Composable AI Tool Stack
Targeted augmentation. Measurable outcomes.
A pragmatic suite of AI-augmented capabilities embedded across analysis, authoring, quality, performance and operational pipelines - selectively integrated where leverage is real.
Image Generation, Upscaling & Training Tooling
Internal command-line tools for AI image workflows used in experiments, case studies, and client work.
Related Case Study
These tools were used to build the AI Influencer workflow - generating training datasets, face-swapped images, and multi-angle character views.
View AI Influencer Case Study
View All Tools & Commands
12 commands spanning generation, enhancement, face swap, and utilities
Click to expand
View All Tools & Commands
12 commands spanning generation, enhancement, face swap, and utilities
| Name | Description |
|---|---|
generate_fusion
|
Face-swapped image generation with LoRA model, face swap models (Easel/CodePlug), and optional restoration |
generate_story
|
Generate narrative story sequences (typically 20 scenes) with consistent character appearance using LoRA models |
cast_story_scenes
|
Two-stage absurdist pipeline: Transform realistic story scenes into absurdist satirical versions using Nano Banana. Adds surrealism, exaggeration, meta-humor while maintaining narrative structure. |
generate_multiangle
|
Generate rotated views of subjects using Nano Banana for 360° turnarounds and character sheets |
generate_lora_images_v2
|
Scene-aware LoRA training with structured prompts from JSON library. Supports multi-angle turnarounds (8 angles). |
generate_model_comparison
|
Generate side-by-side comparisons of same prompt across v2, v3, v4 models. Essential for model evaluation and quality tracking. |
generate_v4_images
|
Showcase generation with edwin-avatar-v4 model. Random sampling across categories for portfolio and demonstration. |
generate_training_images
|
Legacy training dataset generator. Superseded by generate_lora_images_v2. |
ai_editor
|
Targeted image editing with natural language: Make specific changes to images without regenerating entire scenes. LPA-aligned prompts for material upgrades, object additions, lighting adjustments. Interactive + CLI modes. |
refine_training_images
|
Enhance training images with Nano Banana while preserving facial identity and features |
upscale_ai_image
|
Image enhancement with Nano Banana (professional photography quality) + optional Real-ESRGAN upscaling. Default: enhance-only for maximum detail preservation |
enhance_existing_images
|
Batch enhance existing AI-generated images while preserving character identity from custom LoRA training with conservative, balanced, or upscale-only methods |
faceswap_story_scenes
|
Apply face swapping to story scene images using Easel Advanced or CodePlug models with optional GFPGAN face restoration |
convert_training_images
|
Convert WebP training images to PNG format with quality preservation and transparency handling (RGBA to RGB with white background) |
shuffle_training_images
|
Shuffle and randomly rename training images with alphanumeric names for dataset anonymization and preparation |
extract_last_frame
|
Extract the last frame or specific timestamp frame from video files using OpenCV for keyframe creation and video workflows |
generate_cinematic_video
|
Generate videos with cinematic camera movements using prompt engineering (dolly, pan, crane, orbit, zoom) - no proprietary APIs needed, just smart prompts with your existing video models |
apply_visual_effects
|
Apply cinematic visual effects (glow, particles, sparkles, lens flare, film grain, transformations) using img2img prompt engineering - replicates Higgsfield-style effects without third-party APIs |
Command Categories
Generation, Enhancement, Training, Face Swap, Utility
Click to expand
Command Categories
Generation, Enhancement, Training, Face Swap, Utility
Generation
Create new AI images from prompts
Enhancement
Upscale & refine existing images
Training
Dataset prep for LoRA fine-tuning
Face Swap
Identity transfer & face replacement
Utility
Format conversion & preprocessing
Real-World Workflow Examples
End-to-end pipelines with actual costs and processes
View Full Build →
Real-World Workflow Examples
End-to-end pipelines with actual costs and processes
End-to-end pipelines showing actual costs and processes for custom AI character creation
Custom LoRA Training
TrainingTrain an AI model on a single person (84 photos) to generate consistent character images across any scene or pose
- Initial LoRA training $2.00
- Generate 100 base images $0.40
- Nano Banana scene castingCast character into 100 new environments $3.50
- V3 model fine-tuning $2.00
- Complete Pipeline $7.90
Result: Production-ready AI character model trained exclusively on one person
AI Influencer Content
GenerationCreate brand-consistent content with custom-trained model, face swapping, and professional enhancement
- 50 images (custom LoRA) $0.50
- Face swap (Easel)Replace face with real reference photo $2.50
- Face restoration (GFPGAN) $0.10
- Format conversion Free
- 50 Pro Images $3.10
Result: Portfolio-grade content maintaining perfect identity across scenarios
Narrative Story Sequence
StoryGenerate 20-scene visual narrative with consistent character across different settings and angles
- 20 scene generations $0.20
- Multi-angle views (8 angles)360° character turnaround $0.28
- 5 key scene refinements $0.18
- Video extraction utilities Free
- Complete Story $0.66
Result: Cohesive visual narrative ready for animation or presentation
Practical Cost Breakdown: From Training to 30s Video
Actual costs from the AI Influencer buildMost people just prompt Sora or Veo directly and get generic results - anyone's face, inconsistent identity, no brand coherence. This breakdown shows what it actually costs to build a custom AI character that looks like you across unlimited scenarios, with character persistence good enough to build a brand around.
