H3 Infinite - Continue from Latent v1.3

The v1.3 Continue node (H3ContinuousContinueV13) is the version that stopped being just "the thing that extends video" and became a proper directing tool. It kept the proven v1.2 phase-aligned direct AV-latent handover completely unchanged - same freeze handling, same stitching geometry - and bolted on what the suite was missing: auto-growing Qwen References for telling H3 who's on screen and what they're wearing, plus a picture_map output that tells you exactly which Picture slot each input landed in.

Why that matters: H3 is omni-modal, and its Qwen vision encoder reads reference images as <Picture N> slots. But if you can't see the mapping, you're flying blind - is the Last Frame Picture 1 or Picture 2? This node makes the ordering explicit, and it's deterministic: Previous masked/duplicated AV context is not a Picture at all, Last Frame is Picture 1, then each connected Qwen Reference takes the next slot. picture_map prints the real mapping every run, and it's the output you'll check when a character looks wrong and you need to know whether H3 actually saw your reference.

How it works

Same skeleton as v1.2: previous_latent in, carried history out the front of a new target. handover_mode AUTO reads the analyzer metadata, alignment_mode stays on phase_aligned_extended, duration is in seconds and snaps up to the 17k+5 grid. The new bits:

• qwen_reference_1 (optional) - Qwen text/vision reference. Connect it and the node auto-grows qwen_reference_2, then 3, up to nine.
• picture_map output - the authoritative Picture-to-input mapping.
• last_frame - the recommended endpoint/quality reset anchor.

One important honesty note the author is upfront about: Qwen References are Qwen-only. They are not inserted into minimax_refs as native Ref2VA/DiT reference latents. So they steer identity and detail through the vision encoder, not through the persistent-reference mechanism some other H3 tooling uses.

Install

From the Herrgotts-H3-Infinite-Continuation-Suite pack:

cd ComfyUI/custom_nodes
git clone https://github.com/HerrgottMargott/Herrgotts-H3-Infinite-Continuation-Suite.git

Restart ComfyUI, reload the UI. No heavy pip deps, no bundled weights - H3 checkpoints, the video/audio VAEs, and Qwen 3 VL come from Comfy-Org's MiniMax-H3 repo. Note that Qwen References genuinely need the Qwen 3 VL model loaded, which is the biggest single weight in the setup (the ~32B nvfp4 AWQ build is the usual choice).

Gotchas

The direct-latent head is still duplicated into the new clip, so actual_head_context_frames must reach the stitcher or you'll see doubled motion at the seam. And remember the Qwen-reference limitation above: if you're expecting Ref2VA-style persistent reference behavior, this won't give it to you - it's vision guidance, not a native reference latent. v1.3 is a great middle ground: v1.4's masked continuation is the shiny new thing, but v1.3's guide-based path remains registered specifically so you can A/B the two.