☁️BizyAir Kling O1 Video Edit

What it is

This is Kling doing to video roughly what instruction-editing models like Flux Kontext do to still images: hand it an existing clip and a text instruction, and it edits the clip rather than generating something new from scratch. That's a genuinely different job from every other Kling node in this pack - the T2V and I2V nodes all create a clip; this one changes one you already have. If you've got a generated or filmed clip and want to alter its style, subject, or specific elements without redoing the whole shot, this is the node built for it.

The inputs that matter

• prompt - the editing instruction: what to change about the video.
• ref_video - required. The clip you're editing.
• aspect_ratio - 16:9, 9:16, or 1:1.
• keep_original_sound - off by default; turn it on to carry the source clip's audio through to the edited output rather than getting fresh or silent audio.
• fast_mode - off by default, per the node's own tooltip: choose whether to use a faster processing mode. Worth trying when you're iterating on an instruction and don't need final-quality output yet - trade some quality for turnaround speed while you dial in the prompt, then turn it off for the take you actually keep.
• ref_image_1 through ref_image_4 (all optional) - up to four supplementary reference images, for cases where the instruction alone doesn't fully specify what you want changed to (a specific outfit, a specific style reference).

Outputs: video, and bizyair_model_name, reporting kling-o1 as the resolved backend model.

How to install it

ComfyUI Manager: search "BizyAir," install, restart. Manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAir.git

A BizyAir API key is required - Kling has no local execution path at all, so this node is purely a client for the hosted editing endpoint.

Common issues

Edit changes more of the video than intended. This is the video-editing equivalent of the drift people complain about with whole-frame instruction editors on still images - without a mask to constrain the change, the model has latitude over the whole clip. Keep the instruction as specific as possible about what should change, and treat anything outside that description as at some risk of shifting too.

Using fast_mode for a final render and getting a rougher result than expected. That mode exists to trade quality for speed - reserve it for iterating on your prompt and reference images, and switch it off for the generation you actually intend to keep.

Original audio missing from the edited output. Check keep_original_sound - like the reference-video-to-video node elsewhere in this pack's Kling section, it defaults to off, so you need to opt in explicitly if the source clip's audio matters.

Job fails outright. Check API key and account credits first. Video-editing calls that ingest a full reference clip plus reference images are among the heavier requests in this pack, so it's also worth confirming your source video loaded and is in a supported format before assuming the API itself rejected the call.