MultipartCLIPTextEncode

You know the staircase: a CheckpointLoader, three LoRA loaders, two CLIPTextEncode nodes, and a wire from each into the sampler. MultipartCLIPTextEncode is that whole middle section as one node. Give it a CLIP, a model (optional), a list of LoRA tags, and both prompts, and it hands back the modified MODEL, the modified CLIP, and a CONDITIONING for each of positive and negative.

It's the workhorse of this pack. In the author's sample workflows it's the only node between the loader and the sampler.

How it works

Two passes, in order.

LoRAs first. The lora_tag_list is split on newlines and each line parsed for <lora:name:strength> (or <lora:name:model_strength:clip_strength>). Matching tags are applied through core LoraLoader, which patched both model and clip - this is the full loader, unlike the pack's LoadLoraFromLoraList, which is model-only. Duplicate filenames are skipped. A line that is exactly -- stops loading; the source comment says the intent is "don't load both HIGH and LOW" - that's the signal for Wan 2.2 high/low expert setups, where the low-noise tags should go to a second sampling stage instead.

Then encoding. With enable_break on, the text is split on BREAK surrounded by spaces or commas, each piece is encoded separately, and the pieces are stitched back together with a conditioning concat. With it off, each prompt goes through core CLIPTextEncode as one chunk.

The inputs that matter

clip is required. positive and negative are multiline strings, usually wired from a prompt file. lora_tag_list is multiline - one tag per line - and is what you wire from PromptDecode's LoRA output or MultipleLoraTagLoader. enable_break is the BREAK splitter described above. model is optional, and you should almost always wire it: the node returns the LoRA-patched model through its MODEL output, and if nothing is connected there it has nothing to patch.

Outputs: MODEL, CLIP, CONDITIONING (positive), CONDITIONING (negative).

Install

Manager search for the pack name, or:

cd ComfyUI/custom_nodes
git clone https://github.com/morino-kumasan/comfyui-toml-prompt

restart ComfyUI. There's nothing to pip install - requirements.txt is empty - and the encoder, concat and loader classes come from ComfyUI itself. The README's install block still names the pack's old comfyui-utils repo and clones over SSH; use the HTTPS URL above.

Where people get burned

Leave model unwired and you get the "Lora Loaded" lie. The code only calls LoraLoader when a model is connected, but it prints Lora Loaded: <name>: <strength> regardless. So the console cheerfully reports LoRAs that were never applied, and you get base-model output with none of your style. If your LoRAs are doing nothing, check the model wire before anything else.

BREAK means something different here than in core, and it's CLIP-era. Core CLIPTextEncode splits at the 77-token boundary; this node encodes each piece separately and concatenates the conditionings. Either way, on the 2026 Qwen3/Mistral-encoded models there's no chunk boundary to break against and BREAK is dead weight - turn enable_break off there and let the text flow.

A LoRA tag that doesn't match the regex is skipped silently. The pattern wants a plain number, so trailing whitespace or a capital L is a no-op with no warning.

The -- stop is positional, not a filter. Any tag below a -- line is ignored by this node entirely. If you pasted a full high/low list here and only half your LoRAs applied, that's why - split it first.