H3 Infinite - Flexible Start / Conditioning v1.3

The v1.3 Start node (H3ContinuousStartV13) is the answer to a complaint you'll hear constantly about H3 tooling: "why does my Clip 1 need to be FL2VA when I just want text-to-video, or just an image-to-video?" This node kills that friction. First Frame and Last Frame both became optional - so the same node does T2VA (neither connected), I2VA (first only), L2VA (last only), or the full FL2VA (both). That's the "Flexible Start / Conditioning" in the display name, and it means one Clip-1 node in your template instead of four nearly identical ones.

Beyond flexibility, v1.3 added the auto-growing Qwen References that the suite is now known for. The node starts with qwen_reference_1; connect it and qwen_reference_2 appears, then 3, up to nine. Each is a Qwen text/vision input - identity, clothing, objects - and the ordering is deterministic: connected First Frame is Picture 1, Last Frame is Picture 2, Qwen References follow consecutively. The picture_map output prints the exact mapping so you're never guessing which slot H3 saw.

The honest footnote, repeated because it matters: Qwen References are Qwen-only. They don't get inserted into minimax_refs as native Ref2VA/DiT reference latents. They guide through the vision encoder - great for "this is the character, here's the coat" - but they're not the persistent-reference mechanism some other H3 nodes use.

Inputs and outputs

Required: clip, vae, prompt, width/height, duration (seconds, snapped to the 17k+5 grid), ref_image_size. Optional: first_frame, last_frame, qwen_reference_1 (then autogrown). Outputs: positive, latent, and picture_map.

For a long chain, the recommended pattern stays FL2VA with both anchors connected - repeated Last Frames are your quality resets. But now you can also do a T2VA Clip 1 if you don't have a target frame yet, which is genuinely useful for the first segment of an exploration.

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; H3 checkpoints + VAEs + Qwen 3 VL from Comfy-Org's MiniMax-H3 repo. Qwen References only work with Qwen 3 VL loaded, so that weight is effectively mandatory for this node's flagship feature.

Gotchas

If you connect a Qwen reference and a character still drifts, check picture_map first - the slot ordering shifts the moment you add or remove a First/Last Frame, and a reference that landed in Picture 4 when you assumed Picture 2 is silently ignored for identity. And remember: optional Last Frame means the quality-reset behavior is on you - keyframe-free continuation is the less-tested path, so keep Last Frames connected in long chains unless you're deliberately going without.