Nodes/ComfyUI-FL-VLM/FL VLM Image Inference
ComfyUI Node

FL VLM Image Inference

A ComfyUI node in FL/VLM with 19 inputs and 2 outputs.

By filliptm·Created 2 months ago·Updated 2 months ago· 0
FL VLM Image Inference
  • vlm_model
  • image
  • response
  • metadata_json
system_promptYou are a precise vision-language assistant. Answer only from the provided image content.
promptDescribe the image in detail.
batch_modeall_images_one_prompt
image_start0
image_stride1
max_images16
min_pixels64
max_pixels256
max_new_tokens512
do_samplefalse
temperature0.70
top_p0.90
top_k50
repetition_penalty1.00
seed-1
cleanup_temp_imagestrue
resize_preset360p_balanced
CategoryFL/VLM

Inputs (19)

NameTypeDefaultDescription
vlm_modelFL_VLM_MODEL
imageIMAGE
system_promptSTRINGYou are a precise vision-language assistant. Answer only from the provided image content.
promptSTRINGDescribe the image in detail.
batch_modeCOMBOall_images_one_prompt2 options: all_images_one_prompt, each_image_separately
image_startINT00–100000
image_strideINT11–100000
max_imagesINT161–512
min_pixelsINT6464–4096Qwen visual token budget lower bound. Internally multiplied by 28*28.
max_pixelsINT25664–8192Qwen visual token budget upper bound for manual_qwen_budget. Internally multiplied by 28*28.
max_new_tokensINT5121–8192
do_sampleBOOLEANfalse
temperatureFLOAT0.700–5
top_pFLOAT0.900–1
top_kINT500–500
repetition_penaltyFLOAT1.000.1–5
seedINT-1-1–18446744073709550000
cleanup_temp_imagesBOOLEANtrue
resize_presetCOMBO360p_balancedControls Qwen's image resize budget before inference. Existing min_pixels/max_pixels are only used by manual_qwen_budget.

Outputs (2)

NameTypeDescription
responseSTRING
metadata_jsonSTRING