mygens-ai 0.1.0

Track, search, and version every AI image and video prompt. Semantic search, prompt lineage, cross-platform.

MyGens
The prompt library for AI image and video creators

Install · Quick Start · How It Works · Platforms · Architecture

---

You create AI images and videos across Midjourney, DALL-E, Stable Diffusion, ComfyUI, Runway, Sora, Kling, and more. Your prompts are everywhere — Discord, ChatGPT threads, browser tabs, random folders. When you finally nail a perfect result, you can't find the exact prompt a week later.

MyGens gives you three things no other tool does:

1. Semantic search — search "moody dark vibes" and find your prompt about "atmospheric shadows, film noir, low-key lighting." Works by meaning, not just keywords.

2. Prompt lineage — see how any idea evolved from first draft to final version, across platforms, as an interactive tree.

3. Word-level diff — compare any two prompt versions and see exactly what changed, which parameters moved, and which version scored better.

Everything runs locally. Your prompts never leave your machine.

pip install mygens

How MyGens compares

%%{init: {'theme': 'dark', 'themeVariables': { 'fontSize': '14px'}}}%%
graph LR
    subgraph TODAY["How creators manage prompts today"]
        D[Discord messages] -.-> L[Lost]
        N[Notion docs] -.-> L
        F[File names] -.-> L
        C[ChatGPT threads] -.-> L
        B[Browser tabs] -.-> L
    end

    subgraph PT["With MyGens"]
        S[Scan / Sync / Save] --> DB[(Local DB)]
        DB --> SE[Semantic Search]
        DB --> LN[Lineage Tree]
        DB --> DF[Prompt Diff]
        DB --> DA[Dashboard]
    end

    style L fill:#8b3a3a,color:#fff
    style DB fill:#8b6914,color:#fff

	MyGens	DiffusionToolkit	Eagle	Notion
Cross-platform (MJ + SD + DALL-E + ComfyUI + Sora)	Yes	SD only	Not AI-aware	Manual
Semantic search	Yes	—	—	—
Prompt lineage tracking	Yes	—	—	—
Word-level prompt diff	Yes	—	—	—
Video generation support	Yes	—	—	—
API sync (Replicate, fal.ai)	Yes	—	—	—
Local-first, 100% private	Yes	Yes	Yes	No
Cross-OS (macOS, Linux, Windows)	Yes	Windows only	Yes	Web
Open source	Apache 2.0	MIT	$30	—

Install

pip install mygens

No Docker. No Node.js. No native compilation. Python 3.11+ on any OS.

Quick start

mygens init                    # Create your library

# --- Import from local tools (automatic) ---
mygens scan ~/ComfyUI/output   # Reads workflow JSON from every PNG
mygens scan ~/sd/outputs       # Reads A1111 metadata from every PNG

# --- Import from cloud platforms (automatic) ---
mygens sync replicate          # Imports Flux, Luma, CogVideoX, etc.
mygens sync fal                # Imports from fal.ai endpoints

# --- Save manually (5 seconds) ---
mygens log "cyberpunk city at night" -p sora -m sora-v1

# --- Search by meaning ---
mygens search "moody lighting" # Finds "dark atmospheric shadows, film noir"

# --- See creative evolution ---
mygens tree <id>               # Shows how an idea branched and evolved
mygens diff <id1> <id2>        # Word-level diff between two versions

# --- Web dashboard ---
mygens serve                   # Opens http://localhost:9753

How it works

Capture — get prompts into MyGens

%%{init: {'theme': 'dark'}}%%
flowchart LR
    subgraph LOCAL["Local Generation"]
        SD[Stable Diffusion] --> PNG1[PNG with metadata]
        CU[ComfyUI] --> PNG2[PNG with workflow JSON]
        MJ[Midjourney] --> PNG3[PNG with EXIF]
    end

    subgraph CLOUD["Cloud Platforms"]
        REP[Replicate<br/>Flux · Luma · CogVideoX] --> API1[API sync]
        FAL[fal.ai<br/>Kling · Minimax · Flux] --> API2[API sync]
    end

    subgraph MANUAL["Web-based Tools"]
        SORA[Sora] --> WEB[+ Save prompt]
        RW[Runway] --> WEB
        DALLE[DALL-E] --> WEB
        OTHER[Any tool] --> WEB
    end

    PNG1 --> SCAN["mygens scan"]
    PNG2 --> SCAN
    PNG3 --> SCAN
    API1 --> SYNC["mygens sync"]
    API2 --> SYNC
    WEB --> DASH[Dashboard upload]

