Skip to content
The Chat Archive
sonar-pro-high LMArena Search Arena 24K

I cant seem to find any Wan FLF (Frame Last Frame) comfyui workflow for GGUF's.

  1. User

    I cant seem to find any Wan FLF (Frame Last Frame) comfyui workflow for GGUF's.

  2. Assistant

    Wan FLF (Frame Last Frame) ComfyUI Workflows for GGUF Models

    You can find and use Wan FLF (First-Last-Frame) workflows for GGUF models in ComfyUI, but they are relatively new and not as widely distributed as standard T2V/I2V workflows. Here’s what’s available and how to get started:


    Wan FLF GGUF Workflow Availability

    • OpenArt hosts a workflow titled "Wan First Last Frame Native Support" that specifically supports GGUF models for first-last-frame video generation. The workflow references the GGUF model hosted at city96/Wan2.1-FLF2V-14B-720P-gguf, confirming native GGUF support for FLF tasks[6].
    • YouTube tutorials and community guides (e.g., from Prompting Pixels and others) show how to set up GGUF-based workflows for Wan 2.1, including custom nodes and model placement, though most focus on T2V/I2V rather than FLF specifically[2][4][9].
    • Reddit and Wiki discussions confirm that GGUF support for advanced Wan workflows (including FLF and Start-End frame) is available, especially with the latest ComfyUI and custom node updates[5][7][10].

    How to Use Wan FLF GGUF Workflows in ComfyUI

    1. Download the Workflow and Model

    • Download the FLF workflow from OpenArt or similar repositories[6].
    • Download the corresponding GGUF model (e.g., Wan2.1-FLF2V-14B-720P-gguf) from Hugging Face[6].

    2. Install Required Custom Nodes

    • Install the ComfyUI-GGUF custom node via ComfyUI’s custom node manager. This enables loading GGUF models[2][4][8].
    • Ensure you have any other required nodes (e.g., for VAE, CLIP, and video helpers).

    3. Place Models in Correct Folders

    • Place the GGUF model in ComfyUI/models/diffusion_models/ or as specified by the workflow[5][7].
    • Place the text encoder, VAE, and CLIP Vision models in their respective folders as outlined in the workflow documentation[5][7].

    4. Load the Workflow in ComfyUI

    • Drag the workflow JSON into ComfyUI or use the workflow import feature.
    • The workflow should include nodes for:
      • Unet Loader (GGUF)
      • CLIP Loader (GGUF or safetensors)
      • VAE Loader
      • FLF-specific nodes (e.g., WanFirstLastFrameToVideo)
      • Image loaders for first and last frames
      • Video output nodes[6].

    5. Configure and Run

    • Load your first and last frame images.
    • Adjust prompt and settings as needed (e.g., frame count, resolution).
    • Run the workflow to generate the interpolated video.

    Example: Workflow Structure

    Node TypePurpose
    UnetLoaderGGUFLoads the Wan FLF GGUF model
    CLIPLoaderGGUFLoads the text encoder
    VAELoaderLoads the VAE
    LoadImageLoads first and last frame images
    WanFirstLastFrameToVideoHandles FLF video interpolation
    SaveAnimatedWEBP/VideoSaves the output video

    Additional Tips

    • Model Size: GGUF quantized models are available in various sizes (Q4, Q5, etc.), allowing you to balance quality and VRAM usage[5][8].
    • Prompt Engineering: For best results, use detailed prompts that describe both the transition and static elements between frames[9].
    • Hardware: Lower quantization (Q4/Q5) is suitable for lower VRAM GPUs, but higher quantization yields better quality if your hardware allows[8][9].

    References and Downloads

    • Workflow: Search OpenArt for "Wan First Last Frame Native Support" or similar[6].
    • Model: city96/Wan2.1-FLF2V-14B-720P-gguf[6].
    • Custom Node: ComfyUI-GGUF[8].
    • Community Guides: ComfyUI Wiki, YouTube tutorials, and Reddit threads for troubleshooting and setup[2][5][7][9][10].

    Summary:
    Wan FLF (First-Last-Frame) workflows for GGUF models in ComfyUI are available and actively maintained. Download the workflow and model, install the ComfyUI-GGUF node, and follow the workflow structure to generate FLF videos with GGUF efficiency[6][5][8].

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.