Magnific Generate Video

A cloud video generator inside your graph. Prompt, duration, aspect ratio - and optionally a start frame, an end frame, a reference image or a reference video, depending on which model you picked. What comes back is a ComfyUI VIDEO object, not a file path you have to go find.

Read the tradeoff honestly before you build around it: every run is a metered call to Magnific, your prompt and reference media go to their servers, and their content filter applies rather than yours. That's the price of a model with no open weights. The case for it is short and real - you get a five- or ten-second clip without a 20-minute local render, and you get models you can't download.

What you set

prompt (multiline), model (defaults to auto, which lets Magnific pick), duration (5 or 10 seconds - the list narrows to whatever the selected model supports), resolution, aspect_ratio (auto, 16:9, 9:16, 1:1, 4:3, 3:4, 21:9), sound_effects (only models with native audio honour it) and seed.

seed is worth understanding: it's sent to Magnific and reproduces the result only on models that honour it - Seedance is the example named in the tooltip. On models that ignore it, changing the seed just forces a new run, which is not the same thing as a new result being reproducible.

The optional media sockets are where the interesting behaviour lives. start_frame is image-to-video. end_frame is the closing keyframe, on models that support it. reference_image is a softer hint than a start frame. reference_video is a video reference, on models that support it - Seedance 2.0 is the named example - and it isn't combinable with a start or end frame, so pick your mode rather than stacking them. library_references takes a Magnific Library Reference node; video accepts characters, elements and styles but never locations, and the node will stop and tell you which model refused which type before anything is sent.

folder comes from Magnific Save To; without it, results go to your Personal project.

Outputs are video, creation_identifier and metadata. Because the result is a real VIDEO type, it wires into video-aware nodes downstream rather than being an image batch pretending. metadata carries the model, prompt, seed, duration, size and the link to the creation on magnific.com.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Code2Collapse/ComfyUI-CustomNodePacks.git

Or "CustomNodePacks" in ComfyUI Manager, then restart. The Magnific family ships inside this pack (ported from Magnific's own magnific-comfyui 0.7.0); installing the vendor zip on top of it gives you two nodes with the same name, so don't. Sign in through the Magnific menu or the Save To node; tokens live at ~/.magnific/comfyui_auth.json, outside the workflow.

Where people get burned

Duration is a menu, not a request. 5 or 10 seconds, and if the model doesn't do ten, you get what it does. If your graph wants a specific runtime, check the result's length rather than assuming.

Reference modes are exclusive. A start frame, an end frame and a reference video each push the job into a different mode on the service side, and combining the wrong pair fails at submit time rather than producing something odd. Read the tooltip on reference_video before you wire it alongside a keyframe.

Credits. Video is the expensive end of every hosted API, and this is the node where an unattended queue gets costly. Compare against the local alternative before you commit: Wan and LTX on your own card are slower and free, and the KB's Wan chapter is blunt that local video means GPU-tier decisions and 49–81 frames as the community norm. The honest split is local for iteration, hosted when you need a specific model or you're already out of patience.