    SCAN --> DB[(MyGens DB)]
    SYNC --> DB
    DASH --> DB

    style DB fill:#8b6914,color:#fff

Search — find anything by meaning

%%{init: {'theme': 'dark'}}%%
flowchart LR
    Q["Search: 'gloomy dark vibes'"] --> K[Keyword Search<br/>SQLite FTS5]
    Q --> S[Semantic Search<br/>BGE-small embeddings]

    K --> |"exact matches"| M[Merge + Deduplicate]
    S --> |"meaning matches"| M

    M --> R1["'dark moody atmospheric<br/>shadows, film noir'"]
    M --> R2["'cyberpunk city at night,<br/>neon rain, wet streets'"]
    M --> R3["'cinematic lighting,<br/>volumetric rays'"]

    style Q fill:#8b6914,color:#fff
    style R1 fill:#3a5a3a,color:#fff

How it works under the hood: A 33MB local model (BGE-small, ONNX Runtime) embeds every prompt as a 384-dimensional vector. At search time, your query is embedded and compared against all stored vectors using cosine similarity. Takes <10ms for 10,000 prompts. No GPU. No API calls. No cloud.

Lineage — see how ideas evolve

%%{init: {'theme': 'dark'}}%%
graph TD
    V1["v1: 'cyberpunk city'<br/><small>Midjourney · v7</small>"]

    V1 --> V2["v2: '...neon reflections<br/>on wet streets'<br/><small>Midjourney · variation</small>"]
    V1 --> V8["v8: '...aerial drone shot'<br/><small>Midjourney · variation</small>"]

    V2 --> V3["⭐ v3: '...cinematic,<br/>anamorphic lens'<br/><small>Midjourney · hero shot · ★★★★★</small>"]
    V2 --> V7["v7: '...rain, puddles,<br/>35mm film grain'<br/><small>Midjourney · variation</small>"]

    V3 --> V4["v4: translated to<br/>Stable Diffusion<br/><small>dreamshaperXL · translate</small>"]
    V3 --> V6["v6: translated to<br/>DALL-E for client<br/><small>gpt-4o · translate</small>"]

    V4 --> V5["v5: translated to<br/>ComfyUI + ControlNet<br/><small>dreamshaperXL · translate</small>"]

    style V3 fill:#8b6914,color:#fff,stroke:#c4a44e,stroke-width:2px
    style V1 fill:#3a4a6a,color:#fff
    style V8 fill:#3a4a6a,color:#fff
    style V4 fill:#6a4a3a,color:#fff
    style V5 fill:#6a4a3a,color:#fff
    style V6 fill:#3a5a3a,color:#fff

One idea. Four platforms. Eight iterations. Two branches. The lineage tree shows the complete creative evolution — which path led to the 5-star hero shot, where you explored a different direction, and how the prompt changed when translated across tools.

Diff — see exactly what changed

 Generation v1 → v3

 PROMPT:
   cyberpunk city [-at dusk-] {+at night+}, neon reflections on wet
   [-streets-] {+streets, cinematic, anamorphic lens, rain+}

 PARAMETERS:
   seed:     1001 → 3003
   stylize:  500 → 750  [+50%]
   ar:       1:1 → 21:9 [CHANGED]

 RATING: 0 → 5  [+5 ★]

Platforms

Platform	How MyGens captures it	Setup
Stable Diffusion (A1111, Forge)	Reads prompt, seed, model, sampler, CFG from PNG tEXt chunks	mygens scan <folder>
ComfyUI	Walks the workflow node graph — extracts KSampler, CLIPTextEncode, CheckpointLoader	mygens scan <folder>
Midjourney	Reads EXIF Description, parses --ar, --stylize, --v params	mygens scan <folder>
Replicate (Flux, Luma Ray, CogVideoX, LTX-Video, Minimax)	Syncs predictions via API, stores output URLs (no download)	mygens sync replicate
fal.ai (Flux, Kling, Minimax, Luma, SVD)	Syncs request history via API, stores output URLs	mygens sync fal
Sora / Runway / Kling / Pika / Veo	Paste prompt + drop video in dashboard	+ Save prompt button
DALL-E / ChatGPT	Paste prompt + drop image in dashboard	+ Save prompt button
Any tool	Webhook or CLI	POST /api/v1/webhooks/generic

Dashboard

mygens serve   # → http://localhost:9753

Gallery view with prompts grouped by platform (list or grid). Click any prompt → detail panel slides in. View lineage. Compare versions. Rate, tag, search. Dark and light themes. Manual entry with drag-and-drop for images and videos.

