Wan Chunked I2V Sampler ♾️

Wan's native context is 81 frames - about five seconds at 16fps - and anything longer used to mean hand-chaining samplers: generate a chunk, decode it, grab the last frame, feed it back in as the reference, repeat, stitch. Tedious, error-prone, and the frame-count arithmetic (the 4n+1 thing) would bite you at the end. This node collapses that whole loop into a single box. You set total_frames and chunk_frames, connect your model, VAE, and start image, and it internally runs encode → sample → decode → extract reference → repeat until the timeline is done, then concatenates everything into one continuous video.

It's the flagship of the VideoChunkTools pack, and it fits a real 2026 pattern: the "chunking moved into a single node" trend (SCAIL-2 Infinity, Bernini Infinity, scail-auto-extend all did the same thing for their models). The honest caveat that applies to all of them applies here too - the loop is hidden, the 81-frame window isn't, and identity drift across chunk boundaries is still the failure mode. What you get is ergonomics, not magic. But it's a big ergonomics win.

How it works

Under the hood it calls Kijai's WanVideoWrapper nodes (WanVideoEncode, WanVideoSampler, WanVideoDecoder) from ComfyUI's global node registry - lazily, so the pack still loads even if you haven't installed the wrapper. The join strategy is the "concat approach": because each chunk is conditioned on the previous chunk's last frame as its reference, the new chunk's first frame matches it almost exactly, so the sampler just drops that first frame and concatenates. No blending, zero artifacts from a dissolve you don't need.

The inputs that matter

• total_frames / chunk_frames - your target length and the window size. Chunk frames auto-normalize to Wan's 4n+1 rule (81 → 81, 82 → 81), so you can't trip the trap. Output may overshoot to fill the last chunk, then gets trimmed.
• start_image - the reference for chunk 1.
• steps (default 8), cfg (default 2), shift (default 11), scheduler (default euler) - the standard Wan sampling knobs, in line with the distilled/LoRA-friendly settings the community runs.
• seed - one seed for every chunk. Variation comes from the changing reference images, not from reseeding. Set it once and keep it if you're iterating.
• force_offload (default on) - offloads the model to CPU between sampling passes. Keep it on unless you're swimming in VRAM; Wan is a 14B model and offloading is how mid-range cards survive multi-chunk runs.

The interesting optional inputs: model_b + split_step + cfg_b give two-pass sampling (first model handles steps 0→split, second finishes - the classic way to put a speed LoRA on the low-noise pass only). end_image enables FLF (first-last-frame): one image pulls the final frame toward a target, or a batch of images distributes as keyframes across the last chunks. end_blend_chunks (try 9–17) crossfades the FLF boundary instead of hard-cutting. And text_embed_sequence from ChainTextEmbeds swaps in per-chunk prompts, which overrides the uniform text_embeds.

Outputs

video (the joined IMAGE), total_frames (INT), and chunks_generated (INT) - the last is handy for knowing whether your prompt chain actually covered every chunk.

Installing

This node needs ComfyUI-WanVideoWrapper:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper.git
cd ComfyUI/custom_nodes
git clone https://github.com/gregtee2/ComfyUI_VideoChunkTools.git

Restart ComfyUI. The pack itself has no pip requirements; the wrapper is the heavy one (and brings the model files - Wan 2.2 14B needs serious VRAM, 12GB+ with GGUF or offloading, or use the 5B on 8GB). If you haven't installed the wrapper, the sampler will fail with a clear "requires ComfyUI-WanVideoWrapper" error while the rest of the pack keeps working.

Troubleshooting: if your first run OOMs, drop resolution to 832×480, enable force_offload, and consider enable_vae_tiling (it costs you some decode quality). If seams appear between chunks, that's the drift problem, not the concat - per-chunk prompting and a well-chosen end_blend_chunks help more than any blend curve will.