A Little More Sharpness
This is not an upscaler. It does not change your video's resolution or size. It works on the detail already in the frame — enriching it, or reshaping it, depending on how hard you push it.
What it does
Trained mostly on real photography and video, with a strong focus on human subjects.
- On real footage — sharpens and enriches detail: skin, hair, fabric, edges.
- On anime, cartoon or stylized footage — pushes the subject toward a more humanized rendering (LTX 2.5 version).
Both behaviours come from the same mechanism. This LoRA is an extension of anime2real, which was built to turn stylized footage into realistic footage, so that pull toward realism is inherited. Feed it non-real input and expect some of it.
The name is modest on purpose, but be aware the effect is not always subtle — at higher strength it can do more than sharpen.
One of the main goals of this LoRA is to serve as a second pass for videos generated by any model, but especially Minimax H3 (LTX 2.5 version).
Strength is the control
This is the single knob that matters. Low strength leans toward enhancement, high strength leans toward transformation. There is no "correct" value — it depends on your source and the look you want. Start low and raise it until it lands.
Notes
Runs as a v2v pass over existing footage. Trained on human-centric content — results on non-human subjects are untested.
Description
Important note: for this LoRA to have any effect, it must be used with "Add Guide" rather than H3's native reference node; while you can try combining both nodes, it was primarily trained to use guide latents.
"Add Guide" is a native node, but it differs from their native reference-passing nodes; you could try combining the two. The issue is that this LoRA was trained for the ref2va model rather than fl2va. You might achieve different results—or get different reference inputs—by using "Add Guide" alongside "MiniMax H3 Reference to Video." In principle, this model is intended for a second pass rather than for direct generation, though I think it would work for that as well.
The included workflow also uses ostris' AI-Toolkit reference node — clone it into your ComfyUI custom_nodes folder: https://github.com/ostris/ComfyUI-AIToolkit-MiniMaxH3
Full details, weights, and workflow: https://huggingface.co/Alissonerdx/Minimax-H3-ComfyUI
FAQ
Comments (17)
I almost got too excited that I thought this was for Minimax H3.
But you can use it in minimax, hahaha—that's the goal.
I trained one for minimax but the problem is that it's much more expensive than ltx.
now it is
DLSS5?
hahaha no
AlisSS5
Is "Add Guide" node different from adding first/last frame in regular h3 conditioning node?
"Add Guide" is a native node, but it differs from their native reference-passing nodes; you could try combining the two. The issue is that this LoRA was trained for the ref2va model rather than fl2va. You might achieve different results—or get different reference inputs—by using "Add Guide" alongside "MiniMax H3 Reference to Video." In principle, this model is intended for a second pass rather than for direct generation, though I think it would work for that as well.
@NRDX I just meant I thought the Add Guide Node did the exact same thing as first/last frame inputs, just allowing any frame position, so I was wondering why you recommended this node specifically. Especially if it was trained on ref2va, doesn't that mean it was trained for reference inputs and not fixed latent keyframes? Anyway I'll just test and see.
@wewewew Yes, it was trained to use latent guidance directly rather than conventional references, after all, the name "reference" implies that the model might use the input fully or omit parts of it. The goal is a v2v workflow rather than ref2v, since the intention is to use this LoRA in a second pass rather than for initial video generation. Furthermore, it was trained on ref2v because that is the model variant I’m most interested in; what I’m attempting is essentially the same method used to train IC-LoRAs for LTX.
Hi! I wanted to try it out, but the workflow uses a custom node that is not mentioned anywhere and that I can't find - AIToolkitMiniMaxH3RefVideo. Where can we get this from? Thanks in advance!
Check details in the details section in this page
So for H3, this lora requires your workflow? Can it not be used in another workflow to sharpen video output?
You can but you need uses Add Guide node
fun fact:
used at negative value -0.3 / -0.5 improved movements in some scenes.
thanks for sharing