arkennemasis Video Model (H3 or LTX — only one runs)

A canvas that offers two video renderers - MiniMax H3 and LTX-2.5 - can't join them into a plain switch. ComfyUI evaluates everything an output depends on, so a naive switch would stage both models: tens of gigabytes of weights loaded, every shot rendered twice, then one result thrown away. ArkVideoModel is the pack's fix: a dropdown that picks the renderer, with both branches declared lazy, so the branch you didn't choose is never evaluated at all - no model load, no render, no waiting.

How the laziness works

check_lazy_status asks the executor for only the chosen branch's inputs - the H3 branch's h3_video and h3_path, or the LTX branch's ltx_video and ltx_path. Because those inputs are lazy and never requested, the entire upstream renderer subgraph is skipped. That's the same lazy-input trick the pack uses for shot selection, applied one stage later. The dropdown is the switch; you don't need to mute either renderer, because muting would only duplicate what laziness already does.

The model enum has two choices - "MiniMax H3 - local, omni-modal (its own audio)" and "LTX-2.5 - local, 22B distilled." Two 33B/22B-class local models, so it's worth knowing what each is before you commit: H3 is MiniMax's omni-modal model that generates picture and audio in one pass (4–15s clips, ~42.5GB of weights, and a community licence that excludes the US/EU/UK/Korea from running the local weights - check that before you download). LTX is Lightricks' speed-first line; the 2.5/22B distilled generation is the current heavyweight but LTX's whole identity is "seconds per clip on consumer VRAM."

The inputs that carry the film

Required: model and basename (default walkthrough - the finished film's name). The clever bit: the chosen model's tag is appended to the basename, so walkthrough becomes walkthrough_h3 or walkthrough_ltx. Switch renderers and you don't overwrite the film you already made.

Four optional inputs, all lazy: h3_video / h3_path (from the pack's Hailuo Scene node) and ltx_video / ltx_path (from the LTX branch, with ltx_path coming from ArkVideoSave).

Outputs: video (the clip), path (the file on disk - carried through so a single dub node downstream can use it), model (which one ran, as a string), and film_name (the basename_tag that becomes the filename). If you set a model but wire nothing into it, you get a clean ExecutionBlocker message naming the missing input - everything downstream skips quietly instead of the run dying mid-way.

Install

ArkVideoModel is one of the 61 nodes in the comfyui-arkennemasis pack, in arkennemasis/Video:

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis
pip install -r comfyui-arkennemasis/requirements.txt   # then restart ComfyUI

Or ComfyUI Manager → Install via Git URL with the repo URL. The node itself needs nothing, but the renderers it fronts need their weights in ComfyUI's model folders and the hardware to run them - the H3/LTX model download is on you, not the pack.

This is one of those nodes that's trivial until you think about what it prevents: a whole class of "I switched models and it staged 70GB of weights before telling me my other choice was wrong" pain, solved by asking the executor not to touch what you didn't ask for.