Extensions/ComfyUI-Wan-SVI2Pro-FLF
ComfyUI Extension

ComfyUI-Wan-SVI2Pro-FLF

ComfyUI nodes for Wan 2.2 SVI 2 Pro with Keyframe control via First/Last Frame and seamless video stitching.

By Well-Made·Created 6 months ago·Updated 4 months ago· 84
Well-Made/ComfyUI-Wan-SVI2Pro-FLF
Nodes2
On cloudLocal install
Categorylatent/video, conditioning/video_models
Stars84
Updated4 months ago

Nodes (2)

WanCutLastSlot

The one-input trim that saves your stitch

latent/video
WanImageToVideoSVIProFLF

SVI gives you length. This node gives you a destination.

conditioning/video_models
Readme

ComfyUI-Wan-SVI2Pro-FLF

Custom nodes for ComfyUI that combine SVI 2 Pro motion continuity with Wan 2.2 First/Last Frame (FLF) style control over the end of a clip, enabling smooth video generation from a sequence of frames. The included workflows are set up for up to 7 clips, but that's not a hard limit. To extend further, just duplicate a generation group as many times as needed and connect the inputs/outputs the same way as the existing groups.

Feb 20, 2026 - small update: the end_samples input is now optional. Disconnect it and the node falls back to regular SVI 2 Pro behavior. This means you can connect or disconnect intermediate images at any point to give yourself more flexibility during generation. Just remember to plug it back in when you need that last-frame pull.

Feb 20, 2026 - update №2: Added a workflow where you can load an existing video and continue from there either by providing the next last frame, typing a prompt, or both. If you want to continue with prompt only, disconnect the end_samples input in the WanImageToVideoSVIProFLF node and bypass the End Samples subgroup. You may skip the next image, although you might need to manually provide the anchor image to the subgraph node (simply connect the First_Frame output of the Video Preprocess node to the First_Frame input of your current segment or plug in any other image that fits your generation plan).

Feb 21, 2026 - fixed a tricky bug that could sometimes cause generation to completely ignore the Last Frame. If you downloaded the workflow and node earlier, worth updating, your issue might already be resolved.

Mar 02, 2026 - Added a switch to enable/disable end_samples (Last Frame) directly, without manually connecting/disconnecting inputs. This toggles the First/Last Frame functionality (feature suggested by javawock7618).

Important:

  • Sometimes the VAE decoder/encoder splits latents inconsistently, and this isn't easily controllable. Because of that, two clips may not align perfectly at the seam. If you notice ghosting or doubling at the transition, there's an overlap slider in the Image Batch Extend With Overlap node (right side of the subgraph). Adjust it by one or two steps to get a clean stitch. Default settings usually work, but not always.
  • If your generated frames look washed out, right before the Image Batch Extend With Overlap node there's an Adjust Contrast node. It's enabled in the Images-to-Video workflows but set to 1.0 in the Video Extension workflow. Bump it slightly - 1.2 usually does the trick, or adjust to taste.
  • If your generation isn't following the prompt, try increasing the clip length to 81–101 frames and see if that makes a difference. Since generation starts by continuing the previous motion, it needs time to adapt to a new direction. And if the model realizes there isn't enough room left to perform the action, it may drop it entirely. So either increase the frame count, or rephrase the prompt to trigger the action earlier. Instead of "starting to do X", try "doing X" - this can nudge the model to begin the action sooner (though it's not guaranteed).

Examples

svi-2-pro-with-frame-to-frame-stitching-v0-38vs6y4897kg1

https://github.com/user-attachments/assets/90e01396-6885-4fec-99e8-91ff6c72a019

<img width="1655" height="184" alt="image" src="https://github.com/user-attachments/assets/8d663756-a45d-434f-8598-ef8a60818ff9" />

https://github.com/user-attachments/assets/d519b0a4-3fad-437b-8914-5c282cf8e498

<img width="2641" height="352" alt="image" src="https://github.com/user-attachments/assets/d9b0917e-f310-46e7-b9a3-31b196ec94ce" />

https://github.com/user-attachments/assets/a0da3295-3da0-439f-9e4c-c72bfa49bafa

Models

This should work with your current models that already handle standard SVI 2 PRO, but I can't test every single one. Start with your own model combo, and if the results aren't what you expect, use the finetuned model this setup was built around until a more universal solution appears. Here are two options with speed LoRA baked in:

Workflows

This repository includes two example workflows for stitching video segments from images:

  • One uses KSampler (Advanced).
  • The other uses SamplerCustomAdvanced.

