Deprecated:
please use https://civarchive.com/models/1820946/wan2214bsage-torchcompile-llm-autoprompt-workflow

The Wan Img2Video MultiGPU workflow is a powerful and efficient method for generating videos from a single image using the Wan 2.1 model, a state-of-the-art video foundation model. This workflow is typically implemented within a ComfyUI framework, which provides a node-based interface for managing the various components of the video generation process. The multi-GPU aspect is crucial for handling the high computational demands of the larger Wan models, such as the 14B parameter version.

Here's a breakdown of the typical workflow and how multiple GPUs are leveraged:

1. Workflow Initialization and Data Loading:

• The process begins by loading the input image and the necessary models.

• Key components include the "Load image" node and the "Load WanVideo" nodes, which bring the image and video foundation models into the workflow.

• The "WanVideo Loader" and "WanVideo TextDecoder" nodes are used to load and configure the specific models, parameters, and LoRAs (if used).

2. Multi-GPU Distribution:

• To optimize performance and overcome VRAM limitations, the workload is distributed across multiple GPUs. This is where the multi-GPU workflow truly shines.

• Different components of the model can be offloaded to separate GPUs. For example:

  • GPU 1: Might be dedicated to loading the large diffusion model (the core of the Wan 2.1 model).

  • GPU 2: Could be used for the CLIP text encoder, which processes the text prompt that guides the video generation. This is a significant part of the workflow and can consume a substantial amount of VRAM.

  • GPU 3, 4, etc.: Additional GPUs can be used to handle other parts of the pipeline, such as the VAE (Variational Autoencoder) for encoding and decoding, or for specific sampling operations.

3. Video Generation Process:

• Prompt and Parameter Setting: Users provide a text prompt to guide the video's motion and content. They also set key video parameters like num_frames (video length) and frame_rate.

• Diffusion Process: The core of the generation is a diffusion process. The model starts with a latent space representation of the input image and progressively adds temporal information, guided by the text prompt. This is a highly parallelizable task, and the use of multiple GPUs allows different parts of this process to be handled concurrently.

• Temporal and Spatial Coherence: Wan 2.1 utilizes a novel 3D causal VAE architecture, which is specifically designed for video generation. It efficiently compresses spatiotemporal information, ensuring consistency across frames and preserving fine details.

• Video Synthesis: After the diffusion process is complete, the final frames are synthesized from the latent space and decoded into a video.

4. Key Benefits of the Multi-GPU Approach:

• Overcoming VRAM Limitations: The large-scale Wan 2.1 models (e.g., 14B parameters) can require significant VRAM (upwards of 20GB). Distributing the model's components across multiple GPUs makes it possible to run these models on systems that wouldn't be able to handle them with a single GPU.

• Faster Inference: By parallelizing the workload, the multi-GPU workflow significantly reduces the time it takes to generate a video. This is especially important for high-resolution, longer videos.

• Improved Quality: Using larger models and higher resolutions becomes more feasible, leading to higher-quality, more detailed, and more stable video outputs.