seem to work no problem. Is it about the precision? FP8 or FP16? What base model did you use? Just curious.
@leisure_suit_larry I think its mostly about the dataset quality over anything else. And also, I started using more repeats than more epochs.. previously used to make 5-10 epochs. But Now I try to make like 3 to 4 epochs and make sure the repeats are between 5-7. That has worked well for me.
@AmateurAiArtist - Ok, thanks. I think there is actually a compatibility problem with either FP8 models, using NF4 during training, or something to do with onetrainer. I've encountered several flux loras now that work fine for others but do not work for me with Forge and base Flux1-dev for some reason. Cheers and thanks for the models.
@leisure_suit_larry Yeah I mostly just use the Flux1 dev.
Much better likeness with v2. Whatever you did with this you should consider revisiting some of your existing loras and reapplying that change to!
Yeah will definitely do that. Do let me know whichever LoRas you feel would need an update. I made this update based on the feedback from one of the user. So it always helps :)
hi may I know what sampler,schedule type and model you are using? Can't seem to get your results
I use Euler and normal.. and also the base flux dev model..
I do add merges with mid journey loras on most of them but then I always have to crank up the weight for my character lora.. probably like 1.2 and all.. and I mostly keep the step counts low.. 15-20.. even when the image size is large
@AmateurAiArtist I see.. hmm can't seem to get the same result.