ComfyUI-MiniMax-H3-Extend

ComfyUI-MiniMax-H3-Extend

Backports MiniMax H3 video-extend (continuation) support onto stock/public ComfyUI installs. The native MiniMaxH3VideoExtend node only exists on kat3ri/ComfyUI's feat/minimax-h3-video-extend branch (or later) -- confirmed 2026-08-11 against a real "clean" master-branch ComfyUI that continuation isn't a missing node so much as a missing capability in the DiT model's own positional-encoding code, which stock's PackedLayout explicitly rejects:

if pixel_index == 0:
    cond_t = ...
elif frame_count is not None and pixel_index == frame_count - 1:
    cond_t = ...
else:
    raise ValueError("only first/last keyframe anchors are supported")

What this does

Monkey-patches (only applied if the native class is genuinely absent -- never overrides a real fork's better native support):

1. comfy.ldm.minimax.model.PackedLayout.__init__ -- adds handling for keyframe dicts carrying kind="context" (trailing video latent frames of a prior clip, placed at negative RoPE-time positions counting backward from the target's own frame 0) and kind="context_audio" (the audio counterpart). Also generalizes keyframe anchoring from a hardcoded text_len origin to one that accounts for ref_images/etc. being present in the same call -- stock's own two node types never combine keyframes and refs, so it never needed this; when refs is empty, the new origin equals the old one exactly, so stock's existing call patterns are unaffected bit-for-bit.

2. comfy.model_base.MiniMaxH3.extra_conds -- fixes a real bug this surfaced: stock overwrites cond_video_latents/cond_audio_latents from refs instead of appending to what keyframes already contributed. Harmless for stock's own node types (which never combine the two), but silently drops the context frames' actual latent data the moment both are present in one call -- which continuation-with-references does, by design.

3. MiniMaxH3VideoExtendPatched node (nodes.py) -- a standalone, directly-usable node exposing the above, plus injection of MiniMaxH3VideoExtend into comfy_extras.nodes_minimax_h3's own namespace (again, only if absent) so anything that looks it up by that name there -- e.g. ComfyUI-H3-Cast's H3CastToVideoExtend -- finds a working implementation transparently, no changes of its own needed.

4. MiniMaxH3EncodeAVPatched node (nodes.py) -- vae-encodes video frames (+ optional audio) into the AV latent MiniMaxH3VideoExtendPatched's (or the native node's) context_latent input needs, for feeding an externally-loaded prior clip (e.g. VHS_LoadVideo) into either. Vendored from the native MiniMaxH3EncodeAV, which isn't part of stock/public support either. Lives here rather than in ComfyUI-H3-Cast since it's an extend/continuation concern, not a cast/character one -- also injected into the native namespace the same way as MiniMaxH3VideoExtend above.

Confirmed working via live testing (2026-08-11) across text-to-video, reference-to-video, and cast-to-video continuation. Recommended starting settings: context_frames 2, ref_spacing 1-2, ref_decay 0.3, ref_ramp 3-4 (5-6 if the prior clip had more motion than usual).

Install

Same as any custom node pack -- clone into custom_nodes/, restart ComfyUI. No core ComfyUI files are edited; everything is monkey-patched in memory at import time. To remove: delete this folder, restart.