Video to LoRA by https://oneshotlora.com
Source: https://www.youtube.com/@GinnyDi
Description
Video to LoRA by https://oneshotlora.com
Source: https://www.youtube.com/@GinnyDi
FAQ
Comments (6)
I'm having trouble getting this to run using the Flux checkpoint flux1-dev-bnb-nf4-v2, even though other Flux.1 D LORA work great with that checkpoint. When I try and use this one, it seems to completely ignore the presence of the LORA even when I have it tagged and the keyword is present.
Did you somehow grab the SDXL LoRA (55mb) instead of the FLUX one (164mb) by accident?
@loradude Nope, it's the 164mb FLUX LORA, though if I look at the metadata on it, it actually shows that it's for SD 1.5. Here's the properties I get when I look at its properties in Forge. I just downloaded the Flux one again to make sure the properties match what's on this page, and it's the same.
name: gjzzpr191a409d3e7u2d0a
ss_base_model_version: sd_1.5
ss_output_name: gjzzpr191a409d3e7u2d0a
sshs_legacy_hash: e3b0c442
sshs_model_hash: 8A2CA7A962F09979660C63F8FB1448CB0F55E8781750F2F645CDA010E43328C9
Whenever I generate using it just makes anime.
@loradude Ah disregard; I just had to crank the strength of the LORA way up in Forge for some reason. Like 2.8 seems to work fine for this one. Thanks!
@SD304 Thanks for the update. I've confirmed that the metadata says "sd_1.5" even though it definitely doesn't work with sd_1.5. CivitAI reports that it "fixed the hash" and that I should "redownload it" but (other than the hash, and for some reason it's converted from JSON to string) the metadata looks the same. @oneshotlora Any ideas?
@loradude Both the bad hash and the "sd_1.5" in the "ss_base_model_version" field are produced by ai-tookit, which we use under the hood for training FLUX models. We can confirm that CivitAI converts some of the JSON metadata fields to strings when "fixing" the hash, and we are looking into whether or not that trips up Forge with regards to LoRA strength. When generating images using the reference Python provided on HuggingFace, a LoRA strength of 1.0 seems to work fine.