H3 Continuous - Start FFLF (Legacy v0.x)

Every long H3 chain starts here. H3ContinuousStart is the legacy v0.x "Start FFLF" node - the Clip 1 builder that takes a First Frame, a Last Frame, a prompt, and hands back the conditioning and latent for your first MiniMax H3 segment. FFLF = First-Last-Frame-to-Video, and it's the recipe this whole suite is built on: give H3 a strong visual start and a strong visual endpoint, and it fills in the middle with better composition, identity, and image quality than a plain prompt ever gets you.

The quality-reset idea is worth internalizing. In a long chain, every clip's Last Frame is the next clip's launch point, and a strong repeated endpoint pulls drift back toward a clean reference before it accumulates. FL2VA gives you that reset on every segment - which is exactly why the suite picked it over bare continuation.

How it works

Required inputs: clip (the H3 CLIP model), vae, prompt, width/height (default 1344×768), length (frames at 24 fps, default 243 ≈ 10.1 s, snapped upward to H3's 17k+5 grid), first_frame, last_frame, and ref_image_size (default "match"). Optional: reference_image.

Two things to know about the reference inputs, because this is where beginners get burned:

• The first_frame and last_frame are native FL2VA temporal anchors - they're actually conditioning the start and end of the clip.
• The optional reference_image is <Picture 1>, and it's Qwen-only. It goes through the vision encoder as a Picture, but it is not a persistent Ref2VA reference latent. The author calls this "matching the working production behavior," and it means a reference image guides identity but doesn't lock the model to it the way native references would.

Outputs are the simple pair: positive (CONDITIONING) and latent, straight into the sampler.

Install

Part of 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 and reload. No pip deps, no bundled weights - you still need the H3 FL2VA transformer, video/audio VAEs, and Qwen 3 VL from Comfy-Org's MiniMax-H3 repo. That's the heavy download; the node itself is lightweight.

Gotchas

The legacy length is in frames, not seconds - 243 is the default and matches ~10.1 s, but change it in a frame-count panic and you'll get odd lengths because it snaps to the H3 grid. This v0.x node is superseded: v1.0 swapped length for duration in seconds, and v1.3+ made First/Last Frames optional and added Qwen References with a picture_map. For new work use the newer Start. But if you're loading an old workflow or want the minimal Clip-1 node, this one still runs as registered.