(WIP) WanVideo LoRA Optimizer

Stacking LoRAs on Wan video has the same problem it has on image models, plus a compounding one: one user reported likeness getting worse than plain Wan 2.2 once LoRAs were stacked, and that's before you factor in that Wan LoRAs come from a wider mix of trainers than most image ecosystems - Kohya, LyCORIS, diffusers, Musubi Tuner, Fun LoRA, finetrainer variants all circulate. This node is the LoRA Optimizer's engine - TIES-Merging, DARE/DELLA sparsification, per-layer conflict resolution - wired specifically for kijai's ComfyUI-WanVideoWrapper, the third-party wrapper that became the primary way people run Wan with extras like SageAttention and TeaCache. The regular LoRA Optimizer takes a MODEL; this one takes a WANVIDEOMODEL, because that's what the wrapper's loader produces instead.

One honest flag before you build around it: the node's display name literally reads "(WIP)". It's functional and shares the same battle-tested merge engine as the rest of the pack, but expect it to be the least stable corner of this node suite - rougher edges, more likely to change than the image-side nodes.

How it works

Same engine as the standard optimizer, pointed at a different model object with no CLIP branch - Wan wrapper models don't use text-encoder patching, so this node skips it and applies merged patches in-memory. Everything else carries over: the two-pass analyze-then-merge pipeline, per-layer strategy selection (weighted_sum where only one LoRA touches a group, TIES where sign conflicts are real), DARE/DELLA sparsification, and optional merge refinement. Key normalization matters more here than on image models - Wan's T2V/I2V/VACE variants share roughly 90% of their weights but each has unique keys, named differently by different trainers. That's why this node flips a few defaults relative to the plain Optimizer: normalize_keys on (mixed-trainer Wan stacks are the norm), cache_patches off (video models are large enough that caching eats real RAM), and architecture_preset defaulting to dit instead of auto - the right thresholds for Wan are already known, so it skips the guess.

The inputs and outputs that matter

• model (required, WANVIDEOMODEL) - from WanVideoModelLoader, not a regular checkpoint loader.
• lora_stack (required) - same LoRA Stack node you'd use anywhere else in this pack.
• output_strength (default 1.0, -1 for auto) - master volume on the merge, identical behavior to the image-side node.
• auto_strength (default enabled) - turns down individual LoRA strengths automatically when stacking, the same anti-oversaturation lever as the standard optimizer. Worth double-checking it's on here specifically, since Wan stacks are exactly the "3+ LoRAs, high risk of stepping on each other" case this setting exists for.
• Everything past that - sparsification, merge_refinement, strategy_set, decision_smoothing - mirrors the standard LoRA Optimizer's advanced knobs, just defaulted for video.

Outputs: model (a WANVIDEOMODEL, feed it straight to WanVideoSampler), analysis_report (STRING - connect to Show Text to see which layers got merged how), and lora_data (feeds Save Merged LoRA if you want the result as a portable file). There's no clip output here, unlike the standard optimizer - Wan wrapper models don't carry one.

Workflow shape:

WanVideoModelLoader → WANVIDEOMODEL → WanVideo LoRA Optimizer → WANVIDEOMODEL → WanVideoSampler
                                              ▲
                       LoRA Stack ────────────┘

You can also chain it with individually-applied (non-merged) LoRAs coming from WanVideoLoraSelect upstream - the merged optimizer output and unmerged loader patches coexist fine in the same model patcher.

Installing it

Two things need to be present, not one. First, this pack, same as always:

cd ComfyUI/custom_nodes/
git clone https://github.com/ethanfel/ComfyUI-LoRA-Optimizer.git

(or search "LoRA Optimizer" in ComfyUI Manager). Second - and this is the part specific to this node - you need kijai's ComfyUI-WanVideoWrapper installed separately, since that's what actually produces the WANVIDEOMODEL type this node consumes. This pack doesn't bundle it. Restart ComfyUI after both are in place; nodes land under loaders.

Common issues & troubleshooting

Distillation LoRAs (CausVid, Lightx2v) don't belong in the stack. Same warning as the image-side optimizer, and it matters more here - Wan's own community has documented these speed LoRAs degrading motion and lighting and giving "flux level plastic skin" even applied alone, so merging them into a conflict-resolved stack is asking for trouble. Apply CausVid/Lightx2v through a standard loader upstream and keep only style/character LoRAs in this node's stack.

Merge feels slow or heavy on RAM. cache_patches defaults off here specifically because video models are big; that's correct behavior, not a bug - turning it on for repeated re-runs on a large stack will cost you real system RAM.

It doesn't work with your Inline Chain workflow. The pack is explicit that WanVideo wrapper models are not supported by LoRA Optimizer (Inline Chain) - this dedicated node exists precisely because that one doesn't handle WANVIDEOMODEL. If your workflow already runs LoRAs through WanVideoLoraSelect, you'll need to route through a LoRA Stack and this node instead to get the conflict-aware merge.

Results feel unstable or the node's behavior changes between pack updates. That's the honest cost of the "(WIP)" label - treat this node as more likely to shift than the rest of the suite, and keep an eye on the pack's changelog if you're relying on it for production work.