Nodes/MiniMax H3 Contex Loop/MiniMax H3 SelfLift Tiling — Experimental
ComfyUI Node

MiniMax H3 SelfLift Tiling — Experimental

Experimental high-resolution SelfLift tiling. Connect to highres_tiling on Chain SelfLift Sampler or SelfLift Seed Hunt. Off/unconnected preserves existing behavior. Tiles do not share attention; only the first tile's audio prediction is kept. May introduce seams or change motion/audio quality. ControlNet and regional conditioning are unsupported.

By ethanfel·Created 2 months ago·Updated about 21 hours ago· 463
MiniMax H3 SelfLift Tiling — Experimental
    • tiling
    ◄enabledfalse►
    ◄tiles2►
    ◄overlap8►
    ◄axis▾►
    Categorysampling/minimax/context_loop

    Inputs (4)

    NameTypeDefaultDescription
    enabledBOOLEANfalseOpt in to spatial tiling of SelfLift's final high-resolution denoising steps only. Not TST, VAE tiling or tiling of the learned latent lift.
    tilesINT22–8Fixed tile count along one spatial axis. More tiles reduce attention workspace but lose cross-tile context and add overhead. Small grids use fewer tiles.
    overlapINT80–64Context margin at each tile edge, in latent pixels (normally 16 output pixels each). Adjacent margins overlap and blend. Clamped on small tiles.
    axisCOMBOSplit along the longest spatial dimension, or choose width/height. Every tile retains the full video timeline and audio.

    Outputs (1)

    NameTypeDescription
    tilingH3_SELFLIFT_TILING—