Leapfusion Hunyuan I2V Patcher

Tencent's original Hunyuan Video release (December 2024) shipped without an official image-to-video model - text-to-video only, at a time when that gap mattered a lot, because Wan filled exactly that hole a few months later and largely took the local video ecosystem with it. Leapfusion Hunyuan I2V is the community's stopgap for that gap: a LoRA-style patch that gives the text-only-trained Hunyuan model image-conditioning ability it was never trained for out of the box. This node is how you apply it.

Why you'd want this

If you're working with Hunyuan Video specifically and want to animate from a starting image rather than generate from text alone, this is the patch that makes that possible on the original model. It's worth knowing the honest context, though: Hunyuan Video 1.5 (November 2025) eventually shipped with real, official image-to-video support, and by that point a lot of the reasons to stay on Hunyuan at all had already migrated to Wan. So treat this node as relevant specifically if you're on the original Hunyuan Video model and need I2V before or without moving to 1.5 or a different model family - not as the standard way anyone builds an image-to-video pipeline today.

How it works

It patches a loaded Hunyuan model, applying the Leapfusion I2V adjustment the same way a LoRA gets applied - strength-scaled, and only active within a chosen window of the denoising schedule - conditioned against a starting latent rather than pure text conditioning. The output is the same model object, patched, ready to drop into your normal Hunyuan sampling chain in place of the unpatched one.

The inputs and outputs that matter

• model - the Hunyuan model to patch (required).
• latent - the conditioning latent (typically your starting image, encoded) the patch conditions generation against.
• index (default 0, range -1 to 1000) - which frame/latent index the image conditioning applies at.
• start_percent (default 0) and end_percent (default 1) - the portion of the denoising schedule the patch is active over, same idea as a ControlNet or LoRA weight schedule: narrow this if you want the image conditioning to influence only part of the generation rather than the whole thing.
• strength (default 1, range -10 to 10) - how strongly the patch is applied. Negative values are technically allowed by the range, but start at the default and adjust from there rather than assuming a negative strength does something predictable.

One output: MODEL - the patched model, ready to sample normally downstream.

How to install it

• ComfyUI Manager - search KJNodes for ComfyUI, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-KJNodes, then pip install -r ComfyUI-KJNodes/requirements.txt (portable: python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-KJNodes\requirements.txt), then restart.

This node patches whatever Hunyuan model you already have loaded - you'll need the Hunyuan Video checkpoint itself set up separately (it's a large model; expect real VRAM requirements even before adding this patch on top), and the Leapfusion I2V patch weights themselves, which this node applies but doesn't bundle - check the pack's model/workflow references for where to source them if they're not already in your setup.

Common issues & troubleshooting

No visible image-conditioning effect. Check strength isn't accidentally left too low, and that start_percent/end_percent actually cover enough of the schedule for the conditioning to matter - a narrow window near the very end of denoising won't meaningfully steer the generation, since most of the structure is already decided by then.

Generation looks unrelated to your starting image. Confirm index points at the correct latent position for how your workflow is structured - if this is set wrong relative to how your latent batch is laid out, the conditioning is effectively being applied to the wrong frame.

Considering this for a new project. If you're not already committed to Hunyuan Video specifically, it's worth checking whether Hunyuan Video 1.5's native I2V support, or a Wan-based workflow, covers your case with less patching involved - this node is a real fix for a real historical gap, not necessarily the first tool to reach for on a fresh build today.