Open-H3 IC · Prepare Draft (124f / 24fps):
OpenH3ICDraft prepares an existing video draft for the in-context, or IC, version of Open-H3-Regenerate-2K. It standardizes the draft’s timing, frame dimensions, and audio handling so the video can serve as a consistent reference for later regeneration. Instead of generating new scenes or increasing visual detail itself, it selects frames on a fixed 24 fps timeline, resizes the pictures, and prepares accompanying sound. Short drafts can be extended by repeating their final frame, provided they contain enough usable footage to pass validation. In the supplied workflow, this preparation supports both reference-video conditioning and video encoding before a separate latent-upscaling and regeneration stage. This helps you maintain a consistent starting point while leaving the actual high-resolution enhancement to the downstream H3 models.
Open-H3 IC · Prepare Draft (124f / 24fps) Input Parameters:
images
Required IMAGE input containing the source video frames as an ordered image batch, rather than a video-file path. The node treats the batch order as the video’s timeline and selects frames according to the source frame rate and the required 24 fps processing rate. It selects existing frames without generating intermediate motion frames; depending on the source frame rate, this can skip frames or reuse them. When the requested timeline extends beyond the available footage, the last source frame is repeated, but only if at least 113 requested frame positions are covered by the source. An empty batch causes an error. There is no numeric default, minimum, or maximum defined for this socket, although the timing and minimum-footage checks determine whether a particular batch is usable.
source_fps
Required FLOAT specifying the actual frame rate of the incoming image sequence. The default is 24.0, the minimum is 1.0, the maximum is 240.0, and the interface step is 0.01. This value tells the node how much time separates the source frames; it does not change the required 24 fps processing rate. For each output position, the node selects the source frame at the integer part of the output position multiplied by source_fps and divided by 24. A higher source frame rate therefore generally causes frames to be skipped, while a lower source frame rate can cause frames to be repeated. Set this accurately to preserve the intended motion speed and timing relative to the audio; the node does not infer it from the image batch.
target_width
Required INT setting the prepared draft’s width in pixels. The default is 1344, the minimum is 32, the maximum is 8192, and the interface step is 32. The width must be divisible by 32, including when supplied through a connected value rather than the widget; otherwise, execution stops with an error. Together with target_height and fit, it determines the draft’s canvas and how the source image is adapted to it. Larger dimensions require more image-processing memory and can increase downstream encoding costs. This is the preparation resolution, not necessarily the final regenerated video resolution: the supplied workflow prepares a 1344 × 768 draft and separately upscales its latent representation to 2560 × 1440.
target_height
Required INT setting the prepared draft’s height in pixels. The default is 768, the minimum is 32, the maximum is 8192, and the interface step is 32. Like target_width, it must be divisible by 32. Its relationship to the width determines the canvas aspect ratio, which affects cropping with cover or image distortion with stretch. Keeping the default 1344 × 768 canvas follows the supplied workflow’s draft-preparation settings. Increasing the height increases the number of pixels processed and can raise memory requirements, but this node does not add new visual detail through generative enhancement.
frame_count
Required INT setting the number of frames in the prepared draft. The default and minimum are 124, the maximum is 362, and the interface step is 17. Valid lengths follow the rule 5 + 17 × k, giving the interface sequence 124, 141, 158, 175, 192, 209, 226, 243, 260, 277, 294, 311, 328, 345, and 362. At the required 24 fps, the default 124 frames represent approximately 5.17 seconds. Longer footage is limited to the requested timeline; shorter footage is extended by holding its final frame, provided at least 113 requested positions still refer to available source frames. Increasing this setting increases the prepared sequence length and its memory requirements, but does not create new motion beyond the end of the source.
output_fps
Required INT identifying the prepared draft’s playback frame rate. The interface declares a default of 24, a minimum of 1, a maximum of 120, and a step of 1; however, the implementation explicitly rejects every value except 24 because Open-H3-Regenerate-2K-IC sampling is fixed at that rate. Leave this setting at 24 regardless of the original video’s frame rate, and use source_fps to describe the original timing. This setting determines frame selection and the duration used when preparing matched-length audio. The IMAGE output does not carry a separate frame-rate output socket, so keep downstream video creation configured to 24 fps as well.
