ComfyUI-SCAIL2-LongVideoContext

Smart long-video context window nodes for SCAIL-2 workflows in ComfyUI.

This custom node pack adds two model patch nodes:

• SCAIL-2 Smart Long Video Context
• SCAIL-2 Long Video Context

It is designed for SCAIL-2 / Wan long-video workflows that need temporal context windows, overlap handling, and 4n+1 frame alignment.

What it does

The node patches the incoming MODEL with ComfyUI's comfy.context_windows.IndexListContextHandler. It automatically converts pixel frame counts to latent context lengths, clamps frame counts to the 4n+1 pattern, and can preserve the reference condition index through cond_retain_index_list.

It does not replace the official SCAIL-2 nodes such as WanSCAILToVideo, SCAIL2ColoredMask, replacement_mode, previous_frames, or video_frame_offset. It is an additional model-context patch that can be used before sampling.

Installation

Clone this repository into your ComfyUI custom_nodes directory:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/1GirlUniversity/ComfyUI-SCAIL2-LongVideoContext.git

Restart ComfyUI.

Requirements

• Recent ComfyUI build with comfy.context_windows support.
• SCAIL-2-capable ComfyUI core if you are using the official SCAIL-2 workflow.
• No extra Python package is required by this node pack.

Nodes

SCAIL-2 Smart Long Video Context

Automatic mode. Inputs:

• model
• total_frames
• strategy
• manual_segment_len
• manual_segment_overlap
• context_schedule
• context_stride
• fuse_method
• freenoise
• causal_window_fix
• bypass_when_single_pass
• cond_retain_index_list

Strategies:

• auto_smooth_rtx6000
• auto_safe
• auto_quality_rtx6000
• force_official_81_5
• manual

SCAIL-2 Long Video Context

Manual mode. Inputs:

• model
• segment_len
• segment_overlap
• context_schedule
• context_stride
• fuse_method
• freenoise
• causal_window_fix
• cond_retain_index_list

Recommended workflow placement

Place the node after the diffusion model loader and before the sampler:

Diffusion Model Loader
        ↓
SCAIL-2 Smart Long Video Context
        ↓
Sampler

For total_frames, use the same effective frame count that goes into WanSCAILToVideo length, preferably adjusted to 4n+1:

frame_count -> a - ((a - 1) % 4) -> total_frames / length

Initial settings

Good starting point for high-VRAM NVIDIA GPUs:

strategy: auto_smooth_rtx6000
causal_window_fix: true
freenoise: false
cond_retain_index_list: 0

If you get OOM or excessive slowdown:

strategy: auto_safe

If you want conservative official-style chunking:

strategy: force_official_81_5

License

GPL-3.0