凤希AI - LoRa加载器

FxAiLoraLoader (凤希AI - LoRa加载器) is the fxai-toolkit take on the boring-but-essential LoRA stacker, and it lands somewhere between ComfyUI's stock LoraLoader and a full LoRA-block editor. You feed it a base model and a JSON list of LoRA entries, and it applies them in order - order preserved, not alphabetically scrubbed like some stacks do - then stitches all their trigger words onto your prompt for you. If you're running a character-heavy long-video workflow where every shot needs the same three character LoRAs plus a style LoRA, this is the node that keeps the stack consistent between segments instead of rebuilding it each time.

The lora_data input is a JSON array (a hidden widget; the pack's JS gives you a table UI to edit it). Each entry can carry:

• lora_name - file in your loras folder.
• enabled - set false to skip without deleting.
• model_strength / clip_strength - clip_strength of -1 means "use model strength," the stock-UI convention.
• invert - flips both strengths negative (anti-LoRA, handy for style subtraction).
• fade_start / fade_end - multiply model strength by fade_start and clip strength by fade_end. Yes, the fade is per-LoRA and static, not per-sampler-step, but it's a quick way to taper a character LoRA's influence shot to shot.
• trigger_words - a list of tags that get collected and appended to the output prompt in the same order the LoRAs applied. That last bit is the sleeper feature: the node returns MODEL, CLIP, and the assembled 提示词, so the trigger tags follow the LoRAs through your workflow without you hand-maintaining a text file.

Outputs: MODEL, CLIP, 提示词. The prompt output is your base prompt (from the optional 提示词 input) plus all trigger lines - feed that straight into a positive encode and you've closed the loop.

Mechanically it's a thin wrapper around ComfyUI's own LoraLoader().load_lora, so the compatibility story is just "whatever ComfyUI core supports." No model downloads of its own; it reads from the standard loras folder. Missing files raise an error that helpfully points at the author's cloud links, which tells you this pack expects you to grab LoRAs from the community rather than ship them.

Install is pack-standard: ComfyUI Manager, search "fxai-toolkit", or cd ComfyUI/custom_nodes && git clone https://github.com/fxai666/fxai-toolkit, restart. Auto-installs soundfile and psutil; needs FFmpeg on PATH; all labels in Chinese (凤希AI builds for a Chinese-speaking audience - help is in QQ group 775649071 and on Bilibili).

Where people get burned: malformed JSON in lora_data. The node silently falls back to an empty list on a parse error, so a missing comma means "no LoRAs applied and no error" - the classic silent-bad-result. If your output looks like the LoRAs did nothing, check the table, not the sampler. And remember the fade knobs multiply, not add: a fade_start of 1 is "no fade," 0.5 is half strength. Start there.