Continuity
One node for AI video and stills in ComfyUI. Write a prompt, attach media with
@, press Render. Six model families, all through ComfyUI core, all local
open weights. The one thing that can talk to the outside is the prompt
refiner, and only if you point it at your own server or provider - LM Studio,
Ollama, or a hosted API with your key.
What it does
One node, one prompt
Drop the node, type, press Render. Files you attach get an @ name and you
cite them in the sentence, so the prompt says which picture is the person and
which is the room. The pills under the box are the whole setup: duration,
aspect, resolution, the model, and a sampler row for when you want it.
The node walks every control.
References and the cast
A reference contributes what you say it contributes - the person, the place,
the look - and the same face can be saved once and brought back into any
piece by name. Two references cited in one prompt:
The pre-stage
Video models eat stills: start frames, end frames, references. The pre-stage
is a second card on the same canvas that makes them, with any of the stills
families or H3 itself, and hands the result straight into the shot. One Queue
runs both; an untouched pre-stage doesn't re-render.
Timelines
Under the prompt: Write the next shot. Every shot on the strip is a whole
generation with its own prompt, references, LoRAs and even its own family - a
pre-stage on Krea 2 feeding start frames into shots on LTX 2.5 is one strip.
Seams carry picture and sound across the cut, or don't, independently.
Timelines.
Chat
The other way in. A room where you say what you want - a fox in a snowy wood
at dusk, then now a clip of it, she looks up, then bluer - and the
refiner model writes the prompt and renders. Every picture and clip gets a
handle you can cite in the next line, so a still becomes a shot's first frame
by naming it. It is the same node underneath, on the same queue, and
everything lands in the same output folder. Chat.
Tools
The wordmark opens a dashboard of tools that work over the piece: presets, a
ControlNet bench, a blockout bench for staging a scene out of boxes and
walking a camera through it, and an upscale bench. Tools.
The style atlas is 941 looks captioned from real clips, applied as a prefix
to the prompt:
Where it is headed
The name is the goal. A single shot from an open video model is a solved
problem; a piece made of many shots that still looks like one piece is not.
Every continued shot comes out a little brighter and a little harder than the
one it continues, and after a handful of seams the strip has drifted somewhere
you never asked for. Most of the work now goes there: measuring the drift on
real strips rather than guessing, and pushing the number of shots a piece can
hold before it shows.
The other direction is getting the graph out of the way. The chat room and the
blockout bench are both ways of saying what you want without touching a pill,
and the pre-stage plus the timeline are what let one prompt become a scene.
There is no cloud in any of that, and there won't be: everything renders on
your hardware from weights you downloaded.
Install
cd ComfyUI/custom_nodes
git clone https://github.com/roadmaus/ComfyUI-Continuity
That leaves one folder, ComfyUI-Continuity, inside custom_nodes/. Restart
ComfyUI. Nothing to pip install.
Already have MiniMax Creator installed
Don't clone. This pack was renamed, and GitHub redirects the old address here,
so a pull in the folder you already have is the whole update:
cd ComfyUI/custom_nodes/ComfyUI-MiniMax-Creator
git pull
The folder name doesn't matter to ComfyUI, so rename it or leave it. If it came
from the Manager, update it there as usual.
Cloning next to the old folder is what breaks. The node ids stayed the same
through the rename so that old workflows keep loading, so two folders are two
packs registering the same ids, and the result is no node in the search at all,
under either name. If you already have both, delete one and restart. Your
presets, settings, favourites and LoRA memory are in ComfyUI's user/
directory, not in the pack folder.
From a script or a coding agent
The pack ships a command-line client that renders on a running ComfyUI with no
browser. It picks the weights and the turbo settings itself and downloads the
result:
python3 skills/continuity-render/render.py families
python3 skills/continuity-render/render.py h3 "a cat stretches on a sunny windowsill"
python3 skills/continuity-render/render.py --help covers the options, and
docs/headless.md covers the HTTP API underneath. For
Claude Code, copy skills/continuity-render into ~/.claude/skills/ and ask
for a render in plain words.
Documentation
Models
See docs/models.md for which files you need and where they go.
There is also an optional neural refiner: NVIDIA's DLSS 5 neural renderer
as a material pass over finished stills and clips (skin, hair, fabric, contact
shadows), through the open-source MLX-DLSS
port, which travels with the pack. It is not an upscaler and it ships with
nothing of NVIDIA's — the weights are extracted from your own copy of the DLSS
DLL on the settings page. See docs/tools.md.
Thanks
This pack is glue. The work underneath it belongs to other people:
Every node pack on that list is optional: if one is installed, the matching
pills light up, and if it is not, they say what is missing. The models are
optional in the same way — none of them is downloaded for you, and nothing on
this list ships inside the pack except the four vendored libraries it names.
License
MIT. ComfyUI itself is GPL-3.0 and this pack imports it; if you
redistribute the two together rather than as a node pack people install
themselves, that combination is what the GPL has an opinion about.
