VidScribe MiniCPM Beta

You've got a video and you need words for it: a caption for a training set, a scene description to feed a prompt writer, a quick "what's actually happening here" before you decide the next step. VidScribe MiniCPM Beta is a vision-language model that runs on your own GPU and answers that. No API key, no cloud, no per-call cost - the model downloads itself on first use and stays local. The "Beta" in the name is honest: it's the pack's experimental VLM, and it's also the centerpiece of a whole VRAM-management story (see below).

Where it earns its keep is the workflow it enables: describe a clip, then hand that description to a prompt generator and a diffusion model in the same graph. That's the exact "VLM-as-a-node" pattern that's become routine in ComfyUI - the captioner feeding the enhancer feeding the sampler.

How it works

Two backends, selectable with the model dropdown:

• minicpm_v4.5_int4 (default) - MiniCPM-V 4.5 in int4 quantization, roughly 6–8GB VRAM.
• mage_vl_4b - Microsoft's Mage-VL, ~10GB in bf16. It's billed as stronger on video, but it ignores thinking_mode and needs a newer transformers.

Both auto-download on first use. The node smart-samples ~32 frames from longer clips (turn that off with use_all_frames at your own VRAM peril), runs inference, and then - this is the clever part - unloads the model and emits a vram_cleared signal. After 60 seconds idle it unloads by itself; if you're impatient there's a companion UnloadMiniCPM node. A ~7GB VLM sitting in VRAM next to a diffusion model is a recipe for OOM, and this node is designed so it doesn't stay resident.

The inputs that matter

• images - your batch (single image, several for comparison, or a frame sequence).
• prompt - what you ask the model.
• mode - single_image, multi_image (compare several), or video_frames (temporal understanding).
• system_prompt - presets from default and detailed to narrator, technical, accessible, creative, plus none and custom (your own text goes in custom_system_prompt).
• thinking_mode - fast vs deep_thinking (MiniCPM only).
• lock_output - returns the last cached answer without re-running inference. Hugely handy when you're iterating on downstream nodes and don't want to reload the model every run.
• max_tokens, temperature, seed - standard generation knobs.

Outputs

response (the text), images (passes your input through unchanged), and vram_cleared - a string that exists specifically to be wired into the pack's VRAM Gated loaders so they wait for the VLM to finish before loading a big checkpoint.

Troubleshooting

The most common failure is missing deps. MiniCPM needs transformers, accelerate, and bitsandbytes; Mage-VL needs transformers ≥ 5.7. The node tells you exactly what to pip install in the error string, which is more than most nodes do. First run downloads a few GB - do it when you're not mid-queue. And keep in mind the security shape of the category: any node that auto-downloads a model and runs it locally is arbitrary Python with your permissions, so install from the known repo and update it deliberately.

Installing

Part of TrentNodes (all under Trent/). ComfyUI Manager → "Trent Nodes" → Install; if Manager flags the pack (a day-one repo rename left a registry duplicate), clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/TrentHunter82/TrentNodes
cd TrentNodes
pip install -r requirements.txt

Restart, and the pack's VLM requirements are satisfied by the same install.