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-QwenVL-aistudynow > QwenVL Advanced (GGUF)

ComfyUI Node: QwenVL Advanced (GGUF)

Class Name

aistudynow_QwenVL_GGUF_Advanced

Category
🧠aistudynow/QwenVL
Author
aistudynow (Account age: 0days)
Extension
ComfyUI-QwenVL-aistudynow
Latest Updated
2026-06-23
Github Stars
0.02K

How to Install ComfyUI-QwenVL-aistudynow

Install this extension via the ComfyUI Manager by searching for ComfyUI-QwenVL-aistudynow
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter ComfyUI-QwenVL-aistudynow 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

QwenVL Advanced (GGUF) Description

Runs local Qwen VL GGUF models to answer prompts about images, sampled video frames, or text.

QwenVL Advanced (GGUF):

This node runs a locally loaded Qwen vision-language model in GGUF format to generate a written response to a prompt. You can use it to describe an image, ask questions about visual content, or analyze sampled frames from a video supplied as an image sequence. It also works with a text prompt alone. The node selects a model from its configured catalog, downloads missing model files when a repository is available, and uses a compatible vision component to pass images to the model. Its advanced controls let you balance response length, generation style, and the computing resources used. The resulting text is cleaned before it is returned.

QwenVL Advanced (GGUF) Input Parameters:

model_name

Selects the GGUF model to load. The menu contains configured catalog entries that specify a vision component; its exact choices depend on your catalog. The default is the first model name in sorted order. If no eligible entry is configured, the menu shows (edit gguf_models.json) as a placeholder rather than a usable model. Choosing a different model can change both the quality of the response and the resources needed to run it.

device

Chooses where the model is intended to run. Options are auto, cpu, mps, and one cuda:N option for each CUDA device detected by PyTorch; the default is auto. Use cpu when GPU execution is unavailable, or select an available GPU option when appropriate. The actual amount of model work placed on a CUDA GPU also depends on gpu_layers; selecting CUDA with zero GPU layers triggers a warning that the model will run on the CPU.

preset_prompt

Selects a predefined task prompt, such as an available image-description preset. The choices come from the configured preset list; if that list is unavailable, the source supplies a detailed-description choice and an anatomy-focused prompt as fallback entries. The default is 🖼️ Detailed Description when present, otherwise the first available choice. The selected prompt tells the model what kind of answer to produce, unless a nonblank custom_prompt replaces it.

custom_prompt

Provides your own task instructions as multiline text. The default is an empty string, and the source specifies no minimum or maximum length. When you enter nonblank text, it replaces the selected preset prompt rather than being added to it. For example, you could ask for a concise description of a character’s pose instead of using a general description preset.

max_tokens

Sets the maximum number of tokens the model may generate for its response. The default is 512, with a minimum of 64 and a maximum of 4096. A higher value gives the model room for a longer answer but may take more time; a lower value helps keep answers brief and may stop a detailed answer before it is complete.

temperature

Controls variation in the generated wording. The default is 0.6, with a minimum of 0.0 and a maximum of 2.0. Lower settings generally favor more predictable responses, while higher settings allow more variation. For descriptions where consistency matters, start near the default or lower it.

top_p

Limits generation choices by cumulative probability. The default is 0.9, with a minimum of 0.0 and a maximum of 1.0. Lower values narrow the words the model considers and can make responses more focused; values nearer 1.0 leave more possibilities available. It works alongside temperature.

repetition_penalty

Controls the penalty applied when the model repeats tokens. The default is 1.2, with a minimum of 0.5 and a maximum of 2.0. Raising it may reduce repeated phrases, while an excessive setting may make necessary repeated terms harder for the model to use. The value is passed to the model as its repeat-penalty setting.

frame_count

Sets how many frames the node requests from the connected video input for sampling. The default is 16, with a minimum of 1 and a maximum of 64. More frames can give the model more visual material from a changing scene, but they also increase the amount of image content it must process. This setting does not affect a standalone image input.

