WanVideo Looper Lora Sequencer

Here's the thing the main looper node can't do alone: it runs every segment with the same models. If your video is "woman walks in, woman turns into a dragon, dragon flies off," you want the dragon shots to carry a different LoRA than the human ones - and WanVideo Looper Lora Sequencer is how you get that per-segment control.

The pattern is deliberately ComfyUI-native: you don't do any patching inside this node. You build your patched models the normal way - Load LoRA (or whatever you're using) wired into the base Wan 2.2 high-noise and low-noise models - and you connect the finished MODEL and CLIP outputs into this node's inputs. It then bundles them into a sequence the looper understands and hands them over via its single output, model_clip_sequence, which plugs into the looper's optional input of the same name.

How it works

Under the hood it's a list assembler with a friendly face. It exposes ten segments' worth of inputs - model_high_1 through model_high_10, model_low_1 through model_low_10, and clip_1 through clip_10, all optional - and packs each set into a tuple (model_high, model_low, clip) inside a list. The looper indexes that list by segment number: segment 1 uses what you plugged into _1, segment 2 what you plugged into _2, and so on. Any slot you leave unconnected falls back to the looper's clean base models, which is exactly the behavior you want - patch only the shots that need it.

The one genuinely clever bit is the UI. A little JS watches what's connected and only shows inputs up to the highest connected segment, so instead of a wall of thirty ports you get a tidy node that grows as you plug things in. Connect segment 1, you see segment 2's inputs appear; connect those, and segment 3 shows up. It's a small thing, but it makes a 30-input node readable.

Where people get burned

• It only covers the first 10 prompts. The sequencer hard-stops at segment 10. If your prompt_list has 14 lines, segments 11–14 silently run on the base models. The looper logs which models each segment is using, so glance at the console if a late segment comes out looking like it lost its LoRA.
• The looper still needs its own base inputs. model_clip_sequence overrides per segment, but model_high, model_low, and clip on the looper are still required and act as the fallback - leave them wired even when the sequencer is connected, or the whole thing refuses to run.
• Bypassed or wrong-typed connections get ignored. The looper checks that model_clip_sequence is actually a list and quietly ignores it otherwise (it logs a warning). If your per-segment LoRAs aren't applying, that's the first thing to check.
• Wan 2.2 LoRA basics still apply. Speed LoRAs (lightx2v, CausVid) are the one place to be careful: the community consensus is that they degrade motion and detail on the high-noise pass. If you're sequencing speed LoRAs, keep them off the motion model - this node lets you be that precise, which is rather the point.

Install with the rest of the pack: ComfyUI Manager search "WanVideoLooper", or git clone https://github.com/SquirrelRat/WanVideoLooper into custom_nodes/, restart. No models, no extra Python - it's pure routing between the LoRA nodes you already use and the looper. For anyone trying to hold character identity across a multi-shot Wan video, this is the missing piece between "one LoRA for everything" and "render each shot separately by hand."