Loader (Mango)

Loader (Mango) is the pack's answer to the classic ComfyUI starter graph: instead of a Checkpoint Loader, a LoraLoader, and a bunch of wires, you get one node that loads a checkpoint, applies up to five LoRAs, and - here's the part that actually matters - computes a short SHA-256 hash of the checkpoint and remembers every LoRA it applied. That hash and that LoRA list are exactly what the pack's metadata chain needs to produce Civitai-compatible saves downstream.

If you never save with the Mango metadata chain, this node is just a fancier loader. If you do - and the whole pack is built around that - it's the entry point that makes honest metadata possible without any manual bookkeeping.

How it works

Under the hood it calls the same machinery the stock Checkpoint Loader uses: comfy.sd.load_checkpoint_guess_config with the VAE and CLIP pulled out, then comfy.sd.load_lora_for_models for each LoRA you select. What it adds on top is bookkeeping. It streams the checkpoint file and computes a 10-character SHA-256 prefix (the same kind of short hash Civitai uses to fingerprint model versions), and it assembles a lora_stack - a list of (name, weight, weight) tuples - recording which LoRAs were applied at what strength. That stack is the same data structure the sampler needs to log Lora hashes in your saved images.

Inputs and outputs that matter

There are eleven inputs and you'll mostly ignore half of them:

• ckpt_name - dropdown of your checkpoints. The one you'll set every time.
• LoraName1…LoraName5 / LoraWeight1…LoraWeight5 - five LoRA slots with per-LoRA weights. Leave a slot at None to skip it. Weights go −10 to 10; you'll usually live between 0.5 and 1.5.

The outputs are where the magic lives:

• model / clip / vae - wire these to your sampler and VAE decode like normal.
• lora_stack - wire into the sampler so it can record LoRA hashes in metadata.
• ckpt_name / ckpt_hash - STRING outputs carrying the checkpoint's filename and short hash. Feed them into the sampler's text fields for full metadata fidelity.

The order matters for LoRAs: they're applied in slot order, so slot 1 goes in first. If two LoRAs fight over the same style, that ordering changes the result - worth remembering when you're debugging "why is this different than my old graph."

Installing it

Loader (Mango) ships in the Mango Node Pack. You get all the pack's nodes in one go:

cd ComfyUI/custom_nodes
git clone https://github.com/mang01010/MangoNodePack

Restart ComfyUI, or install "Mango Node Pack" from ComfyUI Manager. No special model downloads - it loads whatever checkpoints and LoRAs you already have in models/checkpoints and models/loras. The README's only listed Python dependency is safetensors, which stock ComfyUI ships with.

Gotchas

The big one from the pack README: occasionally a LoRA's hash won't match anything on Civitai, so it shows up as "unknown" in the metadata. The author notes this is usually because the model was updated on Civitai after you downloaded it - re-download the file and the hash lines up again. Also, don't expect this loader to be meaningfully faster than stock - the hash computation streams the whole file once, which for a 7 GB checkpoint is a real but short delay. And if you need more than five LoRA slots, that's what Loader (Mango + 10 Loras) is for.