fit
Required COMBO controlling how selected frames fit the requested canvas. The options are cover and stretch, with cover as the default. Cover uses the node’s cover-resize path to fill the canvas while preserving proportions, which can crop content when the source and target aspect ratios differ. Stretch resizes directly to the exact width and height using bilinear interpolation with antialiasing; it avoids cover-style cropping but can make subjects appear wider or taller if the aspect ratios differ. The stretch path uses the first three image channels and clamps pixel values to the 0–1 range. Choose cover when natural proportions matter most, and stretch when retaining the full composition is more important than preserving its original proportions.
audio_mode
Required COMBO controlling audio preparation, with options trim_pad, passthrough, and silent; the default is trim_pad. Trim_pad requests an audio length matching the prepared video duration, calculated from the output frame count and 24 fps using the supplied audio’s sample rate, so it is the intended choice for trimming or padding sound to the draft. Passthrough copies the supplied waveform without matching its duration, retains its sample rate, and duplicates a mono channel into two channels; despite its name, it is therefore not completely unchanged for mono audio. Silent ignores any supplied audio and requests a silent track at 32,000 Hz for the prepared duration. If no audio is connected, the node creates this duration-matched silent track regardless of the selected mode.
audio
Optional AUDIO input containing the source sound. The implementation expects an audio object with waveform and sample_rate entries, as provided by compatible ComfyUI audio or video-component nodes. Its treatment depends on audio_mode: it can be fitted to the prepared duration, copied without duration adjustment, or ignored in favor of silence. For passthrough, the waveform is copied as floating-point data, and mono sound is duplicated into two channels. If this input is omitted, the node still produces audio by creating silence at 32,000 Hz. There is no numeric default or range for this socket. For reliable synchronization, connect audio from the same source clip and time segment as the image batch.
Open-H3 IC · Prepare Draft (124f / 24fps) Output Parameters:
draft_frames
IMAGE output containing the selected and resized video-frame batch in playback order. For a successful execution, its frame count matches frame_count and its dimensions match target_width and target_height. The sequence is prepared for 24 fps playback, with frame selection based on source_fps rather than motion interpolation. If the requested sequence extends beyond the source, its remaining positions repeat the final available frame after the minimum-footage check passes. In the supplied workflow, this output branches to both MiniMax H3 Reference to Video conditioning and VAE encoding, ensuring that the reference and encoded draft start from the same prepared pictures.
draft_audio
AUDIO output containing the prepared waveform and its sample rate. Its meaning depends on audio_mode: trim_pad prepares sound for the draft’s duration, passthrough retains the supplied audio length and sample rate while duplicating mono audio into two channels, and silent produces silence. Missing source audio also produces silence at 32,000 Hz. This output lets you carry sound alongside the prepared visual sequence and reconnect it during final video creation, as the supplied workflow does. With passthrough, check synchronization and duration yourself because the sound can be longer or shorter than the prepared frames.
report
STRING output containing a JSON-formatted summary of the preparation. Its fields are source_frames, the incoming image count; source_fps, the source rate used for timing; real_frames_at_24fps, the number of requested positions that fall within the source before final-frame padding; output_frames, the prepared image count; output_fps, the required rate of 24; canvas, the width and height in that order; audio_samples, the final waveform length in samples; and audio_rate, its sample rate in Hz. The real_frames_at_24fps value counts valid requested positions, not necessarily distinct source images, because lower-frame-rate material can reuse the same source frame. This report is useful for checking timing, confirming dimensions, and identifying whether final-frame padding occurred.
Open-H3 IC · Prepare Draft (124f / 24fps) Usage Tips:
-
Start with the supplied workflow’s preparation settings: source_fps matching your clip, target_width 1344, target_height 768, frame_count 124, output_fps 24, fit cover, and audio_mode trim_pad. These prepare the draft; the separate latent upscaler controls the final enhancement resolution.
-
Connect frames, audio, and source frame-rate information from the same decoded video whenever possible. In the example workflow, a video-components node supplies all three, reducing the risk of mismatched timing.
