Nodes/ComfyUI-Photoreal-Prompt-Builder/VLM Prompt Refiner (kppb)
ComfyUI Node

VLM Prompt Refiner (kppb)

A ComfyUI node in conditioning/klein with 21 inputs and 4 outputs.

By artokun·Created 6 months ago·Updated 2 months ago· 21
VLM Prompt Refiner (kppb)
  • character_ref
  • scene_ref
  • prop_ref
  • refined_prompt
  • image_caption
  • filename_prefix
  • video_prompt
ollama_urlhttp://localhost:11434
modelhuihui_ai/qwen3-vl-abliterated:32b-instruct-q8_0
modedescribe & enhance
positive_prompt
prompt_json
edit_prompt
system_prompt
temperature0.30
seed-1
preserve_identitytrue
unload_modeltrue
use_claude_codefalse
claude_modelopus
trigger_wordohwx
sfw_promptfalse
generate_video_promptfalse
motion_prompt
audio_prompt
Categoryconditioning/klein

Inputs (21)

NameTypeDefaultDescription
character_refIMAGECharacter reference image — identity/likeness is extracted from this
ollama_urlSTRINGhttp://localhost:11434
modelSTRINGhuihui_ai/qwen3-vl-abliterated:32b-instruct-q8_0
modeCOMBOdescribe & enhance4 options: describe & enhance, image edit aware, caption only, dataset generation
scene_refoptIMAGEAdditional reference for scene, pose, lighting, or outfit
prop_refoptIMAGEProduct/prop reference — handbag, sunglasses, sponsored item, etc.
positive_promptoptSTRINGConnect from Prompt Builder's positive_prompt — used as fallback if VLM returns empty
prompt_jsonoptSTRINGConnect from Prompt Builder's prompt_json output
edit_promptoptSTRINGConnect from Image Edit Composer's edit_prompt output
system_promptoptSTRING
temperatureoptFLOAT0.300–2Lower = more focused, higher = more creative
seedoptINT-1-1–2147483647-1 for random, or set for reproducible output
preserve_identityoptBOOLEANtrueAppend identity lock phrase to reinforce likeness preservation for the diffusion model
unload_modeloptBOOLEANtrueUnload Ollama LLM from VRAM after inference. Turn ON for localhost (frees VRAM for Klein). Turn OFF for remote/RunPod (avoids slow reload between iterations)
use_claude_codeoptBOOLEANfalseUse Claude Code CLI instead of Ollama — does both stages in one shot with images
claude_modeloptCOMBOopusClaude model to use (sonnet recommended for speed/quality balance)
trigger_wordoptSTRINGohwxTrigger word for LoRA training captions (dataset generation mode only)
sfw_promptoptBOOLEANfalseSoften the output prompt to pass content safety filters on hosted models like nanobanana/Gemini. Leave off for local models.
generate_video_promptoptBOOLEANfalseGenerate a Wan 2.6 I2V motion prompt from the scene settings. Turn off to save LLM overhead on smaller models.
motion_promptoptSTRING
audio_promptoptSTRING

Outputs (4)

NameTypeDescription
refined_promptSTRING
image_captionSTRING
filename_prefixSTRING
video_promptSTRING