FL PixelGen LoRA Loader

So you trained a LoRA with the pack's Train node, or picked one up from somewhere, and now you want it to actually do something. That's this node. It sits between the model loader and Generate, and it does the LoRA thing you're used to from every other ComfyUI pack - with one catch that burns people immediately: the lora_path input is a text field, not a dropdown. You paste a path. There's no browser, no models/loras folder scan, nothing to click.

The mechanism is refreshingly direct. The node takes your PIXELGEN_MODEL, deep-copies the cached denoiser (so the base model in memory stays untouched), loads the adapter's adapter_model.safetensors, and applies each low-rank pair by hand: W' = W + strength * (B @ A). No PEFT library involved at inference - it just adds the low-rank update onto the target weights and hands back a fresh model. That's why it can chain: each LoRA Loader outputs the same PIXELGEN_MODEL type it consumed, so stack two of them in series and you get two adapters applied.

The inputs:

• model - from the model loader (or from another LoRA loader, to stack).
• lora_path - the string to paste. It's forgiving about what you give it: a direct adapter_model.safetensors file, an adapter/ folder, a checkpoint folder containing adapter/, or a whole training output directory (it'll grab the final/ adapter, or the highest-numbered checkpoint-epoch-N). Useful, because the Train node hands you exactly the top-level output directory.
• strength - 0.0–2.0, default 1.0, stepped by 0.05. Same mental model as everywhere else, with a modern twist: don't be shy about running 1.0 on this architecture family; the old "drop it to 0.7" instinct is SDXL-era advice.

Output: model, a PIXELGEN_MODEL with the LoRA applied - wire it into Generate or another LoRA Loader.

Install: shared pack install - ComfyUI Manager search "FL PixelGen", or

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-PixelGen
cd ComfyUI-FL-PixelGen
pip install -r requirements.txt

then restart. No extra downloads for this node itself; the models it touches come from the loaders.

Troubleshooting - the real ones: if you see Could not find adapter_model.safetensors in or under: <path>, the path is wrong - check you're pointing at the folder containing the adapter, not one level up, and remember absolute paths are your friend. If the console logs "Skipped N LoRA modules," the adapter was trained for different target modules than the denoiser has - usually means the LoRA is for a different base, and it'll do nothing useful. And if you expected to browse ComfyUI/models/loras/, drop that expectation: this node is a raw-path loader by design, so keep your PixelGen LoRAs somewhere you can type their path from memory.