H3 Infinite - Start FFLF v1.2
H3 Start v1.2 — The Quiet Refresh That Standardized the Chain
- clip
- vae
- first_frame
- last_frame
- reference_image
- positive
- latent
H3ContinuousStartV11 is the "H3 Infinite - Start FFLF v1.2" node, and it's the definition of an unglamorous-but-important release: functionally it's the v1.0 Start with a new class ID and a new home in the "Herrgotts H3 Infinite Continuation Suite" category, because it's the Clip 1 anchor of the v1.2 generation set. If you're running the v1.2 Start → Continue → Analyze → Stitch chain - the first "H3 Infinite" workflow the suite shipped - this is the node that opens it.
Everything that matters about it you already know from v1.0: First Frame and Last Frame in, duration in seconds (default 10, snapped upward to H3's 17k+5 grid → 243 frames ≈ 10.125 s at 24 fps), clip/vae/prompt/width/height/ref_image_size required, optional reference_image as <Picture 1> - Qwen-only, not a persistent Ref2VA reference latent. Outputs are positive and latent into the sampler.
What changed between v1.0 and v1.2 is less visible: the handover story around it. The v1.2 analyzer and Continue nodes got the no-lock freeze fallback, and the Save/Load latent workflow got standardized around sequential clip indices. The Start node itself stayed put - which is exactly what you want from a Clip 1 node. It's the stable entry point; the interesting machinery is downstream.
What it's for
Two jobs, really:
- Build Clip 1 with strong visual anchors (FL2VA) so the first segment starts on the right composition and identity.
- Kick off the chain's quality-reset pattern: repeated Last Frames across segments are what pull image quality back toward a clean reference before drift accumulates.
Install
cd ComfyUI/custom_nodes
git clone https://github.com/HerrgottMargott/Herrgotts-H3-Infinite-Continuation-Suite.git
Restart ComfyUI, reload the UI. No pip deps, no bundled weights - H3 models and VAEs come from Comfy-Org's MiniMax-H3 repo (the big download; the node itself is thin).
Gotchas
The two things that trip people on this family: duration snaps upward (5.0 s can come out as 5.17 s - that's the H3 grid), and the optional reference_image will not behave like a native reference - it's vision guidance through Qwen. If you expected Ref2VA-style locking, that expectation is the bug. For new projects, v1.3+ made First/Last Frames optional and added Qwen References plus picture_map, so that's the forward path; v1.2 remains the reliable baseline that old "Infinite" workflows were built on.
Inputs (10)
| Name | Type | Default | Description |
|---|---|---|---|
| clip | CLIP | — | |
| vae | VAE | — | |
| prompt | STRING | — | |
| width | INT | 134432–16384 | — |
| height | INT | 76832–16384 | — |
| duration | FLOAT | 10.00.25–150 | Requested duration in seconds at H3's native 24 fps. The actual clip snaps upward to H3's 17k+5 frame grid (10.0 s -> 243 frames ~= 10.125 s). |
| first_frame | IMAGE | — | |
| last_frame | IMAGE | — | |
| ref_image_size | COMBO | match | 2 options: match, max |
| reference_imageopt | IMAGE | Optional Qwen-only identity/style reference. Address it as <Picture 1>. It is NOT added to minimax_refs. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| positive | CONDITIONING | — |
| latent | LATENT | — |