小珠光合并视频

Every video workflow ends the same way: a pile of generated frames needs to become an actual video file. XiaozhuguangVideoCombine (小珠光合并视频) is the pack's encoder for that last step - image sequence in, mp4/webm/gif out, with an optional audio track muxed in. If you're generating a clip with a video model, then upscaling or editing its frames, this is the node that turns the result into something you can play, share, and put sound on.

Mechanically it's FFmpeg under the hood via the pack's bundled imageio-ffmpeg, which is the only real runtime dependency in the pack's requirements.txt. The node takes the frames, a frame rate, a filename prefix, a container format, and a CRF (constant rate factor - the H.264 quality knob, lower = better, default 19), then encodes. Two behaviors worth knowing. First, 模式 (mode) has Save and Preview: Preview encodes to a temp file you can watch without committing a numbered file to the output folder - good for checking a clip before you finalize. Second, it's content-aware about re-runs: the pack added a dedup cache that fingerprints your input frames and audio, so re-running the graph with identical content reuses the previously encoded file instead of re-encoding and burning another 30 seconds of waiting. That's the kind of lazy evaluation that actually saves you time, not just VRAM.

Inputs (displayed in Chinese, matching the pack's i18n):

• 图像 (images) - your frame sequence (required).
• 帧率 (frame rate) - FLOAT, default 8. Match this to your generation model's fps or the motion will feel wrong.
• 文件名前缀 (filename prefix) - base name for the output file.
• 格式 (format) - mp4 (default), webm, or gif. GIF for lightweight loops, webm if you want VP9 quality, mp4 for everything else.
• CRF - quality/size tradeoff, 0–51, default 19. Lower is heavier and prettier.
• 模式 (mode) - 保存 (Save) or 预览 (Preview).
• 音频 (audio, optional) - an AUDIO input to mux in. This is the natural pairing for the pack's TTS nodes: generate a voice track, wire it in, and your video comes out with sound.

No outputs - it's a terminal node; the file lands in ComfyUI/output/.

Where people get caught: the frame-rate mismatch. If your model generates at 24fps but you leave the default 8, the video plays back slower than you intended. Set 帧率 to your model's actual fps. Also, the filename prefix can include a subfolder path (xzg_video/name works, and the pack fixed previewing for that case), but keep the base name sane or the output folder fills with near-duplicate numbered files. And if your audio and video lengths disagree, the mux follows the video - trim your audio upstream if it matters.

Install: ComfyUI Manager → ComfyUI-xiaozhuguang → restart, or git clone into custom_nodes. imageio-ffmpeg comes with the pack (and bundles its own FFmpeg binary), so no separate system install. This is the node that closes the loop on every Xiaozhuguang video workflow - and it's genuinely the part people most often forget to plan for.