Sign In
homemodelsimagesvideos3D Modelsarticlescomicschallengesupdatesshop

Ranbell Image - A Local Studio I Built for Myself

0

OimoJet's Avatar

OimoJet

Jun 28, 2026

(Updated: 19 days ago)

announcement
comfyuiollamaranbell imageqdrant
Ranbell Image - A Local Studio I Built for Myself

Hey, I'm OimoJet. First article here. I made a thing and I want to show you.

It's called Ranbell Image. Fully local. Runs on your own GPU. No API keys, no cloud.

Current Supported Langs: English and Japanese

Why I Built It

Honestly? The gallery came first, not the generation.

I wanted to do image blending - pick two of my own images, mix them, generate something new. But to do that, I needed to find the right two images. And finding anything in a folder of thousands... yeah, no.

So I built the gallery I needed. Then the generation pieces fell into place around it.

Refine - Blend Your Own Images

Pin 1–6 of your images. Set weights (70% / 30%, whatever).

The system pulls visual vocabulary from each reference, weighted, resolves conflicts, writes a clean prompt, and sends it to ComfyUI. You watch it happen in real time.

This is the feature I originally wanted. Everything else exists to make this one work properly.

🖼️[IMAGE: Refine/Blend (x5 speed)]

Invoke - Five Takes on One Idea

Give it a phrase or a mood. Five spirits each generate a prompt and an image, in parallel:

  • Mirror / Faithful: Stays on your axis

  • Counter / Rebel: Flips one thing for contrast

  • Wander / Stranger: Sneaks in an unexpected element

  • Surge/ Lunatic: Embraces the impossible

  • Vantage / Oracle: Whatever's memorable

Pick the one you like (Adopt), shake them again (Respin), or send it deeper (To Refine).

This is the part I have the most fun with. Five different paths from the same seed of an idea.

🖼️[IMAGE: The five Invoke spirits side by side (x10 speed)]

Finding Images by Feeling

Type whatever comes to mind:

  • "melancholy girl in afternoon light"

  • "海" (works in any language — Japanese types in, English-tagged images come back)

  • A hex color, and you get perceptual color matching

You can stack semantic + tag + color + mood in one query. It's stupid fast once embeddings are built.

02_search_tags.png

🖼️[IMAGE 2: Semantic search results screen]

Nine Ways to Get Inspired

When you don't know what to make, ask your collection.

Serendipity. Morph between two images. Find the opposite of something. Mix three vibes by percentage. Hunt for outliers. Discover surprising tag combinations.

Nine modes, all asking your library a different question.

04_inspire.png

🖼️ [IMAGE: Inspire panel — morph or blend visualization]

See Your Own Aesthetic

A 2D map of your whole collection (UMAP). Your dominant colors in 3D space. A network graph of your tags.

You can literally see the shape of what you've been making. Your phases. Your blind spots. It's weirdly addictive.

06_02_analyzer.png

06_03_analyzer.png

🖼️[IMAGE: UMAP semantic map with clusters]

A Note on How This Was Made

Ranbell Image was built with Claude (Anthropic) as my coding partner. I made every design decision. Claude wrote the code.

That's how building looks now, and I don't want to pretend otherwise. It doesn't make it less mine. It made it possible.

Try It

v0.3.1 — works, but young.

🔗 github.com/ranbell/ranbell_image

Requirements: Docker Compose, Comfyui, Ollama, Linux (WSL2), NVIDIA GPU

Break it, tell me what's broken. Use it, tell me what to add. I'm building this for myself first, but I'd love to build it with you.

See you in the comments 🍊

OimoJet

0