ComfyUI Node: UniVidX • Load Model

UniVidX • Load Model:

The UniVidXLoader is a specialized node designed to facilitate the loading of UniVidX models, specifically the univid_intrinsic.safetensors or univid_alpha.safetensors. Its primary function is to output an opaque tuple, referred to as UNIVIDX_MODEL, which consists of a model instance and a variant name. This tuple is crucial as it seamlessly integrates into the UniVidXSampler, enabling efficient model sampling. The loader is optimized to cache models based on specific parameters such as variant, checkpoint, device, and data type, ensuring that multi-graph sessions can reuse weights without the need to reload large models, such as the 28 GB DiT. This caching mechanism significantly enhances performance by reducing load times and resource consumption, making it an essential component for AI artists working with complex model architectures.

UniVidX • Load Model Input Parameters:

vram_buffer_gb

The vram_buffer_gb parameter specifies the amount of VRAM, in gigabytes, that should be kept free for activations, KV cache, and VAE decode. This parameter is crucial for managing VRAM efficiently, as it determines how much memory is reserved for these processes. The default value is 4.0 GB, with a minimum of 0.0 GB and a maximum of 96.0 GB, adjustable in 0.5 GB increments. A higher value provides more headroom for operations but may result in slower performance due to increased streaming of data between CPU and GPU. Conversely, a lower value allows for more data to reside in memory, potentially improving speed but risking VRAM overflow. This parameter is particularly important for optimizing performance on different hardware configurations, such as 32 GB cards running BF16 Wan2.1-14B with UniVidX LoRAs.

UniVidX • Load Model Output Parameters:

UNIVIDX_MODEL

The UNIVIDX_MODEL output is an opaque tuple consisting of a model_instance and a variant_name. This output is integral to the UniVidX pipeline as it provides the necessary model data for subsequent processing stages, particularly in the UniVidXSampler. The model_instance represents the loaded model ready for use, while the variant_name identifies the specific variant of the model being utilized. This structured output ensures that the model is correctly identified and processed in the pipeline, facilitating seamless integration and execution of tasks.

UniVidX • Load Model Usage Tips:

• Adjust the vram_buffer_gb parameter based on your hardware capabilities to optimize performance. For systems with ample VRAM, increasing this value can provide more headroom for complex operations.
• Utilize the caching mechanism by maintaining consistent parameters across sessions to benefit from reduced load times and resource usage.

UniVidX • Load Model Common Errors and Solutions:

Model Load Failure

• Explanation: This error occurs when the specified model file cannot be found or loaded, possibly due to incorrect file paths or missing files.
• Solution: Ensure that the model files univid_intrinsic.safetensors or univid_alpha.safetensors are correctly placed in the expected directory and that the file paths are accurate.

Insufficient VRAM

• Explanation: This error arises when the VRAM allocation is insufficient for the model's requirements, leading to memory overflow.
• Solution: Increase the vram_buffer_gb parameter to allocate more VRAM for the model's operations, or reduce the complexity of the tasks being performed.