🔥 Easy LoRA Baker

Every time you load that character LoRA you also have to remember to load the base model, set the weight, and keep track of which LoRAs stack well. There's a simpler life available: bake the LoRA into the checkpoint itself and end up with one file that just works. That's the Easy LoRA Baker. It adds the LoRA's weights into a base checkpoint at the tensor level and hands you a complete MODEL + CLIP + VAE that has the concept permanently.

It's the natural end of a merge pipeline - merge two or three LoRAs in the Easy LoRA Merger, then bake the result into a checkpoint with this node. Handy for distributing a model that includes a style you can't otherwise share, or for locking in a look you never want to forget to load.

How it works

You pick a checkpoint and a LoRA, and the node adds the LoRA deltas into the checkpoint's weights - this is different from stacking LoRAs at load time, which patches the model on the fly. Once baked, there's no LoRA to load; the weights are just in the file. Because LoRAs are tiny deltas, you get to choose how hard you bake: strength (0–2) multiplies the whole adapter, and weight_unet / weight_te / weight_clip / weight_vae scale it per component so you can bake the style but leave the text encoder alone.

The baking_method choices - linear, impact_weighted, orthogonal - are different ways of folding the delta in, and for a first run linear is right. The work streams through in batch_size-sized chunks with a memory guard between batches so it doesn't blow up VRAM or RAM on big checkpoints.

There's also a RAM Guard worth understanding. With save_trigger off (preview mode), the baked result stays in memory with an automatic fallback to disk if you're near the 85% RAM threshold. With save_trigger on, it saves the .safetensors first and then lazy-loads from the file - the low-RAM path, and the one you want for a 12GB checkpoint.

Inputs that matter

• checkpoint - the base model. Its siblings (model, clip, vae) come out the other side.
• lora_data - wire the lora output of the Easy LoRA Merger here; it overrides the lora_name dropdown.
• strength - start at 1.0. Baking is permanent, so test at preview before you commit.
• precision - auto, or force fp8 to shrink the baked file.
• save_trigger - the commitment switch. Off = in-RAM preview, on = permanent .safetensors.

Outputs: model, clip, vae (test them downstream immediately), output_path, and forensic_report.

Installing it

Manager (search "Easy LoRA Merger") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Terpentinas/EasyLoRAMerger

Restart ComfyUI. No model downloads.

Gotchas

The big one is obvious but worth stating: baking is one-way in practice. You can't unbake, so keep the original LoRA and checkpoint around. Also don't bake at strength 2.0 and be surprised when the output is overcooked - preview with save_trigger off, exactly like the rest of this pack. One practical note from the ecosystem: baking at high strength tends to look stronger than the same LoRA at high weight at load time, because there's no negative-weight trickery possible afterward. And if RAM is tight on a big model, use save_trigger on rather than fighting the in-memory path.