H3 Infinite - Flexible Start / Conditioning v1.4

If you build a new H3 long-video project in 2026, this is the Clip 1 node you should be dropping on the canvas. H3ContinuousStartV14 is the v1.4 "Flexible Start / Conditioning" node, and its job is unchanged from v1.3: one node that does T2VA, I2VA, L2VA, or FL2VA starts with optional First/Last Frames and auto-growing Qwen References. The interesting part is what it doesn't do - the v1.4 Masked AV magic lives in the Continue node, not here. This Start is deliberately boring, because a chain's first clip shouldn't be where the experiment lives.

What you get

• Optional first_frame / last_frame - connect neither, either, or both. Full FL2VA (both) is still the recommended long-chain pattern because repeated Last Frames are your quality resets.
• qwen_reference_1 (autogrows to nine) - Qwen text/vision references for identity and details. Picture ordering is deterministic: First Frame → Picture 1, Last Frame → Picture 2, Qwen refs after. The picture_map output shows the real mapping each run.
• duration in seconds, snapped up to H3's 17k+5 frame grid (10.0 s → 243 frames ≈ 10.125 s).
• Outputs positive, latent, picture_map.

Same caveat as v1.3, because it's easy to miss in the v1.4 excitement: Qwen References are Qwen-only. They steer through the vision encoder but are not native Ref2VA/DiT reference latents - if you need persistent reference behavior, that's not what this is.

The v1.4 context

The full v1.4 story (native masked continuation, freeze-safe shared boundary, Net New Content duration, audio-tail carryover) all happens in H3ContinuousContinueV14 and H3ContinuousAnalyzeHandoverV14. This Start node feeds Clip 1 into that machine, and the included Herrgotts_H3_Infinite_v1.4_01_Start.json workflow shows the recommended wiring: Start → sample → Save AV Latent (clip 1), then Continue v1.4 takes over for Clip 2+. And because v1.4 requires ComfyUI with PR #15375's native H3 AV-mask support, the whole chain - Start included - expects a current ComfyUI build. Start itself doesn't probe the mask capability, but your Continue nodes will.

Install

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

or ComfyUI Manager → pack search. Restart, reload. No pip deps, no bundled weights - H3 FL2VA transformer, video/audio VAEs, and Qwen 3 VL come from Comfy-Org's MiniMax-H3 repo, and the Qwen 3 VL weight is effectively required since the Qwen Reference feature is core.

Gotchas

Keep First + Last connected for long chains unless you're explicitly testing keyframe-free continuation (the less-validated path). And when a seam looks wrong later, don't debug this node - check that the same actual_head_context_frames value is wired from Continue into the join node. Clip 1 being solid isn't the hard part; the handover math downstream is.