H3 Continuous - Start FFLF (Legacy v1.0)

H3ContinuousStartV1 is the "Start FFLF (Legacy v1.0)" node, and it's the version of the Clip 1 builder that most people's first H3 long-video workflows actually used: First Frame in, Last Frame in, a prompt, and a duration in seconds instead of a raw frame count. That single change - duration (default 10.0 s) with the node snapping up to H3's 17k+5 grid (10.0 s → 243 frames ≈ 10.125 s) - is the whole reason it replaced the v0.x original. Nobody wants to compute 24 fps × N seconds in their head while mid-workflow.

It's the anchor node of the v1.0 suite: this makes Clip 1, the v1.0 Continue node makes Clip 2+, the analyzer finds the freeze, the stitch node prepares each segment. FL2VA logic is unchanged from v0.x - the First Frame and Last Frame are native temporal keyframes that bound the generation, and repeated Last Frames across the chain are your quality resets that pull identity and image quality back toward a clean reference before drift accumulates.

Inputs and outputs

Required: clip, vae, prompt, width/height (1344×768 defaults), duration (seconds), first_frame, last_frame, ref_image_size (default "match"). Optional: reference_image - and the same caveat as every version of this family: it's <Picture 1> through the Qwen vision encoder, Qwen-only, not a persistent Ref2VA reference latent.

Outputs: positive and latent, which feed the sampler. That's it - the Start node's job is to set up Clip 1, and the whole continuation machinery comes after it in the graph.

Install

Same pack, same commands:

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

Restart ComfyUI, reload the browser. No extra pip dependencies; no bundled weights - the H3 FL2VA transformer, the video/audio VAEs, and Qwen 3 VL come from Comfy-Org's MiniMax-H3 repo. Budget for a large download, especially the Qwen 3 VL encoder.

Gotchas

Duration snaps upward to the H3 frame grid, so a 5.0 s request may render as 5.17 s - that's the model's native temporal grid, not a rounding error in the node. And keep both First and Last Frames connected if you want the FL2VA behavior; v1.0 still requires them (the flexible "either is optional" treatment didn't arrive until v1.3). If you're starting fresh in 2026, jump to v1.3 or v1.4 Start instead - v1.0 is the reliable legacy workhorse for old workflows and A/B comparisons, and it's exactly the right node when you want to understand the suite's clip-1 mechanics without the newer surface area.