Nodes/ComfyUI-MiniMaxH3-Timed-References/Add Timed Video Reference for MiniMax H3 (Experimental)
ComfyUI Node

Add Timed Video Reference for MiniMax H3 (Experimental)

Adds a Qwen-only video reference starting at a target time. Core VIDEO is streamed by real presentation timestamps; IMAGE batches remain available as a constant-FPS fallback.

By ethanfel·Created a day ago·Updated about 10 hours ago· 7
Add Timed Video Reference for MiniMax H3 (Experimental)
  • clip
  • video_frames
  • source_video
  • CLIP
prompt_tagmotion
target_start_seconds0.000
source_fps24.000
video_size512
analysis_fps2.0
Categorymodel/conditioning/minimax

Inputs (8)

NameTypeDefaultDescription
clipCLIP
prompt_tagSTRINGmotionUse this name with a leading # in the Ref2VA prompt, for example #motion.
target_start_secondsFLOAT0.0000–9999Where this reference clip begins on the generated video's timeline. Sub-second values are preserved.
source_fpsFLOAT24.0000.01–240Used only by the IMAGE-batch fallback. It must match the loaded frame rate; core VIDEO reads timing automatically.
video_sizeCOMBO512Approximate pixel-area edge used for Qwen vision encoding. 64 is the technical floor but loses detail.
analysis_fpsFLOAT2.00.5–240Frames per second presented to Qwen. 2 matches core H3; higher rates cost more tokens and memory. Do not exceed the source rate.
video_framesoptIMAGELegacy/VHS fallback: a decoded IMAGE batch with constant frame spacing. Connect either this or source_video, not both.
source_videooptVIDEOPreferred: connect core Load Video. The source stays path-backed and only selected frames become tensors; real presentation timestamps are preserved.

Outputs (1)

NameTypeDescription
CLIPCLIP