ctx

Sets the model’s context length, which is the capacity available for the prompt, image-related content, and generated response. The default is 8192, with a minimum of 1024, a maximum of 262144, and steps of 512. A larger setting may help with longer prompts or more visual content, but it can require substantially more memory. The configured model and backend must also be able to support the chosen value.

n_batch

Sets the backend’s processing batch size. The default is 512, with a minimum of 64, a maximum of 32768, and steps of 64. Increasing it may improve processing speed on a machine with sufficient memory; reducing it is a practical step when loading or processing runs out of memory. It does not set the number of responses.

gpu_layers

Sets the GPU-layer value passed to the backend when the selected device resolves to CUDA. The default is -1, with a minimum of -1 and a maximum of 200. The source passes -1 through to the installed backend without defining its meaning, so check that backend’s behavior before relying on it. When the device does not resolve to CUDA, the node uses 0 GPU layers regardless of this setting.

image_max_tokens

Sets the image-token budget requested from the vision chat handler and, when supported, the model backend. The default is 4096, with a minimum of 256, a maximum of 1024000, and steps of 256. A larger value may allow more image content to be processed but can increase context and memory demands. If the installed chat handler does not support this setting, the node warns that the image budget will instead be controlled by ctx.

top_k

Sets the top-k value passed to the model backend. The default is 0, with a minimum of 0 and a maximum of 32768. This setting can affect how broadly the backend considers candidate tokens during generation. Because the source does not define what 0 means for every supported backend, leave the default in place unless you know how your installed backend interprets it.

pool_size

Sets a pool-size value passed to the model backend when that backend accepts the argument. The default is 4194304, with a minimum of 1048576, a maximum of 10485760, and steps of 524288. It is a backend resource setting, not a measure of image count or answer length. The source does not establish a universal performance effect, so retain the default unless your backend’s documentation gives you a reason to adjust it.

keep_model_loaded

Determines whether the node retains the loaded model after producing a response. The default is True; the options are True and False. Keeping it loaded can avoid a new load on a later run with the same model and loading settings, but it continues to occupy resources. Setting it to False makes the node clear the model after each run.

seed

Provides the integer seed used for PyTorch and passed to response generation. The default is 1, with a minimum of 1 and a maximum of 4294967295. Keeping it fixed can help you compare prompt or setting changes under similar generation conditions, although identical results are not guaranteed across different models, backends, or environments.

image

An optional IMAGE input for a still image you want the model to examine. There is no default image and no numeric range. The node converts a supplied image to PNG data and includes it in the request when a compatible vision component is loaded. If the selected model entry has no usable vision component, the image will not be used; you can also leave this input disconnected for a text-only request.

video

An optional IMAGE input used as the video source for frame sampling. There is no default video and no numeric range. The node samples frames according to frame_count, converts them to PNG data, and sends them as images when vision processing is available. You may connect both image and video; the still image is added before the sampled video frames.

QwenVL Advanced (GGUF) Output Parameters:

RESPONSE

A single STRING containing the model’s generated answer. It reflects the selected preset or replacement custom prompt and any visual content the installed backend can process. The node cleans the generated content and removes surrounding whitespace before returning it. Treat the response as the model’s interpretation of the material, not as a guaranteed factual description; check important visual details against the original image or frames.

QwenVL Advanced (GGUF) Usage Tips:

  • For a specific visual task, write a direct custom_prompt that names what you want examined, such as pose, clothing, or objects in the scene. Leave it blank when you want the selected preset to apply.
  • Start video analysis with the default frame_count of 16. Increase it when important changes occur across the clip; decrease it when you need to reduce the amount of visual content processed.
  • Keep keep_model_loaded enabled for repeated runs with the same loading settings. Disable it when you need the node to release the model after each response.
  • If memory is limited, try a smaller ctx or n_batch and fewer sampled frames before increasing visual budgets. Change one setting at a time so you can tell what helped.
  • For concise captions, lower max_tokens and explicitly request a short answer. For detailed observations, allow more tokens, while staying within the available context and memory.
  • When checking whether a wording change improved a result, keep seed and generation settings fixed for the comparison. This makes the runs easier to evaluate, though it does not guarantee identical output.