The Result
A custom LoRA model that generates your specific face in any scene, plus 50+ portfolio-ready images showing character consistency across professional, lifestyle, travel, and tech scenarios. Then video generation using those trained images as keyframes - not random AI faces.
Phase 1: Foundation
~$7Portfolio of 50+ character-consistent images across scenarios
Phase 2: 30s Video Creation
$0.50–$183× 10s clips @ $0.15-0.50 ea • 1080p • Best value
3× 10s clips @ $0.50 ea • 1080p • Good quality
4× 8s clips @ $3 ea • 1080p • Native audio
3× 10s clips @ $6 ea • Premium quality • Audio sync
All models support image-to-video from your trained character images
Complete custom AI Influencer (API) Build
Custom model + 50 images + 30s video with audio
$7.50–$25
depending on video model
SeeDance Route
8.5/10$7.50
$0.05/s
★ Best Value
17× cheaper, 85% quality
Hailuo Route
9/10$8.50
$0.05/s
★ Quality Leader
Tops benchmarks
Veo 3 Route
9.5/10$19
$0.40/s
Premium
Native audio
Sora 2 Route
10/10$25
$0.60/s
Top Tier
Best overall
Quality vs. Cost Efficiency
Quality scores from benchmark tests & user reviews • Hailuo 02 topped Veo 3 in recent comparisons
Foundation cost is one-time • Video generation costs scale with duration
Alternative: Cloud Subscription Services
For 30s videoIf you prefer subscription-based tools with polished UIs over API access, here's what 30s of video costs on popular platforms. Note: These don't include custom LoRA training for your specific face.
Higgsfield
$9/moFor 30s:
Budget models (Hailuo, SeeDance)
$9.00
+20%✗ No custom face training
Runway
$15/moFor 30s:
Gen-4 Turbo
$15.00
+64%✗ Generic characters only
Kling AI
$6.60/moFor 30s:
Standard quality
$7.00
-7%✗ No face persistence
Luma Ray2
$10/moFor 30s:
Ray2 Flash
$10.00
+36%✗ Random faces each time
The Real Advantage: Custom Face + API Control
Cloud services offer polished UIs but use generic models - you can't train them on YOUR face for character consistency. The API route includes custom LoRA training (50 images of you), giving you a persistent digital twin + automation capabilities. Costs are similar, but the API gives you ownership and creative control.
What is LoRA Training?
LoRA (Low-Rank Adaptation) fine-tunes an AI model on a specific person using 50-100 photos. The result is a custom trigger word (like "TOK") that generates that exact person in any scenario while maintaining their unique features, expressions, and identity. This is how we create AI influencers that look consistently like one real person across thousands of generated images.
What is Scene Casting?
Scene casting uses Nano Banana to take a generated character and "teleport" them into completely different environments while preserving their exact identity. Instead of just upscaling one image, you create diverse training data by placing the same person in European streets, mountain lakes, urban rooftops, Japanese gardens, etc. This dramatically expands dataset variety from a single generation pass.
Layered Prompt Architecture (LPA)
Universal 7-layer prompting framework achieving 90-100% facial likeness for LoRA-based image generation. Separates identity, positioning, environment, lighting, color, camera, and quality into discrete non-interfering layers.
