Save 4 hours! We auto-setup your workflow! Free!

Drop your workflow.json — we handle every dependency, custom node, and model. Just open the link and run.

Auto-Setup Workflow Json (Free) Now!
ComfyUI > Nodes > ComfyUI-llama_Dapao

ComfyUI Extension: ComfyUI-llama_Dapao

Repo Name

ComfyUI-llama_Dapao

Author
paolaoshi (Account age: 0 days)
Nodes
View all nodes(1)
Latest Updated
2026-11-14
Github Stars
0.07K

How to Install ComfyUI-llama_Dapao

Install this extension via the ComfyUI Manager by searching for ComfyUI-llama_Dapao
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter ComfyUI-llama_Dapao in the search bar
After installation, click the Restart button to restart ComfyUI. Then, manually refresh your browser to clear the cache and access the updated list of nodes.

Visit ComfyUI Online for ready-to-use ComfyUI environment

  • Free trial available
  • 16GB VRAM to 80GB VRAM GPU machines
  • 400+ preloaded models/nodes
  • Freedom to upload custom models/nodes
  • 200+ ready-to-run workflows
  • 100% private workspace with up to 200GB storage
  • Dedicated Support

Run ComfyUI Online

ComfyUI-llama_Dapao Description

ComfyUI-llama_Dapao adds local GGUF-powered chat and vision tools to ComfyUI, analyzing images and video frames, generating creative prompts, and managing face, clothing, and pose references without APIs.

ComfyUI-llama_Dapao Introduction

ComfyUI-llama_Dapao brings local language and vision-language tools into ComfyUI. It lets you use a GGUF model to chat, describe images, analyze video frames, and create prompts without relying on an online language-model API. A GGUF file is the model itself; a compatible mmproj file is also needed when a vision model must interpret images or video frames.

For an AI art workflow, the extension can help you turn references into descriptive prompts, organize material for a video prompt, or keep a creative conversation going inside ComfyUI. It also includes visual reference browsers for faces, clothing, and poses. The original llama node was adapted from ComfyUI-llama-cpp_vlm.

How ComfyUI-llama_Dapao Works

Think of a workflow as a small creative team. A model reads your instructions, a vision processor helps it look at images, and each node gives it a particular job—such as chatting, describing a reference image, or drafting a video prompt.

You choose a local GGUF model in a node. For visual tasks, you also choose the matching mmproj file. The node then passes your text and supported media to the model and returns text you can use elsewhere in your workflow. Model files in ComfyUI/models/LLM/ are discovered automatically.

The choice of model matters. A text-only model can work with written instructions, but it cannot inspect an image merely because one is connected to a workflow. For image or video-frame analysis, use a compatible vision model and its matching mmproj. Audio capability depends on the particular model and its llama.cpp handler; it is not guaranteed for every GGUF model.

ComfyUI-llama_Dapao Features

Face, clothing, and pose references

These nodes help you choose visual material before writing or refining a prompt:

  • Beauty face-model references (美女脸模素材@泡老师的小课堂) offers 1,499 original-size face and styling images. You can filter by hairstyle or look, use a six-image quick-selection view, and obtain positive and negative prompts in Chinese and English. See the face-reference guide.
  • Curated clothing references (精选服装参考@泡老师的小课堂) provides a six-image selection view, searches across eight clothing categories, and outputs original-size images plus Chinese and English positive and negative prompts. Personal material is configured locally and is not distributed with the node code. See the clothing-reference guide.
  • Curated pose references (精选姿势参考@泡老师的小课堂) includes 288 groups containing 1,439 compressed pose-reference images. You can search by category, explore five viewpoints, preview multiple angles, and send a selected reference through a standard IMAGE output. It does not require a model or an internet connection. See the pose-reference guide.

Local multimodal chat

The llama smart chat node combines model selection, settings, and conversation in one place. It supports text alongside up to 8 images, 2 videos, and 2 audio inputs. You can keep or disable multi-turn history, switch thinking mode for supporting models, let the node allocate VRAM automatically or set a manual limit, and optionally unload the model after inference to free VRAM.

