Skip to content
[ STACK BUILDER · CONFIGURATOR ]

Build your local AI stack — matched to your card.

Tell us your hardware budget and what you want to make. We pick a consistent, working stack from our catalogue — workflow engine, model, runner, training tools. Empty slots tell you why nothing fits, instead of pretending.

CONFIGURATOR //LIVE
01 / What do you want to do?The job your stack needs to handle.
02 / What hardware do you have?We filter out anything that won't fit your card's minimum VRAM.
03 / PlatformApple Silicon only shows tools with native Metal / MPS support.
[ YOUR STACK ]4 / 4 SLOTS FILLED

Train LoRAs / fine-tune on 24 GB VRAM

Train your own LoRAs or fine-tune base models. The most VRAM-hungry use case. Cards in this tier: RTX 3090, RTX 4090, M-series 32 GB+.

SEE ALL TOOLS FOR 24 GB VRAM →
TRAINING TOOLKIT🏋️ Trainer

Diffusers

Hugging Face's go-to library for every diffusion model.

Role: LoRA / fine-tune trainers.

OPEN SOURCE4–12 GB VRAM

ALT // Kohya_ss — the standard sdxl/flux lora training ui.

FULL REVIEW →
WORKFLOW ENGINE🕸️ Core

ComfyUI

The nodal workflow engine for serious diffusion.

Role: The graph / UI that actually runs your pipelines. Where you build the model chain.

OPEN SOURCE6–16 GB VRAM

ALT // ComfyUI-Manager — install, update, and govern comfyui custom nodes.

FULL REVIEW →
IMAGE GENERATION🖼️ Model

FLUX.1 [dev]

12B parameter open-weight diffusion model.

Role: The image model itself — SDXL, Flux, SD3, etc.

OPEN SOURCE8–24 GB VRAM

ALT // ComfyUI IPAdapter Plus — reference-image conditioning for comfyui.

FULL REVIEW →

HOW WE PICK //

Every pick is a function of three things in our catalogue: minimum VRAM (must fit your budget), power-user score (60% of the weight), and trending score (20%). We add small bonuses for open-source licensing and beginner-friendly setup. No paid placements, no “sponsored” tier — if it's not in our catalogue, it can't appear here.

READ THE FULL METHODOLOGY →