Won't respond to stupid comments.
Most “long H3” graphs either fail on import (ConcatenateVideo / MiniMaxH3AddGuide) or fire clip 2 before clip 1 is actually done. This one doesn’t.
One character photo. Queue once. Three Ref2VA segments (15s + 14s + 11.8s) get joined into a single MP4. Turbo LoRA, 8 steps.
Clip B cannot start until A’s last still exists. Clip C waits on B. That’s wired, not hoped for.
Continuation uses nodes you already have: last still as <Picture 2>, last 39 frames + 1.625s of audio as <Video 1> / <Audio 1>. That 39-frame length is the native H3 grid (n % 17 == 5). We don’t feed 48 frames and let the model silently trim picture while the audio stays 2 seconds.
Final join is ImageBatch + CreateVideo, so it loads on current ComfyUI without the newer ConcatenateVideo pack.
The example prompts are a boring kitchen morning on purpose. Swap them for your story. Keep <Picture 1>, <Picture 2>, <Video 1>, and <Audio 1> on clips 2 and 3.
Description
FAQ
Comments (4)
0.4 megapixels video took hour and a quarter on the rtx5016ti 16gb. I added 4 (Deno) Local LLM Loaders blocks for autopromt on the idea of video and promting all 3 parts separetely. I lked it need to watch what will it be. Thanks for sharing
No problem I guess i should have said i have a 6 rtx6000 pro max-q 96gb
@fuckciviindabut How much time took to make 1megapixel video on that monster system?
@vitoldturkind581 1000 seconds