BerniniR · Apply Patches (para GGUF/UNETLoader)

BerniniR · Apply Patches (para GGUF/UNETLoader) does one thing: takes any native Wan MODEL - specifically one loaded with UnetLoaderGGUF from ComfyUI-GGUF - and retrofits Bernini's two patches onto it: the source-id RoPE and the stream-concat forward. That's the difference between a quantized Wan model and a model that can actually run Bernini's editing streams. One node, one input, and it's the thing that makes the whole GGUF path work.

Why do you need it? The GGUF route loads the model through ComfyUI's native machinery, which knows nothing about Bernini. Without the patches, the model's attention runs on the stock cos/sin RoPE, and the stream-concat logic that Bernini's inference code depends on isn't there. With them, the model understands the bernini_streams that BerniniR · Source Stream injects - each stream's RoPE grid multiplied by its source-id phase - and editing becomes possible on GGUF weights.

Input

model (MODEL) - a Wan model, e.g. from UnetLoaderGGUF. That's the entire interface. It clones the model, applies the patches to the clone (so your original stays untouched), and hands it back.

When you need it vs. when you don't

• Need it: GGUF editing graphs - i2i / v2v / rv2v on the dual-expert 14B. The README's wiring is UnetLoaderGGUF → Apply Patches → Source Stream → Guider, with the high-noise model as model and the low-noise GGUF as model_low on the Guider.
• Don't need it: the BerniniR · Load Model (native) node already applies these patches internally - running it through Apply Patches again is harmless but redundant.
• Genuinely optional for pure t2v/t2i: with no streams, the patched forward is numerically identical to standard Wan2.2 (source-id 0 = identity phase), so a plain UnetLoaderGGUF → sampler works without it. It's the editing streams that need the patches.

Output

model (MODEL) - the patched clone, into BerniniR · Source Stream (or directly into the Guider if you have no streams to add).

Install & troubleshooting

Shared pack install: ComfyUI Manager, or clone into custom_nodes + pip install -r requirements.txt. You also need city96/ComfyUI-GGUF installed for the UnetLoaderGGUF loader that feeds this node.

• "patches Bernini NOT installed" warning? The model still runs as native Wan - t2v works, editing won't. The patch install is fault-tolerant by design; if you see the warning, double-check the model is a Wan model (not something else loaded by accident).
• No-op on t2v? Correct - with no bernini_streams in transformer_options, the forward falls through to stock Wan. That's not a bug, it's the design.
• Forgetting it on the GGUF path is the #1 way to get a silently-ignored edit: model loads fine, generation works, but your source video does nothing.