Wan 2.2 Continuation Record (Nukun)

When a Wan continuation loop runs, each iteration re-captions the current end frame, advances three seeds, and renders the next segment. If you don't record what happened each pass, you end up with eleven videos and no way to know which seed and prompt produced the fifth one. Nukun Wan 2.2 Continuation Record is the loop's memory: it appends one deterministic record - captions, prompts, all three seeds, and the iteration number - to a JSON log. It doesn't render anything. It exists so the run has a paper trail.

How it works

It's a bookkeeping node designed to sit inside the loop of the pack's anima_to_wan2.2_i2v style continuation workflow. You pass in the plan from NukunWan22ContinuationPlan, the current log_json (starting from the plan's empty_log_json, which is just []), the iteration index, and the three seeds the loop is using - vision_seed (for captioning the end frame), prompt_seed, and sampling_seed - plus the caption, positive, and negative for that segment. It returns an updated log_json with the iteration appended. The defaults for the three seeds (33004, 44005, 55006) line up with the example workflow's deterministic seed stepping, so a run is fully replayable from the log.

The inputs that matter

• plan - the continuation plan; keeps the record consistent with the planned frame budget.
• log_json - the running log. Start with [] (or the plan's empty_log_json), and thread the output back into the next iteration's input.
• iteration - 0–9, matching the plan's extension count ceiling.
• vision_seed, prompt_seed, sampling_seed - record what the loop actually used; this is what makes the whole run reproducible.
• caption, positive, negative - the per-segment text. These get written verbatim into the log.

One output: log_json. That's it. Feed the chain from iteration to iteration and it accumulates the full history.

The honest take

This is a deliberately thin node and that's fine - its job is to make the loop's invisible decisions visible. Without it, the continuation manifest at the end would only know the base run, not what each extension actually did. The single output is also a nice lesson in the pack's design: every one of these continuation nodes is a small, honest piece that only does its one thing, and the workflow glues them together. If you're hand-writing continuation loops instead of using the example workflow, this is the node that keeps your bookkeeping from drifting.

Installing it

Part of Nukun_ComfyUI_Nodes - ComfyUI Manager (search "Nukun") or:

cd ComfyUI/custom_nodes
git clone https://github.com/OnekoSL/Nukun_ComfyUI_Nodes

Restart ComfyUI. Dependencies are just numpy, Pillow, scipy, and PyWavelets. The usual failure mode is wiring the wrong log_json - make sure each iteration's output feeds the next one's input, or your log will only ever hold the last iteration.