Basic Refmod creator for Minimax H3
Combine and encode your REF2VA audio/image/video references into a single .safetensor for use with MMH3 Director Workflows
Nodes: search for "Minimax H3 Refmod" in your manager/extensions
Description
FAQ
Comments (16)
no better creating page to creating minimaxh3 for free?
not sure I follow your question
@FirstPrinciples mean is not try creating page to generate video ai for minimaxh3
I’ve seen many people on Reddit recommending this LoRA type that bundles various reference materials. Does this mean that when creating such a LoRA, the source files must all be of the same type—such as bundling only images, only videos, or only audio (even if the specific formats differ, like mixing AVI and MP4 or PNG and JPG)—rather than bundling a folder containing a mix of different media types (like audio, video, and images all together) into a single LoRA?
Well refmods aren't really Lora's, although if that helps conceptually. Normally when you do REF2VA, you would add your image and audio references directly to the timeline, then the workflow encodes them as it goes. What the refmod does is combine all those images and audio into a single package that can be used for a reference in the timeline, and since they are pre-encoded not only does it save file size, but it saves processing time too since they dont need to be encoded at runtime.
Think of them like little character packages where (yeah, conceptually similar to a LoRA) you would have like a front, back, quarter, clothed, nude, expressions, voice sample etc packaged into a single .safetensors that you can then drop in your REF2VA timeline - that way instead of something like <image 1> <image 2> <image 3>, etc, you can just tie the <Subject> to <Refmod 1> which contains everything. The source files don't need to be the same type or even size, you can mix the characters voice samples with different image types - anything that you would normally drop into a REF2VA timeline you can drop into the refmod instead.
Hope that helps.
@FirstPrinciples Thanks for the reply. Yesterday, I tried packaging an image of a person and some video footage (two 7-second MP4 clips) into a single RMOD, and I noticed it was automatically named <VIDEO_1>. How do I specify in the prompt that H3 should use this RMOD? Should I simply write something like <video_1> walking towards the camera?
Do you also recommend using one dedicated RMOD per character or object? If I were to mix different subjects (e.g., a soccer star and a basketball star) into a single "sports star" RMOD, how would I phrase the prompt so that H3 can distinguish between the different assets contained within that mixed RMOD?
@wyxzddsjj919 Anytime you package in a video it will add everything to the video section of the timeline, but the images are still there for you to reference - in your example with a sports star with multiple outfits you could do something like:
<Subject 1> is a (describe the person like tall althetic male with x hair and eyes whatever) fully referenced in <Refmod 1>
<Subject 2> is a white basketball outfit with the number 7 on the jersey fully referenced in <Refmod 1>
<Subject 3> is a red and blue "Richmond" soccer jersey fully referenced in <Refmod 1>
Then the prompt detail would say something like:
At 00:00.000, <Subject 1> is fully clothed in <Subject 2>, he turns to the camera and says <d>[ENGLISH] But I love soccer too! </d>. At 00:04.000, <Subject 1> snaps his fingers and in a poof of white smoke, <Subject 2> is replaced by <Subject 3>
Obv your prompt would be better but I hope you get the idea - you can reference anything in the refmod or use different <Subjects> tied to different parts of the refmod to build your prompt
Is this limited to just 6 images? can you do it from a folder with a bunch of images?
The standard REF2VA limits apply:
Images - Max 9
Videos - Max 3, 2-15 seconds per, 15 seconds combined total
Audio - Max 3, 2-15 seconds per, 15 seconds combined total
Max total references 12
So you could do 9 images, 3 video or 6 images 3 video 3 audio... etc etc.
For some reason I can't get the audio ref to work no matter how I prompt it.
Try tying the references together with the <Subject> tag, like:
<Subject 1> is the X defined by <Refmod 1>
<Refmod 1> provides the audio and visual identity for <Subject 1>
<Audio 1> provides the voice timbre for <Subject 1> as defined by <Refmod 1>
Then in retention, reinforce
<Audio 1>: reference - timbre and delivery for <Subject 1> are followed closely.
Then in the prompt reference <Audio 1> like
<Subject 1> says in the voice and timbre of <Audio 1> <d>[ENGLISH] Hello, world! </d>
So to clarify from what ive learned about refmods: There a similar to loras, but I don't have to save a 1g safetensor. Instead I can just save a tiny package of references(images, audio, and videos) But thoroughly labeled on what's what? Or do i have that wrong? I'm new to making minimax loras. Could you correct me if im wrong?
More or less correct, think about it like packaging your REF2VA timeline references in a pre-compiled .safetensor. So instead of dropping a few images and audio and video into your timeline and processing it all at generation, you bundle them up and process them in the refmod, then add that to the timeline instead.
Referencing them depends on your workflow, I recommend Dasiwa's nodes since you can just add it and reference it directly with <Refmod #>. If you include character sheets you can use number or name labels to help tie it together, or you can use positional references
@FirstPrinciples What's ref2VA timeline? Also so let's say i use your creator to make one. All i have to do is collect and label what maybe 10 images, 20 clips, and a few audio files. Label them, and drag it to comfy for the creating, and use one of Dasiwa's workflows to get the ref mod working?
@Geno ref2va timeline on the MMH3 director node - yes you can use dasiwa's workflow and there will just be a refmod button you can add it with.
The max references is 12 combined items between 9 images, 3 video and 3 audio. You couldn't do 10 images and 20 clips as that would exceed the limit, but you also shouldnt need that many references either.
Hello, I’m getting an error when trying to create a refmod with video; could someone help me?