State-of-the-art open image model from Black Forest Labs.
Image Generation
Flux is the image model that photorealism people quietly switched to in 2024 and never went back from. Built by Black Forest Labs — a team founded by the same researchers who trained the original Stable Diffusion at Stability AI before leaving — Flux launched in August 2024 with three weight variants and a very specific pitch: open-weight, benchmark-beating, and unapologetically photographic. Two years later, in mid-2026, Flux is the reference model most working pros compare against when they say the words "AI photorealism."
What sets Flux apart is not marketing polish — Black Forest Labs barely does marketing — but the pixel-level quality of the output. Skin looks like skin. Hands have the right number of fingers most of the time. Product photography renders with plausible material response. Cinematic stills carry believable lighting and depth-of-field cues instead of the plastic sheen that gives most AI images away in a Twitter feed. On the standard prompt-fidelity and aesthetic benchmarks the field cares about, Flux Pro 1.1 (the current flagship variant) trades top position with Midjourney v7 depending on the category, and beats it outright for anything realism-adjacent.
Flux the product is unusual. There is a hosted API at api.bfl.ai for developers who want a first-party endpoint. There are open weights for Flux [dev] and Flux [schnell] anyone can download and run locally, plus a commercial Flux Pro tier available only via API and partner platforms. That combination — a top-tier hosted API plus real open weights — does not exist anywhere else in the market. Stable Diffusion has weights but the top-line quality gap is real. Midjourney has quality but no weights and no self-serve API. DALL-E 3 has neither weights nor a competitive photorealism ceiling. Flux is the pragmatic answer for anyone building an image pipeline in 2026 who wants photographic output.
The current lineup — Flux Pro 1.1, Flux Pro 1.1 Ultra, Flux [dev], and Flux [schnell] — covers hosted flagship quality, high-resolution 4-megapixel output, permissive-license open weights, and speed-optimized generation, respectively.
Photorealism ceiling. Flux Pro 1.1 renders skin, fabric, glass, metal, and food with a fidelity that trails only high-end commercial photography, and beats every other general-purpose AI model on realistic subject matter. This is the single feature that got Flux adopted by product photographers, e-commerce teams, and stock content producers within its first six months.
Prompt fidelity that respects nouns. Flux follows literal instructions more closely than Midjourney and roughly matches DALL-E 3 on object counts, spatial relationships, and compositional details. Ask for "three brown ceramic mugs on a walnut table with soft window light from the left" and Flux delivers the count, the material, and the light direction reliably.
Text-in-image that works. Flux renders short headlines, product labels, and signage cleanly enough that you can use it for mockups without post-processing. Ideogram is still the best-in-class model for text-heavy work, but Flux closes the gap for anything under twenty words.
Flux Pro 1.1 Ultra for 4-megapixel output. The Ultra endpoint produces images up to 2048x2048 (and higher on custom aspect ratios) with sharpness and detail retained across the frame. Useful for print work, campaign hero images, and anything destined for a physical medium where 1024x1024 is not enough.
Open weights for [dev] and [schnell]. Flux [dev] is a 12-billion-parameter guidance-distilled model released under a non-commercial license, runnable locally on a 24GB GPU or via ComfyUI. Flux [schnell] is Apache 2.0, four-step-optimized, and roughly ten times faster than [dev] with a modest quality trade. Between them you have a stack you can host, fine-tune, or embed into your own product without asking Black Forest Labs for permission.
Hosted API at api.bfl.ai. First-party API access to Flux Pro, Pro 1.1, and Pro 1.1 Ultra with pay-as-you-go pricing, plus the same models available on Replicate, Fal, Together, and Runware for teams that prefer a single-vendor image-model marketplace.
LoRA and fine-tune support. The open [dev] weights accept LoRA training the same way Stable Diffusion does — train a small adapter on ten to thirty product photos or brand references, and Flux will reliably produce output matching that visual identity. Every serious AI-image workflow in 2026 that needs brand consistency ends up training a Flux LoRA.
No consumer app. This matters. Flux has no Discord bot, no web app for casual users, no monthly subscription. It is a model, not a product. You use it through Replicate, Fal, ComfyUI, or your own code. Non-technical creators will find this hostile; technical builders find it liberating.
