Kandinsky S2V Encoder Loader:
KandinskyS2VEncoderLoader prepares the speech and text encoder used by Kandinsky S2V 1.0 for speech-driven video generation in ComfyUI. It loads the Qwen2.5-Omni Thinker component, which the workflow uses to interpret speech and text as guidance for generating video, while leaving the Talker component on disk rather than loading it for this task. This keeps the loading focused on the component needed for conditioning instead of the separate speech-generation component. The node also integrates the encoder with ComfyUI’s model-management system, using its configured text-encoder loading and offloading devices to support memory management. It is a preparation step: it does not generate video or encode your speech and prompt by itself. CLIP-L, which supplies another part of the workflow’s text guidance, is loaded separately through ComfyUI’s standard Load CLIP node.
Kandinsky S2V Encoder Loader Input Parameters:
omni
A required dropdown selection of type COMBO[STRING] that identifies the Qwen2.5-Omni model directory to load. The node discovers suitable directories in ComfyUI’s configured text-encoder locations, with the interface specifically describing a Qwen2.5-Omni-7B directory under models/text_encoders. Your selection determines which local model directory is used to construct the Kandinsky S2V speech and text encoder; it is a model-location choice, not a control for animation strength, prompt influence, or video quality. Options are populated dynamically from entries accepted by the extension’s Omni-directory check. If no suitable entries are found, the dropdown falls back to the single label Qwen2.5-Omni-7B; that fallback does not mean the model has been installed or will be downloaded automatically. The source does not specify an explicit default selection beyond this fallback behavior, and numerical minimum and maximum values do not apply.
Kandinsky S2V Encoder Loader Output Parameters:
encoder
An output of type KANDINSKY_S2V_ENCODER containing the prepared speech and text encoder wrapped in ComfyUI’s model-management interface. Connect it to the encoder input of Kandinsky S2V Encode (speech + prompt), where it is used alongside your audio, prompt, and separately loaded CLIP-L to prepare guidance for the video-generation process. This output is a reusable model object, not an image, audio clip, video, or already-computed conditioning result. Its model-management wrapper associates the encoder with ComfyUI’s configured text-encoder loading and offloading devices so that downstream processing can work within ComfyUI’s memory-management system.
Kandinsky S2V Encoder Loader Usage Tips:
- Place the complete Qwen2.5-Omni-7B model directory under
ComfyUI/models/text_encodersor another configured text-encoder location. For example,ComfyUI/models/text_encoders/Qwen2.5-Omni-7Bmatches the directory layout described by the node’s tooltip. - Select a discovered model directory rather than assuming the fallback
Qwen2.5-Omni-7Bentry confirms a successful installation. If no model directories are discovered, check your files and folder placement before running the workflow. - Connect
encoderto Kandinsky S2V Encode (speech + prompt). This loader only prepares the model; the encoding node is where your speech and prompt are processed into conditioning. - Load CLIP-L separately using the standard Load CLIP node with
clip_l.safetensorsand typestable_diffusion, then connect it to the encoding node’sclipinput. Qwen2.5-Omni does not replace this separate requirement. - If memory is limited, review ComfyUI’s device and memory-management configuration and close other GPU-heavy applications. This node follows ComfyUI’s text-encoder device settings and exposes no precision, quantization, or device controls of its own.
- After adding a model directory, refresh the node’s available options or restart ComfyUI if the directory does not appear. Keep the model’s supplied directory structure intact rather than selecting an arbitrary individual weight file.
Kandinsky S2V Encoder Loader Common Errors and Solutions:
Model directory not found or unavailable
- Explanation: The selected model directory may have been moved, renamed, or placed outside ComfyUI’s configured text-encoder locations. The fallback
Qwen2.5-Omni-7Blabel can also appear when no suitable directory was discovered. The supplied loader code does not define an exact error message for this situation; the wording depends on the path-resolution or loading code. - Solution: Confirm that the model directory exists under
models/text_encodersor another configured text-encoder location. Refresh the available options or restart ComfyUI, then select the discovered entry again.
Model files are missing, incomplete, or cannot be loaded
- Explanation: Finding a directory does not necessarily guarantee that every file needed to initialize the encoder is present and readable. An incomplete download or damaged model files may cause a loading failure. Exact messages come from the underlying encoder-loading code and are not shown in the supplied source.
- Solution: Check that your Qwen2.5-Omni-7B download is complete and that its original configuration and model files remain together in the model directory. Restore or download missing files and inspect the ComfyUI console for the specific file named in the error.
CUDA out of memory
- Explanation: This is a possible runtime error when there is insufficient GPU memory for loading or using the encoder. It is not a custom error defined by this node, and its occurrence depends on your hardware, ComfyUI settings, and other models or applications using memory.
- Solution: Close unnecessary GPU applications and review ComfyUI’s supported low-memory or offloading configuration. Restart ComfyUI if needed to clear a failed session. There is no memory-size or precision setting on this loader itself.
Encoder output cannot connect to the destination input
- Explanation: The output uses the custom KANDINSKY_S2V_ENCODER type. It cannot substitute for a standard CLIP, MODEL, or CONDITIONING connection, even though these objects all participate in a generation workflow.
- Solution: Connect
encoderto the matchingencoderinput on Kandinsky S2V Encode (speech + prompt). Use separate loader nodes for CLIP-L and the diffusion model.
Kandinsky S2V needs CLIP-L alone: load clip_l.safetensors with Load CLIP, type stable_diffusion
- Explanation: This exact message is defined in the surrounding encoding code, not in the encoder loader. It indicates that the separately supplied CLIP object does not match the required CLIP-L setup; it does not by itself indicate a problem with your Qwen2.5-Omni directory.
- Solution: Set the standard Load CLIP node to
clip_l.safetensorswith typestable_diffusionand connect it to the encoding node’sclipinput. Keep this loader’sencoderconnection in place.