Architecture

%%{init: {'theme': 'dark'}}%%
graph TB
    subgraph UI["User Interfaces"]
        CLI[CLI<br/><small>13 commands · Typer + Rich</small>]
        WEB[Web Dashboard<br/><small>React · Vite · TailwindCSS</small>]
    end

    subgraph API["REST API"]
        FA[FastAPI<br/><small>16+ endpoints · Pydantic v2</small>]
    end

    subgraph CORE["Core Engine"]
        GEN[Generation CRUD]
        SEARCH[Hybrid Search<br/><small>FTS5 + BGE-small embeddings</small>]
        LIN[Lineage DAG<br/><small>Recursive CTEs</small>]
        DIFF[Prompt Diff<br/><small>Word-level</small>]
        FRAG[Fragment Analysis]
    end

    subgraph CAPTURE["Capture Layer"]
        PARSE[Parsers<br/><small>A1111 · ComfyUI · Midjourney</small>]
        INT[Integrations<br/><small>Replicate · fal.ai · Webhooks</small>]
    end

    subgraph STORAGE["Storage"]
        DB[(SQLite + FTS5<br/><small>+ embedding vectors</small>)]
    end

    CLI --> GEN & SEARCH & LIN & DIFF
    WEB --> FA --> GEN & SEARCH & LIN & DIFF
    GEN & SEARCH & LIN & DIFF & FRAG --> DB
    PARSE --> GEN
    INT --> GEN

    style DB fill:#8b6914,color:#fff

Deep dive: ARCHITECTURE.md — data model, search internals, parser system, every design decision with tradeoffs.

Tech stack

	Choice	Why
Language	Python 3.11+	Zero-friction install. AI creators already have it.
Database	SQLite + FTS5	Built-in. Zero config. Single portable file.
Search	BGE-small (ONNX Runtime)	33MB. CPU-only. <10ms. No PyTorch (2GB).
CLI	Typer + Rich	Beautiful output. Typed.
API	FastAPI	Async. Auto-generated OpenAPI docs.
Dashboard	React + Vite + TailwindCSS	Bundled in the Python package.
DAG viz	React Flow + dagre	Interactive lineage graphs.

CLI reference

mygens init                    Initialize library
mygens scan <path>             Import images (auto-detects platform)
mygens sync <platform>         Import from Replicate or fal.ai
mygens log <prompt>            Log a prompt manually
mygens search <query>          Semantic + keyword search
mygens list [--json]           List with filters
mygens diff <id1> <id2>        Word-level prompt diff
mygens tree <id>               Lineage tree visualization
mygens link <child> <parent>   Connect prompts in lineage
mygens rate <id> <1-5>         Rate a prompt
mygens tag <id> -t "a,b"       Add tags
mygens stats                   Library statistics
mygens serve                   Start web dashboard

REST API

Available at localhost:9753 when running mygens serve:

POST   /api/v1/generations/upload     Create with image/video upload
GET    /api/v1/generations/           List with filters
GET    /api/v1/generations/{id}       Single prompt + outputs
GET    /api/v1/generations/{id}/tree  Lineage DAG
GET    /api/v1/generations/{id}/diff/{other}  Prompt diff
GET    /api/v1/search?q=...          Hybrid search
POST   /api/v1/webhooks/replicate    Replicate webhook
POST   /api/v1/webhooks/generic      Generic webhook

Full OpenAPI docs: http://localhost:9753/docs

Extend it

Add a parser for a new platform in ~30 lines:

from mygens.parsers.base import Parser, ParsedGeneration

class MyParser:
    name = "My Platform"
    platforms = ["my_platform"]

    def detect(self, buffer: bytes) -> bool: ...
    def parse(self, buffer: bytes, file_path: str) -> list[ParsedGeneration]: ...

See a1111.py for a complete example.

Privacy

• 100% local. Prompts, images, videos, embeddings — everything stays on your machine.
• Zero telemetry. No analytics. No tracking. No phone-home.
• Portable. One folder (~/.mygens/), one SQLite file. Copy it anywhere.

Development

git clone https://github.com/chopratejas/mygens && cd mygens
uv venv && uv pip install -e ".[dev]"
pytest tests/                       # 138 tests, ~15s, zero mocks

License

Apache 2.0

Contributing

The highest-impact contributions right now:

1. Parsers — Flux, Pika, NanoBanana Pro, any tool that embeds metadata
2. Browser extension — auto-capture from Sora, ChatGPT, Runway, Midjourney web
3. API integrations — RunwayML, Luma, any platform with generation history
4. Dashboard — the UI is functional but early

See ARCHITECTURE.md for the full codebase guide.