Morpheus (Video Sampler) 🌙

MiniMax H3 is genuinely great at making a clip, but it tops out around 5–15 seconds and ComfyUI ships no extend/continue node for it. So if you want a minute of video, the only real route is to run it a bunch of times and hand each shot's last frame to the next one as its first keyframe. That handoff - plus every seam-adjacent annoyance that makes the naive version come out wrong - is exactly what Morpheus does. It's the "video sampler" for the whole Morpheus storyboard suite, and it's the node you wire the final chain of Morpheus Dream shots into.

What it actually does

You feed it a chain of shots (the output of the last Morpheus Dream in a row), a model, clip and VAE; it samples every shot, passes the last frame forward as the next shot's first keyframe, and hands back one big images batch, one audio track, and an fps float. Four things a hand-wired graph gets wrong are handled here, and they're the reason this node earns its keep:

• The seam frame. A continuing shot's first frame is the previous shot's last frame, so seam_trim (default 1) drops the duplicate. Raise it to 3–6 and it also swallows the re-acceleration - the model is handed a still frame with position but no velocity, so every shot eases in from rest and the subject visibly speeds up at each seam. Trimming happens at decode time, so cached shots are re-used and tuning it costs nothing.
• Loudness. VAEDecodeAudio normalizes per decode, which steps the level at every seam. Morpheus normalizes once over the whole track and ramps the seams (seam_fade_ms - a fade, deliberately not a crossfade, so audio never drifts against picture).
• RAM. 1344×768 float32 is ~12 MB per frame - roughly 1.5 GB per 5-second shot. The node manages that so you don't OOM a long chain.
• fps is not a parameter. 24 is baked into the model; it comes back out as a FLOAT (24.0) that drops straight into Create Video, with one monotonic progress bar across the whole run instead of a bar that fills and resets per shot.

The inputs that matter

width/height are one canvas for the whole storyboard. H3's own budget is a 768 short edge with a 768×1344 area cap - exceeding it is slow and out of distribution, and the report output will tell you. Height mismatch matters more than most expect: H3 stretches the first frame onto the canvas and doesn't crop, so a wrong aspect ratio distorts every shot.

seed is the base; each shot adds its index and its own seed_offset, so no two shots share noise. shift_video (default 12) is H3's video flow shift, applied automatically unless your model already went through a MiniMax H3 Sigma Shift node. audio defaults to concat - decode, trim to the shot's exact length, ramp seams, normalize once. Without audio_vae wired, the video comes out silent, which is the single most common "why is my Morpheus video mute" answer.

cache: disk stores each shot's sampled latents under user/kinburg-nodes/minimax_shots (~7 MB per shot). Edit shot 5, re-sample 5..N, and shots 1–4 replay instantly. The cache key covers architecture + LoRA patches but not the exact weight file - if you swap to another H3 checkpoint of the same size, bump cache_tag to invalidate. shots_range ("2-4") renders only part of the chain; shots before it are still needed for the handoff frame, but free if cached.

The optional llm_config input is the fancy bit: wire a Local LLM Settings (GGUF) with a Vision Settings (GGUF) mmproj and a vision LLM gets to see the frame each shot really starts on and rework its opening. Per-shot control comes from the refine widget on each Dream node. On 12 GB VRAM with a 26B model, leave llm_keep_loaded off - the LLM is loaded, used and killed around each shot, which is the difference between "runs" and OOM.

Installing it

The node ships in Kinburg/Kinburg-Nodes. Easiest is ComfyUI Manager - search "Kinburg-Nodes" - or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/Kinburg/Kinburg-Nodes

then restart ComfyUI. The pack itself has no pip dependencies; the LLM half wants llama-cpp-python (CUDA), which install.py installs automatically when Manager runs it. What you do need to fetch yourself is the H3 model - model, VAE, text encoder - and one thing before that: check the H3 Community License. It excludes the US, EU, UK and South Korea from its applicable territory, so in those regions you're not licensed to run the local weights at all.

Where people get burned

Silent output (missing audio_vae), distorted shots (canvas/keyframe aspect mismatch), and sluggish runs (canvas over H3's budget). The last_frame output is worth remembering too - it's the cover, or the start_frame of a later run if you're building longer than one storyboard.