I2V Looper High/Low

What it is

This is the flagship of the Native Looper pack, and the node worth building around if you're on Wan 2.2. Here's the thing about 2.2: it doesn't denoise with a single model. It splits the job between two experts - a high-noise model that handles motion, scene composition, and camera, and a low-noise model that refines detail. ComfyUI ships them as separate checkpoints, and stock workflows make you chain two KSamplerAdvanced passes together by hand, every single generation.

I2V Looper High/Low automates that two-pass dance and folds it into the same last-frame-chaining loop as the plain I2V Looper. You get long, multi-segment video with Wan 2.2's sharp output - which is the pack's whole selling point - without manually babysitting a dual sampling setup per segment.

How it works

Same skeleton as I2V Looper: for each segment it encodes the prompt and CLIP Vision features of the current start image, builds WanImageToVideo conditioning, samples, saves the decoded segment to disk, clears VRAM, and promotes the last frame to the next loop's start image. The difference is inside the sampler. Instead of one KSampler pass, it runs two KSamplerAdvanced passes: the high-noise model from step 0 to split_step (with leftover noise kept), then the low-noise model from split_step to the end with add_noise disabled. The source code's own comment says it matches the "APP VIDEO" dual-pass approach - this is the standard Wan 2.2 recipe, just automated per segment.

Inputs that matter

• model_high / model_low - the two experts, tooltips literally reading "High-noise model" and "Low-noise model." The pack's workflows load the GGUF pair wan22I2VA14BGGUF_q8A14BHigh.gguf and ..._Low.gguf.
• split_step - default 4. The step where the high model hands off to the low model. This is the one you'll actually tune: set it too low and the low expert re-does everything (slow); too high and it barely participates. Keep it comfortably below steps.
• steps / cfg - defaults of 8 steps and cfg 1.0, and that cfg is not a typo. Wan 2.2 runs near CFG 1; crank it the way you would for SD 1.5 and you'll fry the image. The KB backs this up - flow-matching family wants low guidance on conservative schedules.
• clip_vision / start_image / width / height / sampler / scheduler / seed - same roles as the plain I2V Looper.
• loop_1 … loop_10 - the I2VLoopHL config cards. These take the I2V_LOOP_HL type, not the plain NATIVE_LOOP, so don't cross-wire the two config nodes.

Outputs

• full_video (IMAGE) - concatenated frames, one tensor. Wire it into VHS Video Combine (VideoHelperSuite) for an actual mp4; the pack's own workflows do exactly that, with RIFE interpolation on top in the "Detailed" one.
• used_prompts (STRING) - per-segment prompt log for debugging.

Install and the model stack

Same pack as the rest of Native Looper:

cd ComfyUI/custom_nodes
git clone https://github.com/masteroleary/ComfyUI-LooperNode

Restart, or install via Manager searching "Native Looper." No requirements.txt - ComfyUI core provides everything. The pack ships two example workflows, "I2V Looper HL Detailed" and "Preview," which are worth dragging in: they show the full stack, including wan_2.1_vae.safetensors, umt5_xxl_fp8_e4m3fn_scaled.safetensors (CLIP), clip_vision_h.safetensors, a ModelSamplingSD3 shift of 8, and - in the detailed one - SageAttention patches and a RIFE VFI upscaler. The Looper itself needs none of those extras; they're just what the author tuned around.

Common issues

• split_step at or above steps means the low-noise pass effectively never runs and you get soft, high-noise-only output. Keep it well below.
• VRAM stacks up: two models plus VAE plus CLIP Vision. GGUF Q8 helps a lot; on a 12GB card, drop the start image resolution - the node resizes to width/height anyway, so a smaller source image costs you little.
• It's brand new - a single commit at the time of writing, no community footprint. Thin code, mostly familiar Wan failure modes, but you're the pioneer. That's the honest trade for a tool this focused.