Current Integration: Automated image generation and style casting with edwin-avatar-v4 LoRA model. Framework powers all generation commands, AI editor, and scene casting workflows.
How It Works
LPA is the prompt construction framework used before generation - not post-processing. Every image, edit, and refinement passes through these 7 layers to build the initial prompt that gets sent to the AI model. The framework prevents common AI artifacts by explicitly defining shadows, depth, materials, and realism at prompt time.
Explore the 7 Layers
Click to see detailed layer specifications and anti-artifact rules
Expand
Explore the 7 Layers
Click to see detailed layer specifications and anti-artifact rules
Identity Anchor (TOK Purity)
PURE identity trigger. Only TOK + pose + clothing + camera engagement. NEVER add body type, skin texture, or realism modifiers here - they dilute the LoRA signal.
TOK male, full body standing, face clearly visible looking toward camera with thoughtful expression, contemporary casual attire, both hands visible at sides, exactly two hands, anatomically correct
Spatial Positioning
Subject's scale, frame position, and compositional placement. Critical for architectural photography where subject is scale element.
, subject positioned as small scale element (1/12th of frame), standing in lower third of composition, dwarfed by towering architecture
Environment & Background
Complete material specification for every surface - walls, floors, ceilings. Add natural imperfections, weathering, texture variation by distance. Prevents vague AI backdrops.
, futuristic curved spiral staircase, polished white marble steps, brushed steel railings, floor-to-ceiling glass walls, fully designed architectural background with complete spatial definition, every surface material specified, natural surface imperfections and subtle weathering, foreground surfaces with visible fine texture detail, background textures naturally compressed by atmospheric perspective, NO miniature model texture appearance
Lighting & Shadow Logic
CRITICAL: Shadows ONLY from identifiable light sources. Prevents phantom shadows from non-existent lights. Add volumetric effects, atmospheric haze, proper shadow gradients.
, architectural interior lighting, even diffused quality, cool white (5000-6000K), shadows cast exclusively from identifiable light sources in scene, NO random shadows from non-existent lights, volumetric light rays visible through atmospheric haze, natural shadow gradients with proper falloff
Color Grading
Color palette, saturation levels, tonal treatment. Keeps color decisions separate from environmental material specs.
, desaturated architectural environment (very low almost monochromatic), white and gray primary tones, minimalist editorial color treatment
Camera Technical
Lens, aperture, depth of field gradient, perspective. CRITICAL: Specify "natural DoF gradient" and "NO tilt-shift miniature effect" to prevent model-like appearance.
, ultra-wide angle (14-24mm equivalent), worm's eye view, natural depth of field gradient from foreground to background, subject in primary focus plane with critical sharpness, progressive focus falloff toward distant background, NO tilt-shift miniature effect, correct relative scale between foreground and background elements
Quality & Anti-AI Tags
Photography style, anti-AI modifiers, photographic realism standards. Prevents synthetic AI smoothing and plastic appearance.
, professional architectural photography, high detail, luxury hospitality aesthetic, editorial quality, authentic photographic capture not AI-generated smoothness, natural film grain and sensor noise, no plastic artificial AI smoothing or synthetic polish
Core Principles
Layer Separation
Each layer handles ONE concern. Mixing concerns dilutes effectiveness.
Non-Interference
Realism tags in L7 only, NEVER in L1. Keeps TOK signal pure.
Explicit Logic
Shadows from sources, DoF gradients, surface materials - nothing vague.
LPA in the Generation Pipeline
1. Build Prompt
Apply 7 layers
2. Send to Model
Flux-dev LoRA
3. Result
90-100% likeness
LPA defines the prompt structure before generation, preventing artifacts like phantom shadows, tilt-shift effects, and AI smoothing at the source
- •Miniature/tilt-shift effects
- •Phantom shadow casting
- •Synthetic AI polish
- •Architectural photography
- •Portrait & lifestyle
- •Product photography
- •Initial image generation (all commands)
- •AI editor prompt construction
- •Enhancement & refinement passes
Add Targeted AI Capabilities
Not blanket automation - precision augmentation mapped to measurable constraints: cycle-time, quality, throughput, incident lag.
