SkyReels V4

SkyReels is Skywork's video generation line, and the reason it keeps getting attention is that it does native audio: the model can generate sound effects synced to what's happening on screen, which most video models simply don't do. V4 is the current tier, and this node runs it in the cloud through the Runware pack. Type a prompt (optionally feed it reference images or a reference video), get back a VIDEO you can preview or save. No local video stack, no VRAM marathon, no frame-interpolation fiddling - a video node that behaves like every other node in the pack.

SkyReels V4 sits in the "text-to-video with bells on" camp: it takes a prompt, does LLM-based prompt expansion by default, can take reference inputs, and has that audio toggle. If your ComfyUI workflows have been dead silent, this is the cheapest way to change that.

How it works

It sends a videoInference request to the Runware API. The returned video is wrapped into a native VIDEO object when your ComfyUI supports it (via comfy_api's VideoFromFile), and degrades to a saved file path otherwise. Outputs are MP4/WEBM/MOV. The video lands as a tensor you can wire into a save or preview node, or read as a path.

Inputs worth touching

• positivePrompt (required) - describe motion, camera, subject. With settings.promptExtend on (it's on by default), the model's LLM rewrites and expands your prompt before generation - better results, but note it affects reproducibility, same as any prompt-enhancer.
• duration - 3–15 seconds, default 5. Longer isn't always better; 5–8s is where the quality/consistency sweet spot usually sits.
• size - a 19-choice dropdown from 480p (1:1) up to 1080p (16:9) and portrait variants. Pick the one that matches your target; "from input" options inherit the reference's aspect ratio.
• settings.audio - off by default. This is the headline feature: flip it on and the model generates sound effects for the video. If you're making anything that gets watched with speakers on, try it - it's the differentiator for this model line.
• referenceImages / frameImages / referenceVideos - three ways in: reference images (via the Runware Reference Images builder), a batch of frames as a plain IMAGE, or a reference video (URL/UUID). This makes SkyReels V4 a competent image-to-video / video-extension model, not just text-to-video.
• numberResults (1–4) - seeded variations. Video is expensive and slow; do 1–2 and pick.
• safety.mode - none (default), fast, or full (when safety.checkContent is on). Video moderation is heavier than image; fast is usually enough if you're running it at all.
• outputFormat / outputQuality - MP4 (default), WEBM, MOV, and a compression slider.

The single output is video → wire to Save Video / Preview Video.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Or ComfyUI Manager → search Runware → install → restart. API key via Settings, RUNWARE_API_KEY, or runware auth login. No local model downloads.

Honest gotchas

Cloud video generation is the priciest and slowest thing in this pack - a 15-second 1080p clip with audio is a real job, and the title bar will happily show you the cost. Budget for it. Also remember settings.promptExtend changes your prompt under the hood, so if you're chasing a specific frame from a specific text, you may want it off and to write the full prompt yourself. And if your video output arrives as a plain file path rather than a VIDEO tensor, that's the older-ComfyUI fallback - update ComfyUI and it'll come back as the native type.