Local AI Images Are Free. Cloud AI Images Are Instant. Your GPU Decides.

Open image models have caught up — and in some categories, pulled ahead. HiDream-O1, an 8B model released under the MIT license in May 2026, ranked among the top open-weight entries on the Artificial Analysis text-to-image arena at a fraction of the size of larger rivals. Alibaba's Qwen-Image renders readable text inside images better than most cloud tools. And the editing models — Qwen-Image-Edit, FLUX Kontext — now change objects, backgrounds, and text inside existing photos from a plain-language instruction, locally, for free.

The cloud side has its own 2026 story: the market consolidated around a few serious players, entry prices dropped to the $8-10/month range, and commercially-safe training data became a real differentiator for business users. Both doors are genuinely good. The question is which one fits you — and unlike video, the hardware bar for images is low enough that the local door is realistic for far more people.

All three run through ComfyUI (or a similar local interface) on your own machine. As with video generation: these are diffusion models, not LLMs — they don't run in Ollama.

The FLUX family is the default for serious local image work. FLUX.2 [dev] (32B) leads on photorealism and high resolution, combining up to 10 reference images while keeping character, product, and style consistent. FLUX.1 [schnell] generates quality images in 1–4 steps on just 8GB of VRAM. FLUX.1 Kontext is the local leader for editing existing images.

License — read this part carefully: the family is split. FLUX.1 [schnell] is Apache 2.0 — unrestricted, commercial use included. FLUX.1/2 [dev] and Kontext use Black Forest Labs' non-commercial license — running them in a commercial product requires a paid license from BFL. "Open weights" does not mean "commercially OK" here.

Hardware: 8GB (schnell), 12–16GB (dev/Kontext), 24GB (FLUX.2 dev, GGUF Q4).

SD 3.5 (8B Large / 2.5B Medium) is no longer the quality leader, but it has something the others don't: the deepest ecosystem in local AI. Years of community LoRAs (small add-on files that teach the model a style, a character, or a product look), ControlNets, and tutorials mean that whatever you want to make, someone has already built the parts.

Hardware: 8–12GB depending on variant; SDXL runs happily on 8GB.

License: Stability Community License — free for commercial use if your annual revenue is under $1M; above that you need an Enterprise License. Fine for freelancers and small businesses; a real constraint at scale.

Alibaba open-sourced Qwen-Image (20B) in August 2025 under Apache 2.0 — no revenue thresholds, no non-commercial clauses, no territory games. Its specialty is something most models still fail at: readable, correctly-spelled text inside the image, in multiple languages. Posters, signs, infographics, thumbnails with headlines — this is the model.

Bonus: Qwen-Image-Edit performs precise, prompt-based edits on existing photos — change an object's color, swap a background, fix text — while preserving everything else.

Hardware: 8GB (GGUF quantized) to 24GB (full precision). License: Apache 2.0 — the only top-tier image model with zero fine print.

Released May 2026 under the MIT license — even more permissive than Apache 2.0 — HiDream-O1 (8B) ranked among the top open-weight entries on the Artificial Analysis text-to-image arena shortly after release, competing with models several times its size. It's young, the ecosystem is thin, and long-term support is unproven (this ranking is single-source as of writing — verify before treating it as settled). But if the trajectory holds, this list gets rewritten within a year.

If our video article scared you off local AI, this is the good-news table: images are simply a much lighter workload. A used RTX 3060 12GB (~$170–220) covers the first two rows below. Compare that to video generation, where 24GB is the comfortable tier.

Not sure where your hardware lands? These guides do the math: VRAM Calculator for exact requirements per model, How Much VRAM Do You Need? for charts across model sizes, Best GPUs for Local AI and Best Budget GPUs for hardware picks, and GPU vs CPU vs Apple Silicon for platform comparisons.

Same honesty as the video article. Local image generation means:

The setup. ComfyUI or a similar interface, model files in the right folders, the occasional dependency error. An evening, not a week — image setups are far simpler than video — but still your evening.

The prompting. No built-in prompt helper, no style presets, no content filter (full control — and full responsibility). You write the prompts yourself. Our guides on system prompts vs. user prompts and prompt engineering for local models cover the fundamentals that transfer directly.

The finishing. Upscaling, face fixes, batch organization — separate tools and nodes you choose yourself. Want a consistent character across 30 images? That's LoRA training: doable, documented, but a project.

Before picking a door, know what you're walking through it for. The realistic use-case map:

• Content sites and blogs: hero images, article illustrations, social preview cards.
• YouTube and social: thumbnails, channel art, post graphics, ad creatives — including fast A/B variants.
• E-commerce and marketing: product mockups, lifestyle scenes, seasonal variants of the same shot.
• Work materials: presentation visuals, pitch-deck graphics, concept mockups.
• Creative projects: book covers, concept art, mood boards, print-on-demand designs.
• Editing, not just creating: with Qwen-Image-Edit or FLUX Kontext — swap backgrounds, remove objects, restyle product photos, fix text in graphics.

Two honest limits: AI images still struggle with exact brand consistency across large batches (local LoRAs help; cloud tools are catching up), and anything requiring real people, real products, or factual accuracy needs photography, not generation.

We picked Adobe Firefly and getimg.ai because they genuinely cover the two most common cloud needs: maximum commercial safety, and the easiest bridge from local to cloud. Midjourney and ChatGPT are also widely used for image generation, but neither fits either of those two specific needs as directly — Firefly and getimg.ai are the more useful picks for this comparison, not a default.

Firefly is trained on Adobe Stock and openly licensed content — meaning Adobe designed it so business users don't inherit copyright risk — and it integrates directly with Photoshop and the rest of Creative Cloud. If client work or brand safety is your concern, this is the cloud door. A free trial exists to test it before paying; paid plans start at $9.99/month for 2,000 generative credits (Standard tier). Best for: professionals, agencies, anyone whose clients ask "is this legally safe?"

Here's the twist most comparisons miss: getimg.ai runs the same open models you'd install locally — FLUX and friends, 20+ models in one interface — on their GPUs instead of yours. No setup, no VRAM math, commercial rights included on every plan. If the local door appeals to you but your hardware says no, this is the bridge. Pricing is paid-only since early 2026 (the free tier was retired) — Entry from $8/month billed annually ($10/month billed monthly) for 3,000 credits; higher tiers scale up from there. Best for: local-curious users without the GPU, and anyone who wants open-model variety without the ComfyUI learning curve.

(Honorable mention: Ideogram — a cloud leader for text-in-image, with a limited free tier that publishes images to a public gallery, and paid plans starting around $20/month.)

The short version, mapped to common situations:

Go local if you have (or will buy) an 8GB+ GPU, generate images regularly, and want privacy, zero marginal cost, and full creative control. Qwen-Image is the safest foundation — Apache 2.0, best-in-class text rendering — with FLUX for photorealism (mind the license split by variant) and SD 3.5 for its unmatched style ecosystem.

Go cloud if you want results in the next five minutes, generate occasionally, or have no GPU. Adobe Firefly is the safe, professional pick with commercially-safe training data; getimg.ai is the bridge for anyone who likes the idea of open models but not the idea of installing them.

And if video is next on your list — that's a different hardware conversation. Read the companion piece: Local AI Video Generation vs. Cloud.