PATRYK KULPOK
← BACK TO SYSTEM MAP

SYSTEM 03

PROMPT DIRECTOR

MULTIMODAL AI PROMPT ORCHESTRATION

Local multimodal tool for analysing image and video references and composing structured prompts for generative AI workflows.

PYTHONQWEN3-VLHUGGING FACEFFMPEGMULTIMODAL AILOCAL AI

SYSTEM IDENTITY

SYSTEM IDPROMPT-DIRECTOR-03

STATUSACTIVE R&D

TYPEMULTIMODAL AI / VISION / LOCAL MODELS

Reference analysis and prompt composition pipeline for image and video generation workflows.

01 / OVERVIEW

PROJECT OVERVIEW

Generative AI workflows often use multiple references, and each reference can cover a different aspect of the desired output.

Prompt Director groups this context from analysis to routing and turns it into a structured prompt shape before the target generation stage.

Prompt flow

  1. PROBLEM
  2. ANALYSIS
  3. ROUTING
  4. STRUCTURED PROMPT

02 / USE CASES

REFERENCE WORKFLOWS

UC-01

ANALYSE IMAGE REFERENCES

Turn local visual references into structured descriptions for later prompt composition.

UC-02

ANALYSE VIDEO REFERENCES

Prepare motion and scene context from local video material when a workflow needs it.

UC-03

ROUTE REFERENCE DETAILS

Map separate references to characteristics such as identity, wardrobe, action or lighting.

UC-04

COMPOSE STRUCTURED PROMPTS

Combine analysis, routed context and user intent into a generation-ready prompt shape.

03 / MULTIMODAL INPUT

MULTIMODAL INPUT

The tool handles local image and video references as input materials, with optional grouped references for richer prompt context.

Input family

  1. IMAGE REFERENCE
  2. VIDEO REFERENCE
  3. OPTIONAL MULTIPLE REFERENCES
IMAGE REFERENCE

Single image or pair of images prepared locally.

VIDEO REFERENCE

Local capture or local file as a motion source.

OPTIONAL MULTIPLE REFERENCES

Optional grouped inputs routed independently by target characteristic.

04 / ANALYSIS PIPELINE

ANALYSIS PIPELINE

Input sources first pass through local routing. Video streams may use frame preparation with FFMPEG when needed, while OCR is applied only when the image content includes text to enrich descriptors.

Analysis flow

  1. IMAGE / VIDEO
  2. MEDIA INPUT ROUTING
  3. QWEN3-VL ANALYSIS
  1. VISUAL ANALYSIS
  2. STRUCTURED DESCRIPTION

Optional branch map

VIDEO STREAM

FFMPEG for frame extraction before Qwen3-VL visual context can be used.

IMAGE STREAM

OCR can run as an optional step when text is detected in visual references.

05 / REFERENCE ROUTING

REFERENCE ROUTING

Each source is mapped to outputs that are meaningful for later prompt composition.

Reference routing topology

REFERENCE A

PORTRAIT

  • FACIAL IDENTITY

REFERENCE B

BODY

  • BODY PROPORTIONS

REFERENCE C

VIDEO

  • HAIR
  • WARDROBE
  • ACTION
  • MOTION
  • CAMERA
  • BACKGROUND
  • LIGHTING

06 / PROMPT COMPOSITION

PROMPT COMPOSITION

Composition combines structured analysis, routing output and explicit intent. The output format is designed for R&D exploration with MiniMax H3 / reference generation prompts.

Prompt composition map

  1. ANALYSIS
  2. REFERENCE ROUTING
  3. USER INTENT
  4. STRUCTURED GENERATION PROMPT
  5. TARGET GENERATIVE WORKFLOW

07 / TECH STACK

TECH STACK

  • Python
  • Qwen3-VL-8B-Instruct
  • Hugging Face Transformers
  • FFmpeg
  • Multimodal AI
  • Computer Vision
  • OCR
  • Local AI

08 / STATUS

PROJECT STATE

SYSTEM ID PROMPT-DIRECTOR-03

STATE ACTIVE R&D

DEPTH MULTIMODAL GENERATIVE AI TOOLING