Join my Discord for help, workflows and news of new LoRAs!
🥊 Aether Punch – Face Impact LoRA for Wan 2.2 5B (i2v)
Aether Punch is a custom LoRA trained for powerful face punches — a single boxing glove or fist slamming into the subject’s face from the left side.
It was trained specifically for image-to-video (i2v) workflows using Wan 2.2 5B, and expects a starting image that defines the subject clearly.
Best results are on human-like subjects, but it seems to work on other things too!
đź’Ą Trigger Phrase
Use this exact prompt to trigger the effect:
A single boxing glove appears from the left, punching the face🧠How to Use (i2v)
Start with a 768Ă—768 image (might work with lower res)
– Should feature a clearly visible subject, preferably a person
– Front-facing or 3/4 angle works bestPrompt using the trigger phrase above
– Keep prompt minimal
– The LoRA tells the model to animate a left-to-right punch across the face
⚙️ Inference Settings
• Mode: image-to-video (i2v)
• Model: Wan 2.2 5B
• Resolution: 768×768 (has been reported to work at 512x512 as well)
• FPS: 24
• Steps: 20
• CFG: 5
• Clip length: 5 seconds (121 frames)
âś… Best Practices
• Use a portrait-style image of a person (centered, well-lit)
• Avoid clutter — punch effect is strongest with clear subject
• Don’t include boxing gloves in the image itself — the animation adds them
Thanks to @bblink787 and @masslevel for creating some gold videos here!
Trained using Ostris's AI Toolkit.
Description
FAQ
Comments (20)
This is nice! are you planning on expanding this to the 14b? I have not tried this yet but does it do a hand without a glove? if not, would it be possible to add it in the next versions if there is any?
Hey! Thanks for the engagement!
I have yet to first try it base 14b already can do this (maybe you have already tried it?). If it can’t I might train for it. To me this is just a fun little concept that I wanted to try out.
And concerning other than gloves: I think it can be done with hand/fist. Try it out if you have the possibility! I’ve seen guys hitting zombies with meat using this so it seem a bit flexible.
what wan lora compatible with each other ? wan2.2 work with wan 2.1 and versa ? and about i2v and t2v which of them work for both i2v and t2v ?
I don’t know, man. I only finetune and haven’t tested out compatibility between 2.1 and 2.2.
I always get the following error with this Lora:
ERROR lora diffusion_model.blocks.18.cross_attn.q.weight shape '[5120, 5120]' is invalid for input of size 9437184
my source image is 640x640
Any suggestions? Not having issues with any other lora
Sorry no idea. It’s a 5b LoRA that works with i2v. Haven’t heard anyone else with issues. Trained with Ostris’s Ai-toolkit
Got the same error when i used Wan 2.2 14B model together with this Lora but i still got a (unsatisfying) result ... well now i am using Wan 2.2 ti2v 5B and get: WanVideoModelLoader - 'WanVideoModel' object has no attribute 'diffusion_model' .. ah well, i don't know enough about AI to get out of this mess.
is it possible to have a fist only no boxinggloves?
Just it in prompt! I’ve seen successful results of it from my beta testers. 5b seems to be a very capable base model 🙂
joachim_s ok tried it and its cool, just a quick question tho, is it possible to control the speed of the punch?
minanimator28 Glad you got some progress! About the speed: I think I know what you mean - you refer to the speed up until the punch? The thing is that didn’t really find clips to train on that has a rapid movement before the impact, so mostly the whole movement will be in slowmo. But but you can try and add common terms like rapid or fast and see how that goes. I will need to get back to you on this. Also: experiment with length of clip. Shorter than 5 secs might help. Not sure.
minanimator28 The owl example among my videos has a bit more rapid movement before impact. Try that prompt as inspiration. Might also be the aspect ratio that affects it or it’s just that seed:
A bright red boxing glove bursts into frame from the left, crashing into an owl. The owl's eyes close quickly. Closed eyes.
joachim_s thank you for the insights! enjoying this lora so far lol, now tryna make the tooth fly out lol! thanks! more great loras in the future!
minanimator28Â Haha. That I surely want to see! Do lmk how that goes. I bet it could be a challenge.
Join my humble discord server if you’d like and discuss further more easily: https://discord.gg/77ah3Sfy
Can also always look me up on discord by the same name and image as here.
Hey man, love the effect, I've been trying this with no luck. Would please share your WF please?
Yes ofc! Join my discord server and I can send one later today:
Hi, this is really amazing work, I got great results with different subjects. Let me know, I will be very grateful if you share information. How did you collect the dataset for training such lora? Is it synthetic data generated on veo3 or natural scenes, and how many such videos did you have to collect? Thank you.
Feel free to connect with me on my Discord server. The link is in the description to this LoRA 🙂
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.