⭐ Star Minimax H3 LoRA Merge

Here's the middle of the ⭐StarNodes H3 LoRA pipeline, and the part that earns its keep. It takes two MINIMAX_H3_LORA objects and blends them into one - a standard LoRA that the stock LoraLoader loads like any other file, in any H3 workflow. The pack even tells you where the code came from: it was merged in from a standalone Star_M3H_Merger pack, so this is the field-tested version, not a weekend prototype.

Why you'd reach for it. H3's weights only dropped in August 2026, so its LoRA library is young and specialized. The classic moves apply: combine a style LoRA with a character LoRA, interpolate between two takes on the same subject, or just fold two half-usable adapters into one file you can share. The old SD community built entire careers on weight-space merging; this brings that habit to a video model.

How the merge works. This is the part people get wrong elsewhere, and this node gets right. It doesn't average the raw up/down matrices - that breaks because each LoRA has its own rank and scaling. Instead it:

1. Reconstructs each LoRA's delta matrix (dW = up @ down, alpha/r scaling included) in fp32.
2. Blends the deltas in weight space: dW_merged = (1-weight)·A + weight·B.
3. Re-composes the blended delta back into up/down factors via SVD, at your chosen rank.

Because the result is re-factorized, it stays a normal LoRA - loadable everywhere, not some Frankenstein tensor soup that only this pack can read.

The inputs that matter:

• lora_a / lora_b - the two MINIMAX_H3_LORA objects from two Loader nodes.
• weight - the blend slider. 0 = only A, 0.5 = 50/50, 1 = only B. Start at 0.5 and nudge.
• output_rank - rank of the merged LoRA (dropdown, 8–128, default 32). If the combined rank of A+B is smaller, it uses that instead, so you never "lose" capacity by picking too high.
• output_dtype - bf16 (default), fp16, or fp32. Storage only; the math is always fp32, so this just decides file size.

Two details worth knowing. LoRA keys are matched by exact base-layer name, so the merge only ever blends on layers both LoRAs share; layers present in just one are carried over, scaled by that side's weight. And "extra" tensors (things that aren't up/down pairs) are blended when shape-compatible, kept from one side otherwise. That's what the report output summarizes - read it once to confirm both LoRAs actually shared layers. If they don't, you're not getting a blend, you're getting a union.

The output minimax_h3_lora feeds the ⭐ Star Minimax H3 LoRA Saver (or anywhere else that accepts the type). The report STRING is your post-merge sanity check.

Where it lives. Install StarNodes once - Manager, search Starnodes, restart - and all three nodes appear under ⭐StarNodes/Video. No base model is loaded during the merge; it's pure CPU tensor math, so it's fast and doesn't touch your VRAM. The usual gotcha applies: restart ComfyUI after installing, or the nodes won't show up.

One final honest note. Merging LoRAs is legal anywhere - but these adapters patch H3 weights, which are licensed for use only outside the US, EU, UK and South Korea. If you're in an excluded territory, the merge is the least of your problems.