Local AI image generation in 2026: Flux, SD & ComfyUI

Local AI image generation in 2026 means running models like Flux and Stable Diffusion on your own GPU through a tool such as ComfyUI — no API bill, no per-image fee, and no images leaving your machine. For a marketing team, the appeal is obvious: brand visuals at $0 per image after the hardware is paid for. The catch almost nobody spells out is licensing.

The cloud comparison is stark on paper. DALL·E (gpt-image-2) costs roughly $0.04 to $0.08 per image on the API; Midjourney runs $10 to $120 a month; Adobe Firefly sits in between. Local generation collapses that variable cost to zero. But “free to download” and “free to use for a client” are not the same thing — and getting that distinction wrong is a contract risk, not a rounding error.

This guide is the map agencies actually need: which models exist and what they cost, how much VRAM each one demands, which licenses clear them for paid client work, the tooling that has won by 2026, and an honest break-even on local hardware versus a subscription. Every figure below is sourced to a primary or benchmark reference, and the speculative ones are flagged as such.

Three things changed by 2026. Open-weight models like Flux closed most of the quality gap to closed cloud tools. Consumer GPUs got enough VRAM to run them. And the tooling — ComfyUI above all — matured from hobbyist novelty into something a production team can actually depend on. The result is that a marketing studio can now generate on-brand visuals locally that, two years ago, only a cloud subscription could deliver.

The cost argument is the headline, and it is real: after the hardware, every image is free. But the more durable argument is ownership. A cloud tool can change its pricing, its content policy, or its terms of service overnight; a model and a fine-tune sitting on your own disk cannot. For agencies handling client IP, the privacy angle compounds it — nothing is uploaded, nothing is retained on a third-party server, nothing is used to train someone else’s model. That is the same logic driving teams to run language models locally, which we covered in the case for on-device AI agents and the infrastructure walkthrough in our local LLM deployment guide.

Where this is heading is worth naming. As open models keep shrinking — FLUX.2 [klein]’s 4B variant runs license-clean on modest hardware — the local option stops being a power-user niche and becomes a default for any team with steady image volume. The studios that win the next two years will treat a trained brand model as infrastructure they own, not a feature they rent.

Two model families dominate local generation. Black Forest Labs ships Flux — a 12B-parameter rectified-flow transformer in its first generation, with the larger FLUX.2 line arriving in late 2025 and early 2026. Stability AI ships Stable Diffusion, now at version 3.5, alongside the still-ubiquitous SDXL. They differ in fidelity, VRAM appetite, and — critically — in how freely you can use them commercially.

Two more Flux models matter for context, even though most teams won’t run them daily. FLUX.2 [dev] is a 32B open-weight model released November 25, 2025 — the largest locally-runnable Flux, with multi-reference support for up to 10 images and high-resolution editing. Don’t confuse it with the 12B FLUX.1 [dev]; they are different models with different release dates and parameter counts. FLUX.2 [max] is the top-tier, cloud-API-only model — no local weights, available as of mid-2026, adding grounded generation and the highest editing consistency in the family. We cover the API tier in depth in our FLUX.2 [max] guide. There is also a purpose-built editing model, FLUX.1 Kontext [dev] (June 2025), for iterative edits with character preservation.

VRAM is the single hard constraint. The model has to fit in your GPU’s memory to run at full speed; quantization (FP8, or GGUF Q4) shrinks the footprint at some cost to quality. The table below maps each model to its minimum VRAM and a measured generation speed on an RTX 4090 — the de facto reference card for local Flux work.

The practical brackets: an 8GB card runs SDXL and quantized Flux; a 12GB card is the comfortable floor for FP8 Flux and SD 3.5 Medium; a 16GB card handles most workflows; and 24GB (the RTX 4090 / 3090 tier) runs everything short of full-precision 32B FLUX.2 [dev] cleanly. One caveat worth repeating: GPU street pricing in 2026 has stayed volatile, partly because of a global DRAM and GDDR7 shortage — the same squeeze that pulled Apple’s high-memory Mac Studio configs and pushed workstation-GPU prices up. Verify current pricing before you buy. For the broader picture of what a well-equipped local AI studio looks like, see our guide to local AI hardware.

