Load Video and Split Video

Load Video builds an ordered playlist with a player inside the node. Trim each clip by its original frame numbers and join its audio with the same cuts into native VIDEO. Find both nodes under Eclipse → Video.

Visual tour

Load Video: arrange and trim the playlist

The example uses two six-second sample clips at 25 and 16 FPS. The first keeps 100 frames starting at frame 20; the second keeps its full range. Source timing retains each clip's playback speed. The metadata checkbox selects a source, but only yields generation data when that file contains usable A1111 parameters

Direct video upscaling with Eclipse

Connect Load Video → Upscale Video [Eclipse] → Save Video with Generation Data directly. The upscaler's model selector switches between PyTorch models and compatible TensorRT engines while keeping the same VIDEO connections. Audio and timing travel inside VIDEO; this route needs no splitter.

See the direct Eclipse chain tour for backend requirements and the optional generation-data connection.

Split Video: process images while retaining their timing

Connect the loader's video to Split Video. This example routes images through a 2× spatial resize while audio and timing go directly to the saver. Keep frame count and order unchanged. This route holds the complete IMAGE batch in RAM; for long videos use the streaming Upscale Video tour.

Add and trim clips

1. Choose Browse videos for the dedicated Input/Output video grid, or Upload videos for multiple local files. Dropping local video files also uploads them.

2. Check files in the desired order and click Add selected. Searching, sorting, and changing folders keep the pending selection. Clicking a thumbnail previews that file without adding it or choosing its generation metadata.

3. Select a playlist row. Start frame is zero-based; Load cap is the number of source frames to retain. Both default to 0: start at the beginning and retain every remaining frame. A cap beyond EOF retains the available frames; a start beyond EOF reports an error. Changing either value seeks the Source preview to the selected start; playback stops at the cap. This quick preview uses the clip's reported FPS. Use Preview selected trim for frame-exact review of variable-frame-rate sources.

4. Use Include, Mute, Duplicate, Remove, the arrow buttons, or drag rows to arrange the join. Disabled clips contribute neither frames nor audio.

For example, start 20, cap 100 retains source frames [20, 120). Every clip has independent values, and its audio uses that exact time interval. Missing audio, leading/trailing audio gaps, and muted clips become silence. The first audio track is normalized to 48 kHz stereo, without changing its tempo.

The editor exists immediately when the node is added. Playlist rows, trims, metadata choice, active source, and pending browser selections are serialized with the workflow and survive a page reload before the first queue. Uploads must finish before their files can be added. A built preview is temporary and can be rebuilt after a reload or expiry.

Resize the node to suit your workspace. Its outer size stays the same when you change options or previews, and the saved size is restored with the workflow.

Global size and timing

width and height apply to every clip. Both 0 use the first included clip's displayed size. With one dimension 0, it follows that clip's aspect ratio. pad fits inside the size with black borders; crop fills and center-crops; stretch fills both dimensions directly. Encoded VIDEO pads an odd dimension by one edge pixel for H.264.

timing_mode = source preserves each frame's presentation duration. Joining one second at 25 FPS with one second at 16 FPS produces 41 frames over 2 seconds. The 16 FPS section keeps its original playback speed. The resulting VIDEO has variable frame timing.

timing_mode = fixed exposes fps and duplicates/drops frames on that output grid. Its last frame may extend the result by less than one output frame; audio is padded to that boundary. The duration output describes this encoded result.

Preview and outputs

Source plays the original file within the selected range. Editing a trim returns to that clip even when a joined result or browser thumbnail was being previewed. Preview selected trim and Build joined preview explicitly render a playable MP4, shown under Joined. Editing the playlist marks that result stale. Cancel preview stops an active request. Browsing, inspecting metadata, and restoring a workflow never encode a video.

Queuing always produces VIDEO, including the joined audio. For separate images or audio, connect Split Video or ComfyUI's Get Video Components.

OutputContentsvideoNative file-backed VIDEO with the chosen timing and soundtrack.generation_dataPrompt/sampling PIPE from the selected metadata source, or None.durationDuration of the joined VIDEO in seconds.reportPer-file media properties, ranges, A1111 text, parsing status, and result details.

Frame counts remain in the report. A count node can also read extracted images.

Upscale before saving

For long videos, connect:

Load Video → Upscale Video → Save Video with Generation Data

Upscale Video includes model selection, a bounded batch size, PyTorch support for non-NVIDIA devices, and optional TensorRT engines. Audio and source timing stay inside VIDEO. Connect generation_data directly to the saver's pipe_opt if wanted; leave audio and timing_opt disconnected.

For shorter videos that fit in RAM, connect video to Split Video, then:

• images → image upscaler → Save Video with Generation Data images / video

• audio → saver audio

• timing → saver timing_opt

• optionally, Load Video's generation_data → saver pipe_opt

Connected timing overrides the saver's FPS, trim, and loop settings; those controls are hidden while it is connected. This preserves mixed frame rates through a spatial upscale. Frame count and order must remain unchanged. The saver rejects a different count; it cannot detect a reorder with the same count. Selection, interpolation, duplication, or temporal trimming needs matching timing/audio changes and is not supported by this connection.

Split Video provides IMAGE, AUDIO, and timing outputs for VIDEO from this loader or ComfyUI's built-in nodes. It has one VIDEO input and no manual FPS or timing input. File-based videos retain their timestamps and native trim/crop views; component videos use their stored frame rate. Native concatenations are handled clip by clip. Unsupported third-party VIDEO representations report an error rather than guessing their timing. report describes the extraction. The splitter materializes the complete IMAGE batch; using it before an ordinary image upscaler does not provide streaming or solve the full-video RAM limit.

ComfyUI's Get Video Components also exposes images and audio, plus a single FPS value. It works for ordinary constant-rate workflows. For a join with mixed frame rates, use Split Video's timing connection to preserve the individual frame durations through an IMAGE processing pass.

A1111 generation metadata

Use metadata selects at most one included clip as the downstream source. Selecting another clears the previous choice. Preview selection is independent; duplicating a row does not copy this flag, and excluding its row clears it.

Only the container's A1111 parameters tag, as written by Save Video with Generation Data for Civitai, is read. The PIPE contains positive/negative prompts and available sampling settings: seed, steps, CFG, sampler, scheduler, denoise, CLIP skip, and dimensions. Model, VAE, LoRA, embedding, and hash fields are not restored. Workflow graphs and JSON sidecars are not read.

GenData previews the active clip's parsed prompts/settings. If the selected source has no usable A1111 tag, the PIPE is None; another clip is never selected automatically. The debug report includes bounded original A1111 text, which can contain resource names that were excluded from the PIPE.

Resource use and limits

VIDEO frames stream into a temporary H.264 MP4. Audio is held as joined PCM; the loader checks available RAM for audio and frame processing. Exact frame ranges require an initial timestamp scan. Cached scans and still-live media can be reused when only the metadata choice changes.

Extracting images with a component node requires a complete batch in RAM: approximately frames × width × height × 3 × 4 bytes before downstream upscaling.

Loading decodes SDR frames to 8-bit RGB. HDR and alpha inputs require explicit conversion/compositing first. Source timing is preserved, but compressed pixels are not bit-identical: VIDEO is encoded, and a saver can encode it again. Use the streaming Upscale Video route for longer videos. Joined previews also render at the configured size.

The initial scope is hard joins of up to 128 videos. It does not include transitions, motion interpolation, speed changes, multiple soundtrack mixing, remote URL downloads, or image slideshows. Original source files are unchanged.