TIDE WAN High-Resolution Extrapolation

WAN 2.2 is where local video generation actually consolidated, but the checkpoints are trained on a 720p (or 480p) native budget. Push your frames to 1080p or beyond and the spatial coherence starts to wobble - motion still works, but the per-frame detail goes soft and structures drift. TIDE WAN High-Resolution Extrapolation is the video sibling in the same pack as the Flux TIDE node, and it attacks one specific cause: attention that's too diffuse when the model is extrapolating beyond what its RoPE positions were trained for.

It's a focused port of the TIDE paper's Dynamic Temperature Control (the same step-aware attention sharpening idea behind the Flux node) applied to WAN 2.1 / 2.2-style video DiTs. Attention gets sharper early in denoising, when the global structure of the video is being decided, and relaxes toward the end as fine detail takes over. The one thing this node deliberately doesn't do is Text Anchoring - and it's worth understanding why, because it tells you something about the architecture. WAN conditions text through separate cross-attention, not Flux-style joint text/image attention. There are no text keys competing with video keys inside one softmax, so a positive bias on text keys would just be cancelled by softmax shift invariance. It would literally do nothing.

How it actually works

Under the hood the node clones your MODEL, installs a ComfyUI WrappersMP.DIFFUSION_MODEL wrapper, and lazily wraps the live WAN model's rope_encode and forward_orig paths. WAN uses a three-axis RoPE layout - (time, height, width) - and the node scales the per-frequency temperature on the spatial axes only, leaving the temporal axis untouched. It doesn't touch your sampler, scheduler, block loop, or VRAM handling; it just multiplies the RoPE frequencies once per forward path.

Inputs that matter

• model - any ComfyUI WAN 2.1/2.2-family MODEL.
• width / height - your final frame dimensions in pixels (defaults 1280x720). Must match the video size your workflow actually produces.
• temperature_strength - 0.0 disables the patch entirely; 1.0 uses the TIDE/YaRN curve.

The interesting optional is base_resolution_mode, which defaults to aspect_adaptive_720p. That's the right instinct for WAN: instead of forcing one fixed reference rectangle, the adaptive modes preserve your target aspect ratio while keeping a 1280x720 (720p) or 832x480 (480p) native pixel budget, so a portrait or unusual image-to-video aspect still gets a sane reference point. manual uses base_width/base_height literally. Then the curve knobs (alpha_low, alpha_high, tau_max, frequency_mode) behave exactly as they do on the Flux node, plus apply_to_native_or_smaller to force patching at native-or-smaller sizes.

Output is a single patched model - model loader → this node → sampler.

Installing it

Same pack, same drill. ComfyUI Manager search for "TIDE", or:

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/ComfyUI-TIDE

Restart ComfyUI. requirements.txt is just torch - no extra dependencies, no video model downloads, nothing.

Gotchas

One real trap: the README and the shipped node have drifted on the base resolution. The README's usage section talks about a 640x640 default, but the actual code defaults to base_resolution_mode=aspect_adaptive_720p with 1280x720 base dims. Trust the node's own defaults, and if you see older write-ups citing 640x640, treat them as stale. Also remember this is a temperature-only adaptation - anyone promising "text anchoring for WAN" from this pack is misreading it. And the author's own README is upfront that WAN quality needs live testing across 2.1/2.2 variants, frame counts, and samplers. This is a brand-new, low-traffic node; you're an early adopter. A/B it against your normal WAN workflow at your target resolution before you build it into your pipeline.