Nodes/ComfyMepi/Mepi Checkpoint and Loras Point
ComfyUI Node

Mepi Checkpoint and Loras Point

The one node in the pack with actual logic

By rhplus0831·Created 2 years ago·Updated about a year ago· 0
Mepi Checkpoint and Loras Point
    • MODEL
    • CLIP
    • VAE
    ckpt_name
    loras

    The one node in the pack that isn't pure pass-through

    Mepi Checkpoint and Loras Point is the heavy lifter of ComfyMepi, a small mobile frontend for ComfyUI by rhplus0831. The pack's other nodes are dumb string pipes; this one actually loads a model. The author's own comment in the source calls it a "Dirty copy of CheckpointLoaderSimple," and that's the right mental model - it behaves like ComfyUI's built-in loader, with a LoRA list bolted on the side.

    The checkpoint half is textbook: ckpt_name is a dropdown populated from your ComfyUI/models/checkpoints folder, and loading uses the same load_checkpoint_guess_config routine as core ComfyUI, so it auto-detects whether the file is SD 1.5, SDXL, Flux, or whatever else you throw at it. Out the other end you get the standard trio:

    • MODEL - the denoising model, into the KSampler.
    • CLIP - for encoding your prompts, into CLIP Text Encode.
    • VAE - for decoding latents to pixels, into VAE Decode.

    The loras field is the interesting (and dangerous) part

    The second input, loras, is a multiline string. It's not a dropdown and there's no file picker - you type a comma-separated list in groups of three: filename, model_strength, clip_strength.

    my-lora.safetensors,1.0,1.0
    

    Two LoRAs, same format, just keep chaining:

    style-lora.safetensors,0.6,0.6,char-lora.safetensors,1.0,1.0
    

    The code splits on commas and groups them into triplets, then applies each with the standard load_lora_for_models call - same strength semantics as core LoraLoader, where the first number is how hard the LoRA pushes the model and the second is the CLIP/text influence. It also caches the loaded LoRA tensors per node instance, so re-running with the same filename doesn't re-read the file from disk.

    Where people get burned: the field must always contain complete triplets. A trailing comma, a lone filename, or an empty field makes the parser index past the end of the list and the whole queue bails with an IndexError. The mobile UI writes well-formed triplets for you; hand-editing this box is where it falls apart. When in doubt, paste the exact format above.

    Install

    No Python dependencies, no model downloads - it loads whatever checkpoints and LoRAs you already have in your ComfyUI models/ folders. Install the pack via ComfyUI Manager (Install Custom Nodes → search "ComfyMepi") or:

    cd ComfyUI/custom_nodes
    git clone https://github.com/rhplus0831/ComfyMepi
    

    Then restart ComfyUI. If your checkpoint isn't in the ckpt_name dropdown, it belongs in ComfyUI/models/checkpoints/ - refresh the list and it'll appear.

    Troubleshooting

    The classic failure here isn't a crash, it's a LoRA that silently does nothing. The loader has no idea your LoRA is incompatible - it just applies the patch and hopes. Two things to check:

    • Base-model compatibility. A Flux LoRA on an SDXL checkpoint, or an Illustrious LoRA on SD 1.5, loads fine and visibly changes nothing. Confirm the LoRA was trained for the checkpoint you loaded. If in doubt, run the checkpoint with no LoRAs and compare.
    • Strength. 1.0 is often too strong. The community's default working range is 0.5–0.8, so if output looks overcooked, dial the first number of the triplet down before blaming the model.

    And the usual reminder: this pack is a personal project with essentially no community presence and light maintenance, so don't expect a bustling issue tracker. But since the loading path is a copy of core ComfyUI, the failure modes are the same ones you'd hit on the built-in loader - and those are well documented.

    Categorymepi

    Inputs (2)

    NameTypeDefaultDescription
    ckpt_nameCOMBOThe name of the checkpoint (model) to load.
    lorasSTRING

    Outputs (3)

    NameTypeDescription
    MODELMODELThe model used for denoising latents.
    CLIPCLIPThe CLIP model used for encoding text prompts.
    VAEVAEThe VAE model used for encoding and decoding images to and from latent space.