They are intended to be equivalent in this context and are provided so you can pick whichever sampler node better fits your existing setup or personal preference.

  • The third workflow extends existing video by specifying subsequent target frames or prompts to guide the continuation.

Installation

  1. Go to your ComfyUI custom_nodes folder, for example:

    cd ComfyUI/custom_nodes

  2. Clone this repository:

    git clone https://github.com/Well-Made/ComfyUI-Wan-SVI2Pro-FLF.git

  3. Restart ComfyUI.

Alternatively, search for Wan-SVI2Pro-FLF in ComfyUI Manager, click Install, and restart ComfyUI.

The nodes will appear under:

conditioning/video_models → Wan SVI 2 Pro FLF

latent/video → Wan Cut Last Slot

Nodes

Wan SVI 2 Pro FLF

Class: WanImageToVideoSVIProFLF
Category: conditioning/video_models

Combines:

  • Start: Wan SVI 2 Pro–style motion continuity (like WanImageToVideoSVIPro), using:
    • anchor_samples as the anchor latent(s) for the current segment.
    • An optional tail from prev_samples (last motion_latent_count temporal slots) to continue motion between segments.
  • End: FLF-style control (like WanFirstLastFrameToVideoLatent / FLF2V):
    • The last temporal slots are hard-locked to end_samples using concat_mask.

This gives you:

  • Strong content anchoring at the beginning of each segment.
  • Smooth motion continuity between segments via prev_samples.
  • Deterministic control over the last frames of each segment via end_samples.

Inputs:

  • positive / negative – standard ComfyUI conditionings.
  • length – target video length in frames. Wan video uses stride 4, so this is converted to latent T.
  • prev_samples – latents from the previous segment (B,C,T,H,W). The last motion_latent_count slots are used as motion tail.
  • anchor_samples – anchor latent(s) for the current segment (B,C,T,H,W), usually the first frame or a short clip.
  • end_samples – optional target end latent(s) (B,C,T,H,W). The last T_end slots are copied into the tail of the segment and hard-locked.
  • motion_latent_count – how many last temporal slots to take from prev_samples (0 disables motion continuity; 1–2 is typical).

Outputs:

  • positive / negative – conditionings with concat_latent_image and concat_mask injected.
  • latent – empty latent ({"samples": zeros}) for Wan’s video sampler.

Typical usage:

  • First segment:
    • anchor_samples = first frame latent, end_samples = last frame latent.
  • Subsequent segments:
    • prev_samples = latents from previous segment,
    • anchor_samples = anchor for the new segment,
    • end_samples = new target last frame,
    • motion_latent_count = 1–2.

Wan Cut Last Slot

Class: WanCutLastSlot
Category: latent/video

Utility node that trims temporal slots from the end of a Wan video latent clip.

In Wan 2.2, one temporal slot corresponds to 4 frames (stride = 4). Cutting 1 slot effectively removes the latent information for the last ~4 frames.

The main motivation is to avoid a conflict between hard‑locked last frames (FLF logic) and SVI Pro motion continuity. When the very last slot is fully fixed by end_samples, SVI still tries to “steer” the motion trajectory through that slot. These two behaviors can clash and produce heavy artifacts or “crumpled” frames exactly at the hard‑locked end. The simplest and most robust workaround is to drop that last temporal slot entirely before using the segment as prev_samples for the next one.

Inputs:

  • latents – Wan video latents ({"samples": tensor[B,C,T,H,W]}).
  • slots_to_cut – how many last temporal slots to cut (T dimension). At least one slot is always preserved internally.

Output:

  • trimmed_latents – same structure as input, but with reduced T.

Typical usage:

  • After generating a segment, cut 1 temporal slot from the end before feeding it as prev_samples to the next WanImageToVideoSVIProFLF node. This removes the problematic “hard‑locked + SVI” slot and makes segment stitching more stable.

This is a pragmatic workaround, not a theoretical limitation. If you ever come up with a cleaner way to reconcile SVI 2 Pro motion with fully hard‑locked end slots, you’re very welcome to share your findings with the community.

License

This project is licensed under the GNU GPLv3.

Some logic is adapted from:

ComfyUI (WanFirstLastFrameToVideo)

ComfyUI-KJNodes (WanImageToVideoSVIPro)

Both upstream projects are licensed under GPLv3, and this repository follows the same license.

Related resources