Krea LoRA Converter Pixaroma

Krea LoRA Converter Pixaroma fixes a specific, maddening problem: you trained a Krea 2 LoRA on fal.ai, tried to load it in ComfyUI, and it just... doesn't load. Not a wrong-model error, not a crash - the file is valid, the weights are fine, and no LoRA loader will touch it. The culprit is layer naming, and this node converts the file into one ComfyUI actually recognizes. It's a narrow tool, but if it's your problem, it's the only tool.

Here's the mechanism, and it's worth understanding so you trust the result. fal.ai's Krea 2 trainer names the layers the PEFT/diffusers way (base_model.model.…), but ComfyUI's Krea 2 support expects the diffusers names its own krea2_to_diffusers utility produces. The names don't match, so the loader silently ignores the LoRA - it "loads" and does nothing. This node renames the keys in the file's header and copies the weight blocks through byte-for-byte. The weights are never touched, so the converted file is identical to the original - just loadable. It is, in the truest sense, a lossless conversion.

The safety design is the other half of why you can trust it: the node only reads your file and writes a new one. It never modifies the original, never downloads anything, and it runs entirely locally. The project is explicit that it's an independent tool, not affiliated with or endorsed by Krea or fal.ai - this is a fan's fix for a real interoperability gap, not an official feature.

The inputs

Three, and you'll touch maybe two:

• lora_name - pick the LoRA from your models/loras folder. The readout on the node confirms whether it's actually a fal Krea 2 LoRA before you convert, so you don't burn a click on a file that isn't. If it detects a Krea 2 LoRA, it shows the layer counts and the lora_alpha, and suggests an output name.
• output_name - leave blank to use the original name plus _comfyui (the sane default). Otherwise type your own.
• overwrite - off by default, so you never clobber an existing file by accident. Turn it on only when you know what you're overwriting.

There are no outputs - this is an action node. You pick, you read the confirmation, you click Convert, and a new .safetensors appears next to the original in models/loras. Refresh the LoRA list and load it like any other LoRA.

After you convert

Two things worth knowing once it loads. First, a converted file gets a pixaroma_converted_from metadata marker, so you can tell at a glance which files have been through the converter. Second - and this is the part that trips people even after a successful conversion - Krea 2 LoRAs generally need their trigger words in the prompt to activate. The community threads about Krea 2 LoRAs in ComfyUI are full of "no impact no matter the strength" posts, and the usual resolution is the trigger word, not a loading bug. If your converted LoRA loads but does nothing, check your prompt before blaming the converter.

Installing it

It's part of the ComfyUI-Pixaroma pack - ComfyUI Manager search Pixaroma, Install, restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/pixaroma/ComfyUI-Pixaroma

No extra Python dependencies - conversion happens with the pack's own safe-tensors handling. House rule as always: after updating the pack, hard-refresh your browser tab (Ctrl+Shift+R / Cmd+Shift+R) so a cached old UI doesn't hide the new readout.

Gotchas

The honest limits: it converts Krea 2 LoRAs trained on fal.ai - a different Krea LoRA or a LoRA trained elsewhere won't match the fal naming pattern, and the readout will tell you so instead of mangling the file. And if the file already uses ComfyUI-style names, there's nothing to convert. It's a targeted fix for a targeted pain, which is exactly what a utility node should be.