TARA node keeps MiniMax H3's motion reference native and lines it up in time with the output
TARA is a new, experimental ComfyUI node in the free ComfyUI-BFSNodes pack that changes how MiniMax H3 copies motion from a driving video. Instead of injecting the clip as guide latents, it feeds it as a normal video reference and nudges each output moment to look at the matching moment of the guide.
What it does
The node sits on the MODEL path after your LoRA loader and works with the MiniMax H3 Reference to Video node: the driving clip goes in as <Video 1>, its timing is aligned to the output clock, and attention to guide frames far away in time is softly penalised. It adds no trained weights; four controls (free window, falloff, strength, cap) set the rule, and the Shot Planner offers the same thing as a guide_mode option.
Results and limits
The author's six-column comparison shows an existing step-10500 motion-transfer LoRA trained for Add Guide following the source poses and framing under TARA while keeping the reference character's clothes, with all runs on the same seed; the docs call this an experimental observation, not a benchmark, and the requested background swap stays partial. ComfyUI Wiki reports a user's RTX 5090 run at 544×928 and 360 frames, and a Banodoco user running it on an RTX 4090 and an H100.
Worth knowing
It is opt-in and experimental: the Planner default is still Add Guide, and visual-quality validation is incomplete. Guide and output must both be 24 fps with matching frame windows (107 frames is a valid count), and you should not stack Add Guide, first-frame guides or another temporal patch on the same model. There is no matching native-temporal LoRA to download yet. The node pack is GPLv3.