/vt/ - Virtual Youtubers

>>43502196
You're quite welcome! Sorry for being a cranky bitch last night! told ya I was sharper during the day

HEAR YE, HEAR YE. If any of you need the "unpruned" models from the huggingface that have EMA weights still in them no fucking idea why you would unless you were thinking about training on top of a model merge like a true dumbass. or you like wasting space grab them NOW.

Now that >we know fp16 models will autocast/upcast and work fine for just about everyone, I'm gonna PRUUUUUUUUUUUUUUUNE. I'm probably not going to do everything, especially not in one go, since anything I replace will need to have Colab code amended, and might need rentry - entries fixed too, as well as new hashes calculated.

It's nice that hf gib all that free space, but there's no point in wasting it. EMA weights in a merged model aren't really necessary, and only seem to be necessary for further training. And using fp16 models for Colab will mean faster loading times. Storing them yourself locally means saved space.

I'll pipe up again once something is commit -> push'd.

>>43492499
>Here's the mega: https://mega.nz/folder/rpVHlI5a#-gNPnXuCM3he75xL6embpQ
I think you went to sleep, but thank you, once again. Wanted to give you a heads up that I renamed them - kind of.

- AOM3_hll1.safetensors [911416BEED] Abyss Orange Mix 3 - AOM3 + HLL1 AD 1.0
- AOM3_hll31.safetensors [41BAF1466A] AOM3 + HLL3.1 AD 1.0
- HoloOmelette.safetensors [FCD30E3299] (AOM3 + HLL1 AD1.0) + (7th_anime_v2_A + HLL1 AD1.0) WS0.5
- ScrambledHolos.safetensors [2219AB65B4] (AOM3 + HLL3.1 AD1.0) + (7th_anime_v2_A + HLL3.1 AD1.0) WS0.5

>pruned
redundant, we only had the teeny bit of EMA models in HLL to begin with, and I think having it in column B actually throws the EMA weights out, because I've never had anything to prune when merging HLL with a pruned model.
>AOM mixes
AOM3_hll1 instead of hll1-aom3
I say we let silly names die for basic, AD mixes...props to you for making me realize that