For example, you could provide several character references and ask for a consistent description. Whether the model can interpret each media type still depends on the selected model and processor.

llama smart chat node

Local H3 video prompt generation

The H3 local video prompt generator uses a local GGUF multimodal model to organize text, images, video frames, and audio into a MiniMax H3-format prompt. It does not call an online H3 node or language-model API.

It automatically recognizes T2VA, I2VA, FL2VA, L2VA, and Ref2VA input modes. It accepts up to 9 source images, 3 video-frame batches, and 3 audio inputs, with a combined limit of 12 media items. Video arrives as time-ordered IMAGE frame batches, so this node does not need an additional video-decoding dependency.

By default, audio contributes locally calculated information such as its frequency spectrum, energy, and silence. Only enable raw audio input after confirming that your model supports it. The node checks field order, media numbering, and timestamps, and can optionally attempt one structural repair. For work involving several images, the author recommends a multi-image-capable vision model such as Qwen3.5 or Qwen3-VL with its matching mmproj.

Image and video reverse prompting

The llama image reverse-prompting node describes images or turns them into prompts. It handles batches of 1–8 images and loads prompt-style templates from the extension’s prompt/ folder. You can use the supplied styles or add your own text templates. If you enter an additional instruction, it takes priority over the selected built-in style—use this when you want a specific kind of description.

llama image reverse-prompting node

The reverse-prompt extra options node adds your selected instructions to image reverse prompting. Its 17 optional enhancements include character details, lighting, camera angle and settings, composition, depth of field, artistic-quality assessment, and safety rating. Selecting only the details relevant to your task can keep the result focused.

reverse-prompt extra options node

The extension also includes a llama video reverse-prompting node.

llama video reverse-prompting node

Local versions of prompt tools

The extension includes local-model versions of prompt tools adapted from ComfyUI-dapaoAPI. They cover e-commerce product-detail prompts, H3 prompt and video-prompt creation, general image prompts, Music3 music prompts, Seedance2 directing, and visual-style prompts. These nodes use the GGUF model you select rather than requiring an API key.

The tools retain their original inputs, media analysis, structured checks, and outputs. For visual work, select both a suitable main model and its matching mmproj. When you provide multiple images, they are passed one by one to the model during the same local inference.

Skill-based, multi-turn conversation workspace

The nodes under 🍭大炮-llama-cpp offer a more structured way to chat and manage creative tasks:

NodeWhat it does
Local model loader (🤖大炮本地模型加载器)Selects the GGUF model, mmproj, context, and model-unloading policy.
Conversation enhancement settings (⚙️对话增强设置)Sets the system prompt, number of context turns, sampling settings, and a 2K image-size limit.
Skill loader (🧩大炮Skill加载器)Selects Skills, manages their display names, accepts ZIP or folder uploads, and can use the connected local model to improve displayed names.
Multi-turn media library (📦大炮本地多轮对话素材库)Prepares up to 20 images, 5 videos, and 5 audio items for use in chat.
Multi-turn chat (💬大炮本地多轮对话)Handles conversation, Skills, media mentions, session controls, and final-result output.

Connect the model loader to multi-turn chat and, if you use Skills, to the Skill loader. Connect the Skill loader, conversation settings, and media library to multi-turn chat as needed.

Preparing media is not the same as sending it. To include a library item in the current message, type @ in the chat box and select a mention such as @图片1, @视频1, or @音频1. The conversation history stores the item number and written analysis, not the media itself. If you want the model to inspect the same item in a later turn, mention it again.

During inference, images are proportionally resized with Lanczos so their longest edge is no more than 2048 pixels, then handled as PNG. Videos are represented by evenly sampled frames, and audio is converted to 16 kHz mono WAV. Images and video require a matching mmproj; audio support depends on the model and handler, and the node reports when they do not support it.

The chat node has three text outputs: assistant reply, conversation history JSON, and Skill final result. Its controls serve different purposes:

  • Send runs the chat node and the necessary upstream nodes.
  • Clear context leaves the visible chat record in place but stops passing older messages to the model.
  • Send final state does not run the local model. It passes the Skill final result to downstream nodes, or the latest assistant reply if there is no Skill result.