QwenVL Advanced (GGUF) Common Errors and Solutions:

[QwenVL] llama_cpp is not available. Install the GGUF vision dependency first. See docs/GGUF_MANUAL_INSTALL.md

  • Explanation: The node cannot import the backend required to run the GGUF model.
  • Solution: Install the project’s required GGUF vision dependency in the environment running the node, following docs/GGUF_MANUAL_INSTALL.md, then restart that environment.

[QwenVL] GGUF model not found locally and no repo_id provided: …

  • Explanation: The selected catalog entry’s model file is missing from its expected local location, and the entry does not provide a repository from which the node can download it.
  • Solution: Check the catalog entry and place the named GGUF file in its configured model directory, or configure a valid repository for that entry.

[QwenVL] mmproj not found locally and no repo_id provided: …

  • Explanation: The model’s vision-component file is missing, and no download repository is configured.
  • Solution: Check the vision-component filename in the catalog and place the matching file in the model directory, or configure a repository that contains it.

[QwenVL] Missing Qwen VL chat handler in llama_cpp. Install the correct fork/wheel. See docs/GGUF_MANUAL_INSTALL.md

  • Explanation: A vision-component file is present, but the installed backend provides neither of the Qwen vision chat handlers this node tries to use.
  • Solution: Install a compatible multimodal backend build as directed by docs/GGUF_MANUAL_INSTALL.md, then restart the environment.

[QwenVL] Warning: images provided but this model entry has no mmproj_file; images will be ignored

  • Explanation: You connected visual content, but the selected catalog entry has no configured vision-component file. The request proceeds without those images.
  • Solution: Select a properly configured vision model entry and confirm that its vision-component file is available.

[QwenVL] Warning: installed llama_cpp Llama() does not accept chat_handler; images will be ignored. Update llama-cpp-python to a multimodal-capable build.

  • Explanation: The node found a vision chat handler, but the installed model backend cannot accept it. Visual content will not reach the model.
  • Solution: Install a multimodal-capable llama-cpp-python build compatible with the project and restart the environment.

[QwenVL] Warning: installed llama_cpp chat handler does not support image_max_tokens; image token budget will be controlled by ctx only.

  • Explanation: The installed vision handler does not accept the image_max_tokens control. This is a warning, not necessarily a failed run.
  • Solution: Use ctx to manage the available context within your memory limits, or install a compatible handler build if you need direct control of the image-token budget.

[QwenVL] Warning: device=cuda selected but n_gpu_layers=0; model will run on CPU.

  • Explanation: CUDA was selected, but the GPU-layer setting resolved to zero, so the model will not use GPU layers.
  • Solution: If you intend to use a CUDA GPU, choose an appropriate gpu_layers value for your backend and available GPU memory. Otherwise, select cpu to make the intended configuration clear.

LLM not initialized properly

  • Explanation: The loaded model object does not provide the chat-completion method this node requires.
  • Solution: Check that the installed backend build is compatible with the node’s GGUF chat workflow, then reload the model after correcting the installation.

Model download fails or a repository file cannot be found

  • Explanation: When a configured file is absent locally, the node attempts to download it from the catalog’s repository choices. The download may fail because of connectivity, access restrictions, or a filename that does not match the repository.
  • Solution: Check your connection and repository access, verify the catalog’s repository ID and exact filenames, or place the required files in the expected local directory.

Memory error while loading or processing

  • Explanation: The selected model or settings may require more system or GPU memory than is available, particularly with a large context or substantial visual content.
  • Solution: Try a smaller model, reduce ctx or n_batch, lower frame_count or image_max_tokens where supported, and close other memory-intensive tasks before running again.

QwenVL Advanced (GGUF) Related Nodes

Go back to the extension to check out more related nodes.
ComfyUI-QwenVL-aistudynow
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.

QwenVL Advanced (GGUF)