How to Use
This is a style LoRA, designed to be the base layer of your LoRA stack. It creates the foundational aesthetic of realism, upon which you can add character or concept LoRAs.
Note: The included ZIP archive contains both the high-noise and low-noise LoRA variants, along with our recommended ComfyUI workflows.
Trigger Word:
InstacamRecommended Strength:
1.0. Start here and adjust in small increments.
I would also like to thank Danrisi, who originally taught us how to train LoRAs and helped make our work possible.
Description
FAQ
Comments (19)
Can this also be used for text-to-image?
yes
t2i ***is*** t2v
It was specifically made for text to image, not trained on any videos
@00quebec The images are looking great! I am struggling using Wan for t2i, getting bad dithering patterns (e.g. 80s big hair perm, blond curly hair, dark background). How can i avoid these dithering patterns?
ok, after some more tests I found out, that if you are using turbo loras it's important to set shift=1.0 and use euler not lcm
00quebec That doesn't matter unless you are trying to train a motion or action. This is a T2V lora.
playtime_ai WAN T2V Lora training question: So if I were to train on an /image dataset/ only (not videos) of someone, say a celeb, I would be able to use that /character/ lora trained on WAN and accurately have the person face when using T2I?
Follow up question, would it be possible to produce a video trained on images in WAN? Especially this lora: Instagirls. Probably not so, eh? Not sure how it works still.
ProvenFlawless To answer your first question, yes. You can absolutely train a lora of a person or character on images only and use that lora to create picture of said person or character (but don't post loras or images of real people to Civitai. It isn't allowed.)
To answer your second question, yes... you absolutely can use a lora trained on only images to create a video. You only need to train on video if you are trying to train a motion or an action... like a dance move or something like that. You can train many things... a person's likeness, an outfit, a body type... on just pictures.
Looks good! Danrisi is a legend, has helped me out a bunch too. Anyways I suspect you don't have diversity of ethnicities in your dataset?
I am
working on that for gen 2, first dataset only had a few different people.
How do you keep the character consistent ? Does this works with character loras ?
Updated so quickly?,,
New one releasing in a few hrs
Can generally complete epochs in 4-8 hours depending on how big the dataset is, etc. Doesn't have to indicate anything
So this is mainly for T2i?
Yes he mentioned this in a reply to another user 5734... bunch of numbers
Although someone mentioned t2v IS t2i, just a bunch of frames of it.
Gamerman420 I asked this because you can train WAN off images or short clips. If they're short clips, it leans towards some specific motion (e.g BJ, sex lora, tittydrop etc)
It's great thanks
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.