The interface also supports safe Markdown display, code copying, conversation import and export, editing and resending history, deleting messages from a selected point, a 60-second undo window, and adjustable node height.

Skills live in the extension’s skills/ folder. The supported layouts are skills/<skill-id>/SKILL.md and skills/<repo>/skills/<skill-id>/SKILL.md. Uploads are checked for unsafe paths, symbolic links, duplicate paths, file counts, and extracted size; a folder with the same name is not automatically overwritten. Display names are stored separately in data/skill_display_names.json. Changing a display name does not alter a Skill’s contents, ID, triggers, or behavior.

ComfyUI-llama_Dapao Models

The processor setting tells the node how to work with the family of model you selected. It is not a replacement for the actual GGUF model file. Choose a compatible processor and, for vision models, the corresponding mmproj.

ProcessorIntended use
NoneText-only models; no mmproj needed.
LLaVA-1.5 / LLaVA-1.6LLaVA vision models.
Moondream2Lightweight vision model.
MiniCPM-v2.6 / MiniCPM-v4.5MiniCPM vision models.
Gemma3 / Gemma4Gemma vision models.
Qwen3.8Qwen3.8-VL multimodal model; requires a matching mmproj.
Qwen2.5-VL / Qwen3-VLQwen vision models.
Qwen3.5Qwen3.5 multimodal model.
GLM-4.6V / GLM-4.1VGLM vision models.
LFM2-VLLiquid vision model.
Granite-DoclingDocument-understanding model.

A practical starting point is None for text-only prompt drafting, a supported vision processor for describing an image, and a multi-image-capable vision model for comparing several references or building an H3 prompt from them. Output quality and speed depend on the particular model, files, and hardware you use.

Qwen3.8 note: Its GGUF architecture identifier is qwen35. The node offers reasoning-strength choices translated as Off, Auto, Low, Medium, and High, with Off as the default; Auto and High correspond to the model’s native xhigh setting. Processors with a -Thinking suffix enable thinking mode by default.

Troubleshooting ComfyUI-llama_Dapao

The model cannot see my image or video frames

  1. Check that you selected a vision-capable GGUF model, not a text-only model.
  2. Select the processor appropriate to that model.
  3. Confirm that its matching mmproj is available and selected.
  4. In multi-turn chat, type @ and select the media item for the current message. A connected media library alone does not send it.

A media item worked earlier, but the model cannot revisit it

Mention the item again—for example, @图片1—in your new message. Earlier chat turns retain the item number and written analysis, but do not repeatedly inject the original media.

Audio is rejected or cannot be interpreted

Check whether your particular GGUF model and llama.cpp ChatHandler support audio. The multi-turn chat node reports unsupported audio. For the H3 prompt generator, leave raw audio input off unless you have confirmed model support; its default local audio analysis can still provide spectrum, energy, and silence information.

Qwen3.8 fails with missing tensor 'blk.64.ssm_conv1d.weight'

This error is associated with an older llama-cpp-python build, such as 0.3.35, that cannot load Qwen3.8 GGUF files containing MTP/NextN layers. Qwen3.8 requires JamePeng llama-cpp-python 0.3.47 or newer. See the compatible builds on the JamePeng/llama-cpp-python releases page. After updating the runtime used by ComfyUI, restart ComfyUI.

I need to free GPU memory after a response

Enable the option to unload the model after inference in the relevant node, or adjust the unloading policy in the local model loader. If needed, use the manual VRAM limit instead of automatic allocation.

Frequently asked questions

Do these local prompt tools need an API key?
No. They use the local GGUF model you select.

Does “Clear context” delete the visible chat?
No. The visible record remains, but older messages are no longer sent to the model as context.

Does “Send final state” generate another reply?
No. It forwards the Skill final result—or the most recent assistant reply—to downstream nodes without running the local model.

Will renaming a Skill change what it does?
No. A display name is separate from the Skill’s content, ID, triggers, and functionality.

Learn More about ComfyUI-llama_Dapao

ComfyUI-llama_Dapao Related Nodes

RunComfy
Copyright 2025 RunComfy. All Rights Reserved.

RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.

ComfyUI-llama_Dapao detailed guide | ComfyUI