Style Transfer — MiniMax-H3
Restyles an entire video while keeping its motion, framing and timing intact. The source video enters as an aligned guide, so every frame of the output corresponds to the same frame of the input — only the look changes. Trigger: style_transfer:
Most video restyling is driven by words: you name a style and hope the model knows it. That breaks down for anything specific — a particular illustrator, a look you found on a moodboard, the palette of one film. This LoRA takes the style from an image instead, so what you can restyle to is whatever you can show it.
The instruction deliberately never names the style. It points at the reference and describes only what a painter would copy: linework, brushwork, palette, shading, texture, finish. Naming the artist would let the model solve the task from text alone and ignore the image entirely.
There are two ways to use it.
The first, and the main one, is with a style reference image. The reference goes in as an image reference, which H3 labels <Picture 1>. Refer to it by that tag and describe the look in three to five attributes:
style_transfer: Re-render this video with the style of <Picture 1>: bold graphic style, vibrant colours, clean lines, flat shading, halftone pattern.
Or drop the attributes when you want the image to do all the work:
style_transfer: Transfer the visual style of <Picture 1> onto this video.
The second is text only, with no reference image, for styles the base model already knows:
style_transfer: Make it cinematic style.
style_transfer: Change this video to studio ghibli style.
style_transfer: Transform it into art deco.
Do not mix the two. A style name and <Picture 1> in the same sentence makes the model follow the text and ignore the image.
Works at any H3-valid frame count (17n+5: 73, 107, 141, and so on). Rank 64.
One limitation worth knowing: a reference showing a person or character can bleed identity into the result, because the base model's prior treats an image reference as a subject to insert. Describing the style attributes in the prompt pushes against this, and a reference with no prominent figure avoids it entirely.
Description
FAQ
Comments (17)
Can you stop killing it for a second man, I need a pause, haha.
This stuff is just moving so fast, Sometimes I feel like I need to stop and catch my breath. Minimax Is an absolute game changer.
Wow this looks amazing. Does it transfer audio as well or just video?
only video
works really well. nice job
wtf ? this is way too interesting
Thank you 🪵🐸🪵
Great for preserving complex styles during animations
@NRDX Thanks for this, I like to understand what's happening in workflows. I'm curious about why the video including any audio, does not get plugged into the main node at all - instead, only used for the Set Guide.
I mainly want to know if you handle it this way based on some other research/workflow etc, and if you have more info on this. I can see you very carefully manage the positive conditioning for this use case.
Because this is a LoRa edit, to be used as a V2V, if you have a video with audio, you will only change the style and will not alter the audio in the end. In other words, it's simply adding the original audio over the final video instead of having to regenerate the audio. In the end, this will only take your original video, pass it through with guide latents (hence the Add Guide node), and modify the original video. It's not a reference to the model; it's the main guide.
@NRDX Thanks for the reply. I wasn't so much curious about the audio, mainly why the video input circumvents the "MiniMax H3 Reference to Video" node altogether (I'm assuming affects the positive). On my end, I tested both ways (1. video connected to the node + set as a guide; 2. video set as guide only) and the result was basically identical. BUT! I do see that when I connected the video into the reference node, it added 4 minutes to the generation time.
So I guess in a nutshell: only assets that will be prompted "as references" need to be connected into the "MiniMax H3 Reference to Video". Setting assets merely "as guides" saves a lot of compute, and clearly will yield same results so long as it is sufficient for the use case.
I'll need to try and find a writeup or something that explains all the situations its advantageous to simply set something as a guide only
@altoiddealer I wrote about this here; in the end, it's all very similar to how we do IC LoRa for LTX. That's where I got the idea from, along with some experiments by AbleJones. Ref2V is not the same as V2V. https://gist.github.com/alisson-anjos/b300f2b90e65cf85846519d78b660cd3
@altoiddealer The advantage is basically that if you use something as a reference, you are providing the model with a reference, as the name suggests. It won't always copy the reference video; the result might simply invent parts, skip parts of your reference video, and besides, the result is 100% tied to how you write your prompt, 100% prompt-dependent. Using a guide, the model follows more absolutely and without desync, even if the video has camera cuts, this will show up in the result; everything will be more synchronized.
@NRDX Awesome! Thanks for the pointer. I basically skipped from Wan 2.2 into MiniMax, so never got my hands dirty with the IC Loras.
I've been working every night on a workflow that I think is going to be a hit (coming soon); I had already made it very easy to set Guides for input. But after seeing this, it seems like an "is_ref" toggle for everything (to bypass inputting it into the Ref node) could actually be very useful.
Lastly! Just wanted to comment that your LoRA popped up literally after I tried (and failed) to do a style transfer. I had used Klein 9B to actually edit my first video frame into the style I wanted, and maybe my prompting sucked or something but the video almost immediately shifted into the source video.
Just used your LoRA today and it worked, although the style kind of degraded compared to my ref image. I solved it 100% perfectly by also replacing the first frame of the video - with your LoRA it maintained the exact style. It's high movement/motion blur involved too, I'm really excited about how easy it was with your LoRA
@NRDX Just figured I would double back and let you know that I did indeed finish up and publish that workflow I was working on. It's flexible enough to easily reproduce the specific Reference/Guide combination you use for style transfer with just clicking 1-2 boolean switches. I have a feeling you'll like it - https://civitai.com/models/2947601/simply-advanced-minimax-h3
do you have to have the input image as 4 seperate, similar images in the cubed form?
No, it's in the examples because I got the images for testing from here: https://github.com/willwulfken/MidJourney-Styles-and-Keywords-Reference/blob/main/Pages/MJ_V5/Style_Pages/Just_The_Style/Design_Styles.md, an old repository of MidJourney prompts.