QwenVL-F Advanced (GGUF)

This is the node the fork is actually about. QwenVL-F Advanced (GGUF) is the screenshot on the README, the one with a knob for everything: the llama.cpp GGUF vision engine, an explicit mmproj_name selector, context and batch sizes, GPU offload, image token budgets, three image inputs, and even Gemma 4 support on top of the Qwen line. If the simple GGUF node is "run the defaults," this one is "build it your way."

The pitch: run Qwen3-VL vision-language models from single-file .gguf quants, so a Q4_K_M on a modest card behaves, and you control exactly how llama.cpp spends your memory. The cost is that every file and every knob is on you.

The dials that matter

• mmproj_name - the fork's signature addition. Every *mmproj*.gguf on disk shows up here; auto (default) picks the projector sitting next to your model. You only touch it when you have several projectors lying around or the auto-pick grabs the wrong one.
• ctx and n_batch - context length (default 32768, up to 262144) and batch size (default 8192). Bigger ctx means longer conversations/captions but more KV cache memory. n_batch is a speed/VRAM tradeoff.
• gpu_layers - how many layers offload to the GPU. -1 means all of them; on a low-VRAM card, start there and step down until it fits. Set this before anything else if llama.cpp is crashing on load.
• image_max_tokens / image_min_tokens - the token budget for the vision encoder (defaults 8192 / 1024). These control how much visual detail survives into the model; too low and fine details blur into caption mush.
• top_k, temperature, top_p, repetition_penalty - the sampling set. top_k 0 means off.
• frame_count for video, pool_size (llama.cpp's memory pool, default ~4 MB), stop_words, enable_thinking, keep_model_loaded, seed.

And the three IMAGE inputs (image, image2, image3) for multi-image analysis, plus video. One RESPONSE string out - wire it to a text encoder, a display, or a caption pipeline.

Install - same fork rule, heavier dependency

Model discovery is identical to the rest of the pack: the dropdowns scan models/text_encoders and models/LLM recursively, nothing downloads itself, and a missing model raises an error listing the folders checked. The dependency is the part to get right:

cd ComfyUI/custom_nodes
git clone https://github.com/id-fa/ComfyUI-QwenVL-F
pip install -r requirements.txt
python tools/install_helper.py --python "C:\AI\ComfyUI\python_embeded\python.exe" --run

The helper resolves your CUDA version (from nvidia-smi, falling back to nvcc, then torch.version.cuda) and installs the matching JamePeng llama-cpp-python vision wheel plus a CUDA PyTorch. It installs nothing unless you pass --run - run it once without to see the plan. The wheel must include the Qwen3VLChatHandler / Qwen25VLChatHandler chat handlers, or every run dies with an ImportError.

Common issues

• Gemma 4 .gguf - auto-detected by filename and switched to Gemma4ChatHandler, but that needs the JamePeng fork at v0.3.35+. Older llama-cpp-python raises ImportError on load; that's the error message telling you to update, not a model problem.
• OOM on load - drop gpu_layers, shrink ctx, or move image_max_tokens down before blaming the model.
• Wrong projector - if captions are garbage or the model won't load, check which mmproj is actually being used instead of trusting auto.
• Censored output - the Qwen3-VL Instruct quants carry the usual assistant refusals; if that matters for your use case, find an abliterated build.

Qwen3-VL-4B or 8B GGUF at Q4_K_M/Q8_0 is the sane starting point; on the GGUF ladder Q8 is essentially fp16 quality at half the size, so take it when it fits.