Wan 2.2/2.1... but it's Bad Apple style!!
This t2v LoRA was trained on the legendary black and white silhouette animation Bad Apple for 42 epochs (28 for W2.2), capturing its unique shadow-play aesthetic.
The examples are raw output, no thresholding, segmentation or whatever postprocessing applied!
The trigger word is "bad apple style, black and white silhouette animation.".
Some example prompts:
bad apple style, black and white silhouette animation. This animation showcases a white side-view silhouette of an anime girl шт kimono sitting in lotus pose under sakura. The petals are falling off cinematically. The background is solid black.bad apple style, black and white silhouette animation. This animation showcases a Dart Vader playing a violin. The background is solid black.The default settings for Wan2.1 are cfg from 6 to 8, and flow shift of 5. I used kijai nodes for these examples generation.
For Wan2.2 use cfg of 3.5, default flow shift 8.
Limitations: may struggle with dynamic action prompts. Unprocessed VACE references often add color, however, you can absolutely bring your own silhouette using segmentation models - like "Anime Remove Background" on Huggingface. (example VACE workflow included with one of the videos)
Looking forward to your generations :)
Description
Update of my legendary Bad Apple!! LoRA for Wan2.2 T2V 14B. The prompting is the same, but this one was harder to train and the lora is slightly unstable (the movement can be weird), so inference on lower resolution (~512x512) is advised. It's black and white only, you can upscale with nearest neighbors :) As always, no postprocessing was applied to get the iconic two-colors look. Example prompts are here.
FAQ
Comments (8)
Would it be possible for you to upload the safetensors files directly?
Hello! Wan2.2 loras consist of two files: low noise and high noise, and they have to be supplied in .zip bundles, because Civitai doesn't allow to put two files of the same type as the download in the upload form. Probably, for the API convenience.
If you do not upload the safetensors directly, no application can retrieve metadata about the files. The vast majority of models on civitai are links to safetensors files, and so API calls can get info about the models. When I unzip your zip and try to use the LoRAs in comfy, I want to see the triggers and sample prompts, but none of my software can identify your LoRAs, because the hash does not yield any results. If the end point for your upload is a zip and you do not upload safetensors files, essentially all software designed to catalogue and organize LoRAs using civitai's API is broken.
I have trained over 100 Wan 2.2 LoRAs... I know exactly what they are and how they work. However, I use musubi-tuner so I only create a single LoRA that works in both high and low.
You may notice that virtually everyone else is simply uploading two separate safetensors files.
@leisure_suit_larry Alright, I will look into that (by the way, if you have the experience of uploading both loras, you can dm me), thank you for this explanation note!
What software do you use? If it's open-source, maybe, it can be patched to index zips?
@kabachuha - civitai removed all of my loras in one fell swoop in May and I don't feel comfortable uploading here anymore. People use the versioning system to upload separate loras for Wan 2.2 I believe. As for the software, it's literally all of them. There is no other automated way to query the internet for information about loras, it's all about civitai.
Indexing zips is not a logical solution. There is metadata available for zips, but no one is loading zips into their LoRA managers and none of the many, many scripts that rely on API metadata are written with non-model filetypes in mind.
This is the scenario.
I download your zip. I unzip it. I move it to my lora collection.
Then, later, when I launch comfy, I launch swarm and lora manager to use as references for lora information. Or in Forge it's civitai helper (or something to that effect, there are many extensions for auto1111-based GUIs that query civitai's API for metadata about safetensors files).
Swarm can not find a URL that matches your hash. Lora manager can't. Civitai helper can't.
It doesn't matter what software I use, no app can get any information about the safetensors files via the civitai API unless the safetensors files are uploaded to civitai.
Here is a visual to help you grasp what is happening:
https://pixeldrain.com/u/2EGzG5dc
This is easily automated, and most apps are automated. I used this manual process to show you the dynamics.
You and I can not change the way civitai works and there is literally no other way to automate metadata retrieval.
The ComfyUI Lora Manager has a feature where you can relink a civitai url to a model when you right click it, which has worked so far for the ones it didn't get, as well the other zip uploads I've manually extracted and put in my folders, but for some odd reason this one gives a "Error: Failed to re-link model: Not Found" message...
@kabachuha Many people upload the low and high noise as separate versions on the same page.
As others have said, when you do it this way we can't get metadata properly which makes it a huge PITA to set them up inside of local tools.
@Latent_Dreamscape Yes, thank you! I got the two versions system and will publish my future releases this more correct way