Load LingBot Prompt Rewriter (Yogurt LingBot Video)

Nobody warns you about this with LingBot-Video: it doesn't prompt like a normal model. When the weights dropped, the community put it bluntly - "no natural language prompting; all json." LingBot wants a structured caption, and hand-writing those is a chore. This node loads the translator: a Qwen3.6-27B base model with LingBot's rewriter LoRA (Robbyant--lingbot-video-rewriter-lora) bolted on, turning your rough idea into the structured JSON the generator actually understands.

So this loader is a different beast from the video model loader. It's not a video checkpoint - it's a large vision-language model, 27B params, loaded in bfloat16 with device_map=auto. That makes it a second resident on your GPU, and it is going to fight the MoE video model for VRAM. The pack handles this by treating the rewriter as a temporary guest: the rewrite node unloads it after every job by default. The README even ships a standalone rewrite workflow so you can polish a prompt, copy the JSON out, and only then load the video model. Running both at once on a 24 GB card is where people get burned.

The inputs are minimal and mostly boring:

• base_model and adapter_model - dropdowns with a single official choice each (the Qwen3.6-27B base, the rewriter LoRA), both resolved under ComfyUI/models/LingBotVideo/. Note these are two separate directories - the base and the LoRA are not one folder.
• download_model - defaults to false, same explicit-only policy as the video loader. Flip it only when you're ready to pull a 27B.

Output is one rewriter handle of type YOGURT_LINGBOT_VIDEO_REWRITER, wired straight into the LingBot Prompt Rewrite node. There's nothing to tune here; the real knob lives downstream, in that node's release_rewriter_after_rewrite toggle.

Install

Clone into custom_nodes and install requirements into the Python environment ComfyUI actually runs:

cd ComfyUI/custom_nodes
git clone https://github.com/yogurt7771/ComfyUI-YogurtNodes-LingBotVideo
pip install -r custom_nodes/ComfyUI-YogurtNodes-LingBotVideo/requirements.txt

Deps include transformers>=4.57,<5, peft>=0.19.1, accelerate, and json_repair. Restart ComfyUI after. Model layout:

ComfyUI/models/LingBotVideo/Qwen--Qwen3.6-27B/
ComfyUI/models/LingBotVideo/Robbyant--lingbot-video-rewriter-lora/

Gotchas

Disk space first - the base model is big, so make sure you actually have room before enabling download_model. First load is slow, and every reload after a VRAM release is slow too, which is the tradeoff behind that release toggle. And again: don't try to park this next to the MoE video model on a single 24 GB card. The rewriter is a tool you use in passing, not a permanent resident.