SV3D Batch Schedule (SV3D_BatchSchedule) - ComfyUI-KJNodes

The SV3D Batch Schedule node receives an initial image, extracts its visual features via a CLIP Vision encoder, and encodes it into the latent space using a VAE.

The SV3D Batch Schedule node is a conditional scheduling tool specifically designed for Stability AI's SV3D model. Its core function is to allow users to batch-generate a series of latent space representations with different camera viewpoint conditions (azimuth and elevation) for an input initial image, thereby laying the groundwork for subsequent generation of multi-view consistent 3D videos or image sequences.

Node Function

This node receives an initial image, extracts its visual features via a CLIP Vision encoder, and encodes it into the latent space using a VAE. Its core mechanism lies in allowing users to precisely define the camera azimuth and elevation values for each frame (or specific keyframes) within the generation batch through a specific string format. The node smoothly interpolates viewpoint parameters between these keyframes according to the specified interpolation method, ultimately outputting corresponding positive/negative conditioning prompts and a unified latent space representation for the entire batch.

Node Parameter Description - SV3D Batch Schedule

Connection Inputs (Inputs)

Parameter Name	Data Type	Required	Default Value	Range/Options	Description
`clip_vision`	CLIP_VISION	Yes	-	-	Input for the CLIP vision model used to encode the initial image, extracting its semantic features to generate conditioning.
`init_image`	IMAGE	Yes	-	-	The input initial reference image. The node will generate subsequent latent space representations from different viewpoints based on this image.
`vae`	VAE	Yes	-	-	Variational Autoencoder, used to encode the input initial image into the latent space, creating the starting point for the subsequent denoising process.

Control Parameters (Parameters)

Parameter Name	Data Type	Required	Default Value	Range/Options	Description
`interpolation`	COMBO	Yes	linear	"linear", "ease_in", "ease_out", "ease_in_out"	Selects the method for angle interpolation between keyframes. Options include "linear" or various easing functions, used to control the speed curve of viewpoint changes.
`width`	INT	Yes	576	16 - MAX_RESOLUTION (step: 8)	Sets the width of the output latent space image, ranging from 16 to the system's maximum allowed resolution, with an adjustment step of 8.
`height`	INT	Yes	576	16 - MAX_RESOLUTION (step: 8)	Sets the height of the output latent space image, ranging from 16 to the system's maximum allowed resolution, with an adjustment step of 8.
`batch_size`	INT	Yes	21	1 - 4096	Defines the size of the generated latent batch, i.e., the total number of frames to generate, ranging from 1 to 4096.
`azimuth_points_string`	STRING	Yes	`0:(0.0),\n9:(180.0),\n20:(360.0)\n`	-	Defines the keyframe schedule for azimuth (horizontal rotation angle) using a specific format string, e.g., "0:(0.0), 10:(90.0)".
`elevation_points_string`	STRING	Yes	`0:(0.0),\n9:(0.0),\n20:(0.0)\n`	-	Defines the keyframe schedule for elevation (vertical rotation angle) using a specific format string, e.g., "0:(0.0), 10:(30.0)".

Output

Parameter Name	Data Type	Description
positive	CONDITIONING	Output containing positive conditioning prompt information, used to guide the generation process.
negative	CONDITIONING	Output containing negative conditioning prompt information, used to constrain the generation process.
latent	LATENT	The encoded and processed batch of latent space representations, serving as input for subsequent samplers.

Usage Scenario

You can use this node when creating a 3D video loop that rotates around an object. First, load a single image of the object as init_image. Then, set keyframes from 0 to 360 degrees in azimuth_points_string, while keeping elevation_points_string at 0. After connecting the CLIP Vision and VAE models, the latent and conditioning information output by the node can be passed to a KSampler for denoising sampling, ultimately generating a smooth rotating sequence.

Notes

Please note that SV3D is fundamentally still a video model, so viewpoint scheduling must always progress forward along the timeline. The input azimuth and elevation string formats must strictly adhere to the frame_index:(angle_value) specification; otherwise, errors may occur.

SV3D Batch Schedule Node Source Code Link

The SV3D Batch Schedule node is from the ComfyUI-KJNodes node package.

SV3D Batch Schedule(SV3D_BatchSchedule)