WINT8 Power LoRA Loader

Here's the classic quantized-model LoRA pain, in one sentence: applying a LoRA to a quantized model normally means dequantizing every layer, patching in the LoRA weights, and requantizing it back - a slow, VRAM-hungry round-trip that on INT8 models frequently breaks outright. ComfyUI's native INT8 (ConvRot) support was still smoothing over "LoRA specific issues" at launch, with even the maintainers pointing at custom-node workarounds. WINT8 Power LoRA Loader is that workaround. It's the node you reach for when you've got an INT8 model loaded and a stack of regular LoRAs you want to throw at it without everything falling over or slowing to a crawl.

How it works

Instead of patching weights the standard ComfyUI way, it registers a dynamic forward hook on the diffusion model and applies LoRA patches at forward time, from a dynamic_loras list it threads through the model options. No dequantize-patch-requantize cycle, no per-layer conversion. Two things fall out of that design that are easy to miss and genuinely nice: it uses your existing standard LoRA files - no INT8-converted LoRAs needed - and each LoRA is composed on the fly, so the model clone stays clean between runs.

The frontend is the same multi-row Power LoRA UI as the author's winnougan-nodes pack: a live search dialog over your loras folder, individual on/off toggles and strength per row, plus a global toggle-all. Each row sends the standard payload - on/off, the LoRA filename, strength - and the node chains them through the hook.

Inputs and outputs

The schema is deliberately loose, because the lora rows are added dynamically by the UI rather than fixed inputs. The two you'll wire up by hand:

• model - your INT8 (or any) model. This one is effectively required: leave it unplugged and the node raises "No model connected."
• clip - passed through completely unchanged. Hook it up if your graph needs a CLIP alongside; it isn't touched.

Outputs are MODEL and CLIP, both out the right side in order. The MODEL output feeds your sampler exactly as you'd expect.

Install

Same pack as the other two WINT8 nodes: ComfyUI Manager (search "Winnougan"), or

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/WINT8-ComfyUI

then restart. It rides on the WINT8 suite's internals, so no extra pip dependencies - and because it's self-contained, you don't need any other LoRA pack installed for it to work.

Where people get burned

• Forgetting that model is required even though the schema shows it as optional. The node errors loudly, which at least beats silent failure.
• LoRA files that won't resolve: the loader matches by filename against your loras folder. If a row silently does nothing, that's almost always a name mismatch, not a quantization problem.
• Expecting a strength multiplier that isn't there - each row has its own strength control and an on/off toggle; if a LoRA has a second strength field (the UI supports it), the node reads it. Rows with strength 0 or toggled off are skipped entirely.
• And the honest caveat in the other direction: if you're on a card where fp8 fits and you aren't quantized, ComfyUI's built-in LoRA loaders are faster and simpler. This node earns its keep specifically on INT8 models - which is exactly the scenario native ComfyUI is still sorting out.