-
Keep output_fps at 24. For a 30 fps source, set source_fps to 30 rather than changing output_fps; the node will select source frames for a 24 fps timeline.
-
Supply enough footage to cover at least 113 requested positions at 24 fps, roughly 4.7 seconds. Ideally, provide enough footage for the entire requested sequence so the ending does not become a held still frame.
-
Use cover for natural-looking proportions, then inspect subjects near the frame edges for cropping. Use stretch only when keeping the entire source image is worth possible distortion.
-
Use trim_pad when you want sound prepared for the draft duration. Choose passthrough only when you deliberately want to retain the source audio length, and check the result when assembling the final video.
-
Inspect report with a compatible text-display node. If real_frames_at_24fps is smaller than output_frames, the ending includes repeated copies of the last source frame.
-
Increase frame_count only when your downstream H3 workflow supports the intended length and you have enough memory. Every additional frame increases image-processing and subsequent video-model workload.
Open-H3 IC · Prepare Draft (124f / 24fps) Common Errors and Solutions:
The draft video contains no frames
- Explanation: The images batch contains zero frames, so the node has no visual material to prepare.
- Solution: Check that the source video loads successfully and that any frame-range selection includes actual footage. Connect the decoded IMAGE batch rather than a video-file path.
target_width and target_height must be multiples of 32
- Explanation: At least one requested dimension is not divisible by 32. Connected numeric values can trigger this even if the standard widgets normally advance in valid increments.
- Solution: Use dimensions such as 1344 × 768 or other multiples of 32 within the declared range. Check both width and height, including values supplied by upstream nodes.
Open-H3-Regenerate-2K-IC sampling is fixed at 24 fps; set output_fps=24
- Explanation: output_fps is not 24. Although the widget declares a wider range, this node’s processing explicitly requires 24 fps.
- Solution: Set output_fps to 24 and enter the original clip’s actual frame rate in source_fps. Also configure downstream video assembly for 24 fps.
frame_count must be a MiniMax-H3 valid length: 5 + 17*k (124, 141, ...)
- Explanation: The requested length does not follow the required frame-count pattern.
- Solution: Choose 124, 141, 158, or another valid value in increments of 17 up to 362. If a connected number overrides the widget, correct that upstream value too.
The draft yields only N real frames at 24 fps. Open-H3-Regenerate-2K-IC needs at least 113 (~4.7 s).
- Explanation: N represents the number of requested timeline positions covered by the source, and that number is below 113. The clip may be too short, too few frames may have been loaded, or source_fps may be incorrect. Final-frame repetition is not allowed to compensate for failing this minimum.
- Solution: Load a longer segment or more source frames, and verify source_fps against the actual source rate. Do not lower source_fps merely to bypass the check, because that changes playback timing.
KeyError: 'waveform' or KeyError: 'sample_rate'
- Explanation: These Python errors can occur if connected audio lacks fields expected by the implementation. The node needs a compatible AUDIO object, not a filename or an arbitrary data object.
- Solution: Use audio from a compatible ComfyUI loader or video-components node. If you do not need the source sound, disconnect audio or select silent.
IndexError or RuntimeError while processing images or audio
- Explanation: The exact message can vary, but malformed tensor dimensions can cause indexing, resizing, or channel-handling operations to fail. For example, the node expects an image batch rather than a single unbatched image, and passthrough audio handling expects a waveform with batch, channel, and sample dimensions.
- Solution: Use standard ComfyUI IMAGE and AUDIO outputs directly from compatible nodes. If you are using custom conversion nodes, check that they preserve the expected batch and channel structure.
CUDA out of memory
- Explanation: This possible runtime error means there is insufficient GPU memory for the selected frame batch or processing operation. Large canvases and long sequences increase memory demand; it is not a custom validation message from this node.
- Solution: Return to 1344 × 768 and 124 frames, avoid unnecessarily large source batches, and free memory used by other workflows. If the failure occurs later during H3 regeneration, reduce that stage’s target resolution or model-memory requirements rather than assuming draft preparation caused it.