This is the section that matters most for client work, and the one most guides skip. A model being free to download tells you nothing about whether you can legally use its output in a paid deliverable. The table below combines license and commercial eligibility in one place — verify the current terms on each model’s own page before you ship, because licenses do change.

If you do want the higher-fidelity dev-tier models for client work, Black Forest Labs sells four commercial tiers. Builder covers 10,000 images a month on a single domain with fine-tuning and LoRA rights; Platform raises that to 100,000 a month and adds FLUX.2 [klein] 9B and dev; Professional keeps 100,000 a month but extends to three domains, aimed squarely at creative agencies producing images for named clients; Enterprise is custom volume across all models. Choosing a model is half the decision — matching the license tier to how you bill clients is the other half, and it is exactly the kind of thing we sort out inside a content-engine engagement.

ComfyUI is a node-based interface where every step of the pipeline — load checkpoint, encode the prompt, sample, decode, save — is a visible, wirable node. That sounds fiddly, and the learning curve is real, but it buys two things that matter for production: the best day-one support for new models like Flux, and meaningfully faster execution on complex workflows. In community benchmarks it generates SDXL images ~25% faster than Automatic1111 and handles ControlNet plus upscaling workflows ~60% faster.

A LoRA (Low-Rank Adaptation) is a small add-on you train on top of a base model to teach it a specific style, product, or look. For an agency this is the whole game: train a LoRA on a client’s brand — their product photography, their palette, their art direction — and every generated image inherits that style consistently. The output is not a subscription feature; it is a file you keep.

Dataset size sets the quality ceiling: products want 10–20 images from multiple angles; an art style wants 20–50 representative examples. Once trained, the LoRA loads in seconds and runs offline forever. That permanence is the strategic point. A brand LoRA you own cannot be deprecated, re-priced, or have its terms rewritten — it is closer to a typeface license you hold outright than to a SaaS seat you rent.

Here is where most local-vs-cloud articles oversell. Yes, local generation costs $0 per image — but the hardware is real money up front, and whether it pays off depends entirely on volume. The table below is the recurring monthly cost at three realistic agency volumes; the break-even on the hardware follows.

Two things shift that math in local’s favor in the real world. First, capex is the lever: many studios already own a capable gaming GPU, which makes the hardware a sunk cost and collapses the break-even to near-immediate. Second, the per-image cloud numbers assume you generate exactly what you need — but creative work is iterative, and at $0.04 a try, exploring 60 variations of a concept costs $2.40 every time, where locally it costs nothing. For the full cloud-side pricing picture, see our AI image API pricing comparison, and for the broader local-versus-subscription ROI argument, our local-vs-cloud ROI analysis.

The pieces assemble into a clear default stack for 2026. Pick a license-clean base — FLUX.1 [schnell] for speed, SDXL for the most mature ecosystem, or a BFL commercial tier if you need dev-grade fidelity for clients. Run it in ComfyUI on a 16–24GB card. Train a brand LoRA per client and keep the file. That combination delivers on-brand images at zero marginal cost, fully offline, with a clean commercial license — and the brand model is an asset on your books, not a line item you rent.

The forward signal is iteration speed. If FLUX.2 [klein]’s sub-second generation figures hold up under independent testing — and that is still a vendor claim, not a verified one — the creative loop changes shape. Evaluating dozens of concepts a minute, rather than waiting 45–60 seconds per Midjourney render, turns image generation from a request-and-wait task into something closer to live sketching. The studios that internalize that shift will move from “generate an image” to “explore a space of images,” and the output quality follows the number of iterations you can afford — which, locally, is unlimited.

None of this removes the human. Brand judgment, art direction, and knowing which of 60 variations is actually on-strategy is the work that matters; the local stack just removes the meter from the iteration. That division of labor — senior judgment plus owned, agent-assisted capacity — is exactly how we structure visual production inside an AI transformation engagement.