MiniMax H3 Unified To Video (best-effort)

MiniMax H3's native ComfyUI support is spread across several nodes: EmptyMiniMaxH3LatentAV makes the latent, MiniMaxH3ImageToVideo handles keyframes, MiniMaxH3ReferenceToVideo handles references. Fine for a clean text-to-video, but the moment you want keyframes and reference images and a reference video with its own audio and a standalone audio clip, you're juggling three nodes and manually keeping tag numbering consistent.

This node collapses that whole stack into one: it builds the AV latent, tags every piece of media into a single <Picture N> / <Video N> / <Audio N> sequence, and hands you ready-to-sample positive conditioning plus the av_latent in one shot.

The essential inputs:

• clip and video_vae - your H3 text encoder and VAE.
• prompt - the H3 prompt text.
• mode - auto (use whatever's wired), keyframe (ignore all reference inputs), reference (ignore first/last frame), hybrid, or text (ignore all media).
• width, height, duration, fps - generation size and length; duration is snapped to H3's frame grid.
• ref_image_size - match scales reference images to the generation's own pixel area (cheaper); max uses H3's higher-fidelity 2048px-short-edge reference pipeline (slower, better).

Optional media inputs: first_frame / last_frame (anchored as literal keyframes at frame 0 and the final frame, like the native I2V node does), picture_1–4, video_1, video_audio_1, audio_1/audio_2 - and a references bundle (type H3_REFS) you can wire straight from the "Fantastic H3 Prompt Builder" pack's MiniMaxH3MediaLoader or MiniMaxH3PromptBuilder. Direct sockets win per-slot; the bundle fills in the rest, up to its own capacity (9 pictures, 3 videos, 3 paired video audios, 3 audios).

Outputs: positive (CONDITIONING), av_latent (LATENT), plus conditioned_prompt, media_map_json (a JSON summary of what was attached and how it was tagged, with a warnings list), and report for humans.

Provenance, straight up: this is a best-effort reimplementation of a node type from a privately-distributed H3 workflow with no public source. The frame/canvas/latent math is copied verbatim from ComfyUI's real native H3 nodes so shapes stay byte-for-byte compatible; the merging of keyframe and reference pathways into one unified tag sequence is inferred from a single saved prompt in the source workflow. The pack's model card rates it "medium" confidence. Treat it as a strong starting point, not a verified clone.

Two gotchas. First, the migration one: reference sockets were renamed from the old ref_image_0/ref_video_0 style to the 1-indexed picture_1/video_1/audio_1 names that match MiniMaxH3ReferenceSplitter. If you have an existing workflow, delete the node instance and place a fresh one - reloading in place won't pick up the new socket names cleanly. Second, a reference video longer than the requested output gets trimmed to fit, and both report and media_map_json tell you exactly how many frames were dropped - no silent truncation.

Install:

cd ComfyUI/custom_nodes
git clone https://github.com/jjmaden/comfyui-jjmaden-random-nodes

Restart ComfyUI, or search "JJMaden Random Nodes" in ComfyUI Manager. You need a current ComfyUI with native H3 support, plus the separately-installed siblings this pack doesn't ship (MiniMaxH3SigmaShift in core, the SageAttention patch in KJNodes, MiniMaxH3ReferenceSplitter in the Fantastic H3 Prompt Builder pack). No extra Python deps; the H3 weights are a separate large download. For the mixed-keyframe-plus-references workflow, this is the node that makes the graph readable again - one node instead of three, and one tag sequence instead of your best guess.