Morpheus Dream 🌙

Morpheus Dream is deliberately boring: it's one shot of a MiniMax H3 storyboard. A prompt, a duration, an optional start/end frame, and the rule for how the shot begins. Chain several together and the whole row feeds Morpheus (Video Sampler), which does all the hard seam work. On its own it's the node you reach for when you want to hand-build one opening shot and let an LLM write the rest of the story around it.

How a shot decides where it starts

That's the whole design, and it's a three-way priority:

1. A wired start_frame always wins - that image is the first frame (stretched to the canvas, no crop, so match the aspect ratio).
2. Otherwise, link: continue - the shot inherits the previous shot's last generated frame. A continuous take.
3. Otherwise (link: cut, or it's shot 1) - no keyframe at all: pure text-to-video. That's your hard montage cut.

end_frame is the target the shot moves toward (cover-cropped to the canvas). Worth knowing: the next shot still inherits the generated tail frame, not this image, which is what keeps the two visually identical.

The duration widget is in seconds but snaps up to the model's frame grid - 17k+5 frames at 24 fps, i.e. steps of 0.71 s, so you'll see values like 5.17, 5.88, 6.58. The trained range is 124–362 frames (5.17–15.08 s); anything shorter runs but is out of distribution. The info output tells you the length you actually got, so check it once rather than assuming.

seed_offset is the "re-roll just this shot" dial: it's added to this shot's seed and folds into the shot's cache key, so every earlier shot still replays from cache while later ones re-sample (their first frame changed). keyframe_strength (0.999) is how hard the model is held to its keyframes - it's H3's visual_cond_noise_aug. Lower it a hair on a long chain where re-encoded handoff frames accumulate color drift, but otherwise leave it alone; it's an advanced knob with a good default.

The optional stuff

• shots - wire the previous Dream here to chain them. Left→right is shot order; leave it empty for the first shot.
• refine - whether the sampler's vision LLM (its optional llm_config) reworks this shot just before sampling, when the real first frame finally exists. auto rewrites the opening when the shot starts from a frame the writer never saw (an inherited tail), and writes the whole shot when there's no proper prompt. off is right when you pasted a finished prompt yourself. Ignored entirely if the sampler has no llm_config.
• Outputs: shots (the chain, to pass on) and info (the snapped length and what the shot got wired).

Installing and gotchas

Same pack as everything else: ComfyUI Manager → search "Kinburg-Nodes", or git clone https://github.com/Kinburg/Kinburg-Nodes into ComfyUI/custom_nodes and restart. No per-node install. The heavy lift is the H3 model itself (and note the H3 Community License excludes the US, EU, UK and South Korea - check your region before downloading the weights).

The mistakes people make here are the usual suspects: keyframes whose aspect ratio doesn't match the canvas (stretched, not cropped), and durations below 5.17 s producing shots that feel rushed because the model wasn't trained there. For a first storyboard, chain a few Dreams with link: continue, feed the row into Morpheus, and let the seams be someone else's problem.