Flux pricing is per-generation via the hosted API. There is no monthly subscription and no free tier from Black Forest Labs directly, though most partner platforms include some free credits.
| Model | Price per image | Resolution | Notes |
|---|---|---|---|
| Flux [schnell] | ~$0.003 | 1024x1024 | Apache 2.0, self-hostable free |
| Flux [dev] | ~$0.025 | 1024x1024 | Non-commercial weights, self-hostable |
| Flux Pro | ~$0.05 | 1024x1024 | Hosted only |
| Flux Pro 1.1 | ~$0.04 | 1024x1024 | Current flagship, hosted only |
| Flux Pro 1.1 Ultra | ~$0.06 | Up to 4MP | Hosted only, print-ready |
Pricing on partner platforms (Replicate, Fal, Together) is within 10% of these numbers in either direction. Bulk API contracts are available directly from Black Forest Labs for high-volume customers. There is no monthly plan — you pay for what you generate.
For solo creators, this means Flux via a Replicate account costs roughly $10-30/month for a typical iteration workload — cheaper than a Midjourney Standard subscription if you generate under ~500 images per month, more expensive if you generate thousands. For product teams, the API model is the correct choice: you pay for generations that ship to real users, not for a seat that sits idle.
Self-hosting Flux [dev] on a rented GPU (24GB VRAM, ~$0.50-1.00/hour on Runpod or Vast) breaks even against the hosted API at roughly 20-40 images per hour depending on GPU tier — worth it for continuous workloads, not worth it for spiky ones.
Pros
Cons
Midjourney. Higher aesthetic ceiling for illustrative, editorial, and stylized work. Better --sref for brand-consistent asset runs without training a LoRA. No API, no open weights, no self-hosting. $10-120/month subscription. If your output is not photographic and you are not a developer, Midjourney is probably the better tool.
Stable Diffusion. The other open-weight option. Broader ecosystem — ControlNet, IP-Adapter, thousands of community LoRAs and checkpoints. Lower base quality than Flux [dev] but richer control surface for advanced ComfyUI workflows. Free to self-host; API from ~$0.002/image via Replicate.
DALL-E 3. Best-in-class prompt fidelity, embedded in ChatGPT, safest for casual users. Photorealism ceiling is meaningfully lower than Flux. Included in ChatGPT Plus at $20/month. The right choice if you are already in a ChatGPT thread and the sentence you wrote is the picture you need.
For a deeper comparison across the three big image models, see our full write-up: Midjourney vs DALL-E 3 vs Stable Diffusion.
black-forest-labs/flux-1.1-pro model page takes a prompt, an aspect ratio, and returns a 1024x1024 image in about 5-10 seconds. Try a realistic scene first — Flux rewards photographic prompting.guidance_scale) around 3-4 for realism, higher for stylized. Seed to reproduce a specific output. Everything else is optional until you know why you need it.Is Flux better than Midjourney? For photorealistic subjects — yes, most of the time in 2026. For stylized, illustrative, or editorial output — no, Midjourney's aesthetic ceiling is higher. Working pros run both.
Can I use Flux commercially? Yes, via the hosted API (Pro tiers) or via Flux [schnell] (Apache 2.0 weights). Flux [dev] weights are non-commercial — you can develop with them locally but not ship them in a product. Output generated via the paid API is commercially licensed to you.
Does Flux have a consumer app? No. Access is via API (api.bfl.ai, Replicate, Fal, Together, Runware) or self-hosted weights. Non-technical creators typically use Flux through a partner platform's UI rather than directly.
Can I fine-tune Flux? Yes, Flux [dev] accepts LoRA training the same way Stable Diffusion does. A brand or product LoRA takes ~30 minutes to train on 20-30 images on a rented GPU and produces reliable brand-consistent output.
What's the difference between Flux Pro 1.1 and Ultra? Ultra outputs up to 4 megapixels at higher sharpness — worth it for print, hero images, or anything at large display size. Pro 1.1 at 1024x1024 is the right default for web and social.
Flux is the correct default for any 2026 image workflow where photorealism matters. Product builders should use the hosted API and not think about it further. Working creators doing photographic subjects — portraits, products, cinematic stills, food, interiors — will find Flux Pro 1.1 the sharpest output available and cheaper per image than most alternatives at real usage volumes. Illustrators, art directors, and editorial creators whose output leans stylized are better served by Midjourney's house aesthetic and --sref workflow. Technical teams building any custom image pipeline should download Flux [dev] and train a LoRA — it is the one open-weight model in 2026 that competes with commercial API quality. If you need one hosted image API to embed in a product this year, this is the one.