MiniMax H3 Full-Chain Latent Video Adapter

MiniMax H3 Full-Chain Latent Video Adapter is the bridge between "I have a finished H3 chain" and "I want to upscale the whole thing with SeedVR2." It takes a selected checkpoint lineage, decodes every scene from the original H3 video latents, trims the repeated context overlap, applies the boundary blends, and writes one lossless, audio-bearing, file-backed VIDEO. Then the whole-video upscaler (SeedVR2) chunks a continuous timeline instead of restarting at every scene boundary.

The key word there is file-backed. A long chain decoded to one IMAGE tensor would be a RAM disaster - this adapter specifically avoids it. The full frame sequence never becomes a single ComfyUI IMAGE tensor; it streams scene by scene into a movie file. That design choice is the difference between a tool that works on a 5-scene production and one that OOMs on scene two.

The inputs

• manifest - a selected generated lineage from Checkpoint Manager. Partial runs are supported, and no source Plan is required. That last bit is deliberate: you can upscale a branch without dragging the whole generation graph back.
• video_vae - the original H3 video VAE. Every checkpoint latent is decoded directly; the saved H.264 scene files are not used. Same VAE caveat as PNG export: mismatch equals color drift.
• audio_source - plan (automatic: generated uses checkpointed H3 sound, source recovers the track), or explicit source/generated/none.
• blend_schedule - boundary crossfades applied before SeedVR2. plan uses each saved incoming-scene value; a list like 5,30 overrides the boundaries.
• decode_buffer - disk-backed (default) lets MiniMax's temporal decoder write each scene into a temporary mmap, so RAM holds only boundary frames. memory is a compatibility fallback holding one decoded scene at a time. Neither holds the whole production.
• reuse_cache - default true. The continuous source is content-addressed (cache key includes the checkpoint lineage, VAE implementation, blend schedule, prelude and audio policy) and cached at output/h3_chains/<run>/upscaled/seedvr2/source/<cache_key>.mkv. Reuse the cache and identical lineages skip re-decoding entirely.
• audio_bitrate - AAC bitrate embedded in the lossless RGB movie so SeedVR2 preserves it (default 256).

The outputs

video - one native file-backed continuous VIDEO for the SeedVR2 Direct Video Upscaler, already containing the chain audio. video_path - the cached movie's absolute path. status - cache/decode mode, scene count, frame count, blends and output path.

The recommended finishing graph

The docs are explicit:

Checkpoint Manager.selected_manifest
        + original H3 video VAE
        ↓
Full-Chain Latent Video Adapter
        ↓ native, file-backed VIDEO
SeedVR2 Direct Video Upscaler
        ↓
core Save Video

Use the audio-preserving SeedVR2 Direct Video Upscaler from ethanfel/ComfyUI-SeedVR2_VideoUpscaler; it reads the movie in chunks and returns a file-backed H.264 VIDEO with the audio embedded.

Installing it

Part of ComfyUI-MiniMaxH3-Contex-Loop (ComfyUI Manager: "MiniMax H3 Contex Loop"), or:

cd ComfyUI/custom_nodes && git clone https://github.com/ethanfel/ComfyUI-MiniMaxH3-Contex-Loop.git

Restart ComfyUI. It doesn't bundle SeedVR2 - that's a separate install (ethanfel/ComfyUI-SeedVR2_VideoUpscaler). Standing requirements: H3 weights, current ComfyUI with native Add Guide (PR #15439), and ffmpeg for the movie muxing.

Common issues

If RAM still spikes, check decode_buffer is actually on disk-backed - it's the whole point. If you change the lineage or blend schedule and nothing re-decodes, that's the reuse_cache fingerprint working; disable it to force a fresh decode. And same VAE warning as everywhere else in this pack: the adapter decodes latents, so the VAE must be the original one.