MiniMax H3 Video Extend (Backported)

The 5-second problem, solved sideways

MiniMax H3 gives you excellent 4–15 second clips with native audio, but "excellent 15 seconds" still isn't a scene. Continuation - feeding the tail of one clip back in and generating the next chunk off it - is the standard way people get longer takes out of video models, and stock ComfyUI simply doesn't have it for H3. Not as a missing node, as a missing capability: ComfyUI's PackedLayout positional-encoding code hard-rejects any keyframe anchor that isn't frame 0 or the last frame, raising ValueError("only first/last keyframe anchors are supported"). You literally cannot express "continue from this clip's ending" on a stock install.

That's the gap this pack exists to close. MiniMaxH3VideoExtendPatched is the backported, directly-draggable version of MiniMaxH3VideoExtend, which only ships on kat3ri/ComfyUI's fork. It takes a prior clip's AV latent plus a prompt, and hands back the conditioning + latent you feed to the sampler to produce the next segment. One trap to head off: this node doesn't sample - it emits positive and an empty latent for the sampler. Drag it in alone and nothing happens.

How it works

At import time the pack monkey-patches two places in ComfyUI's H3 code - PackedLayout.__init__ and MiniMaxH3.extra_conds - and only if the native extend class is genuinely absent, so it never fights a real fork.

• The trailing latent frames of context_latent become context keyframes, placed at negative RoPE-time positions counting backward from the new clip's own frame 0 (audio frames land on their own negative track). This is the part stock rejects.
• extra_conds gets a real bug fix: stock overwrites cond_video_latents from references instead of appending to what keyframes contributed, silently dropping context whenever both are present. The patched version appends - that's what makes extend-with-references work at all.
• pin_last_frame decodes context_latent's true trailing pixel frame and pins it as this call's frame 0. That matters because context_frames only carries whole latent frames, each spanning 1–4 pixel frames - without the pin, the continuation can re-play a moment that already happened.

The canvas is inherited automatically (latent dims × 16), so you can't mismatch resolution between clips.

The inputs that matter

Only a few of these will you actually touch on a first run:

• context_latent - the AV latent of the prior generation to continue from. On stock ComfyUI that comes from MiniMaxH3EncodeAVPatched, which is how you feed in an externally loaded clip via VHS_LoadVideo.
• length - frame count at 24 fps for the continuation only, excluding context. Default 124 (~5s); note the step of 17, H3's temporal token cycle.
• context_frames - how many trailing latent frames of the prior clip carry over. Default 2. This is the one knob from the README's recommended set you can actually reach here - the richer ref_spacing/ref_decay/ref_ramp tuning from the README belongs to the fork's native node, not this backport.
• pin_last_frame - on by default; disable only if you're feeding an exact first_frame.
• first_frame / last_frame - exact images to pin frame 0 and the segment's final frame. last_frame is this pack's own addition (the native fork has no end-anchor at all) - the "land on the next shot" lever.
• ref_images - reference images as a single batched socket (use ImageBatch for more than one); each frame becomes its own <Picture i> reference. Not the native node's autogrow slots.
• audio_vae + ref_audio - for reference audio; connect audio_vae or you'll get a ValueError.

Outputs: positive (CONDITIONING) and latent (LATENT). Decode with VAEDecode + VAEDecodeAudio, save with SaveVideo/CreateVideo.

Install

One clone, same as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/kat3ri/ComfyUI-MiniMax-H3-Extend
# restart ComfyUI

Or grab ComfyUI-MiniMax-H3-Extend from ComfyUI Manager. No extra pip deps, nothing edits core ComfyUI - everything is monkey-patched in memory, so uninstalling is deleting the folder and restarting. You still need the H3 stack (MiniMaxAI/MiniMax-H3, ~42.5 GB, plus the H3 video VAE, audio VAE, and CLIP) and a ComfyUI that already supports H3. And the licence caveat applies regardless of install method: the H3 Community License excludes the US, EU, UK and South Korea from running the local weights.

Common issues

• You still see only first/last keyframe anchors are supported. That's the exact error this pack exists to fix. If it shows, the patch didn't apply - you're on a ComfyUI with native H3 extend already (the pack self-skips) or your ComfyUI is too old for H3 at all.
• MiniMax H3 supports batch size 1. Don't batch context_latent; batch ref_images instead - that's what it's for.
• The continuation replays a moment from the old clip. Usually pin_last_frame was switched off. Keep it on; it decodes the true trailing pixel frame.
• It's slow or OOMs. H3 is 33B - context and references cost real VRAM on top of the base load. Drop context_frames, trim ref_images, and be honest about your card.

Confirmed working by the author across text-to-video, reference-to-video and cast-to-video continuation - but this is a brand-new, one-person backport of subtle positional-encoding math, so treat your first render as a smoke test. Start at context_frames 2 and add references one at a time.