Nodes/ComfyUI-QwenVL-F/QwenVL-F (Advanced)
ComfyUI Node

QwenVL-F (Advanced)

A ComfyUI node in QwenVL-F with 21 inputs and 1 output.

By id-fa·Created 5 months ago·Updated 9 days ago· 1
QwenVL-F (Advanced)
  • image
  • image2
  • image3
  • video
  • RESPONSE
model_name(no models found — see console)
quantizationNone (FP16)
attention_modeauto
use_torch_compilefalse
deviceauto
preset_prompt🖼️ Detailed Description
custom_prompt
max_tokens512
temperature0.60
top_p0.90
num_beams1
repetition_penalty1.20
frame_count16
enable_thinkingfalse
stop_words
keep_model_loadedtrue
seed1
CategoryQwenVL-F

Inputs (21)

NameTypeDefaultDescription
model_nameCOMBO(no models found — see console)Pick a Transformers checkpoint already present under models/text_encoders or models/LLM. Nothing is downloaded automatically — copy the model folder in yourself, then reload ComfyUI.
quantizationCOMBONone (FP16)Precision vs VRAM. FP16 gives the best quality if memory allows; 8-bit suits 8–16 GB GPUs; 4-bit fits 6 GB or lower but is slower.
attention_modeCOMBOautoauto tries flash-attn v2 when installed and falls back to SDPA. Only override when debugging attention backends.
use_torch_compileBOOLEANfalseEnable torch.compile('reduce-overhead') on supported CUDA/Torch 2.1+ builds for extra throughput after the first compile.
deviceCOMBOautoChoose where to run the model: auto, cpu, mps, or cuda:x for multi-GPU systems.
preset_promptCOMBO🖼️ Detailed DescriptionBuilt-in instruction describing how Qwen-VL should analyze the media input.
custom_promptSTRINGOptional override—when filled it completely replaces the preset template.
max_tokensINT51264–32768Maximum number of new tokens to decode. Larger values yield longer answers but consume more time and memory.
temperatureFLOAT0.600.1–1Sampling randomness when num_beams == 1. 0.2–0.4 is focused, 0.7+ is creative.
top_pFLOAT0.900–1Nucleus sampling cutoff when num_beams == 1. Lower values keep only top tokens; 0.9–0.95 allows more variety.
num_beamsINT11–8Beam-search width. Values >1 disable temperature/top_p and trade speed for more stable answers.
repetition_penaltyFLOAT1.200.5–2Values >1 (e.g., 1.1–1.3) penalize repeated phrases; 1.0 leaves logits untouched.
frame_countINT161–64Number of frames extracted from video inputs before prompting Qwen-VL. More frames provide context but cost time.
enable_thinkingBOOLEANfalseEnable thinking mode for Qwen3-VL Thinking models. When disabled, the model skips chain-of-thought reasoning and responds directly. Has no effect on non-Thinking models.
stop_wordsSTRINGComma-separated list of stop words/sequences. Generation stops when any of these strings is produced. Leave empty for default behavior.
keep_model_loadedBOOLEANtrueKeeps the model resident in VRAM/RAM after the run so the next prompt skips loading.
seedINT11–4294967295Seed controlling sampling and frame picking; reuse it to reproduce results.
imageoptIMAGE
image2optIMAGE
image3optIMAGE
videooptIMAGE

Outputs (1)

NameTypeDescription
RESPONSESTRING