Black Forest Labs' open-weight model that beats DALL-E 3 on hands, text, and prompt fidelity. Run via fal.ai, Replicate, or local for full control.
Image Generation
Flux.1 is Black Forest Labs' text-to-image family — the models the ex-Stable Diffusion core team shipped after leaving Stability AI in 2024. The lineup runs from Flux.1 Schnell (Apache 2.0, four-step distilled, generates a 1024px image in about a second on an H100) through Flux.1 Dev (non-commercial open weights, the quality sweet spot devs actually download from Hugging Face) up to Flux 1.1 Pro Ultra (API-only, 4-megapixel outputs, closed weights). By late 2026 Black Forest Labs has moved the flagship banner to FLUX.2 and previewed FLUX 3, but Flux.1 remains the workhorse the community fine-tunes, the model most ComfyUI workflows are built around, and the family with the deepest LoRA ecosystem on Civitai.
The honest pitch: if you need an image model you can run on your own GPU, ship inside a product without a per-request Adobe or OpenAI fee, or train custom LoRAs against, Flux.1 is the strongest option in that lane. If you want to type a prompt into a web app and get a beautiful image with zero setup, this is not that product — you either use one of the hosted API providers (Replicate, Fal, Together, Runware) or you install ComfyUI and learn the node graph. There is no official consumer web app.
Positioning against the rest of the market: Midjourney still wins on default aesthetic taste out of the box. DALL-E 3 (via ChatGPT) wins on prompt understanding for casual users. Ideogram wins on legible in-image text and typography. Adobe Firefly wins on commercial safety for enterprise legal teams. Flux.1 wins on the axis none of them touch — open weights, controllable inference, on-prem deployment, and per-image API costs an order of magnitude below Midjourney's subscription math for high-volume workloads.
Three-tier open family with a real license split. Flux.1 Schnell ships under Apache 2.0 — you can put it in a commercial product, sell the output, retrain it, no royalty. Flux.1 Dev is non-commercial only unless you buy a commercial license from Black Forest Labs (contact sales, quote-only, [verify pricing] for your seat count). Flux.1 Pro and 1.1 Pro are API-only, no weights released. This split matters because teams keep confusing "Flux is open" with "any Flux model is free to ship." It isn't.
Prompt adherence that closes the gap with DALL-E 3. Flux.1 uses two text encoders — T5-XXL alongside CLIP — instead of the CLIP-only encoder Stable Diffusion 1.5 and SDXL used. In practice this means it actually renders "a red cube stacked on top of a blue sphere" the right way around, holds multi-clause prompts across 200+ tokens, and handles compositional geometry that made SD 1.5 hallucinate. Not at DALL-E 3's level for surreal creative interpretation, but for descriptive prompts it is the strongest open model in 2026.
Legible in-image text. Flux.1 renders words inside images correctly to about 15-20 characters — signage, book covers, product labels, poster text. Ideogram is still tighter for typography-heavy layouts, but Flux closed most of the gap SDXL never could. If you need "OPEN" on a diner sign or "Chapter 1" on a book cover, it works on the first try.
Flux Kontext — instruction-based image editing. Released 2025, Kontext takes an input image plus a text instruction ("change the shirt to red", "put this person on a beach") and edits in place while preserving identity and style. Kontext Pro runs $0.04 per image on Fal; Kontext Max sits around $0.08 per image [verify pricing on your provider]. This is the feature that made Flux relevant to product teams building consumer photo editors — you don't need to fine-tune, you just prompt.
Guidance-distilled inference for speed. Schnell generates in 1-4 steps instead of the 25-50 steps standard diffusion needs. On an A100 you're looking at sub-second 1024px generations, which makes it the only open model realistic for real-time interactive workflows — Krea-style live canvas, chat-based image loops, in-editor previews.
Native LoRA training and the largest fine-tune ecosystem after SDXL. Civitai has thousands of Flux.1 Dev LoRAs — style transfers, character consistency, aesthetic packs. Training a LoRA on Replicate costs roughly $2-5 for a 1000-step run; on your own 4090 it is a few hours and free. This is the reason ComfyUI power users picked Flux over Midjourney: you can bend it to a specific brand style or character.
Multiple resolutions and aspect controls. Standard Flux.1 Dev handles 1024x1024 comfortably; Flux 1.1 Pro Ultra generates up to 2048x2048 (4MP) natively without a separate upscaler pass. Aspect ratios from 21:9 down to 9:21 are supported directly — no tiling artifacts, no seam issues you had to fix with SDXL.
Deployment optionality no closed model matches. The same weights run on Replicate (~$0.003 per Schnell image, ~$0.025 per Dev, ~$0.04-0.05 per 1.1 Pro), Fal (similar bands, $0.05/megapixel for Pro confirmed), Together AI, Runware, or self-hosted via ComfyUI, Diffusers, or vLLM-compatible servers. If one provider's latency spikes or prices change, you switch endpoints without changing your model choice.
Flux.1 has no single unified pricing page because pricing depends on where you run it. Black Forest Labs' own API (bfl.ai) is pay-as-you-go with no per-seat fee; hosted providers each set their own rates. Approximate 2026 per-image costs on the major hosts:
| Model | Approx. price per image | License | Best for |
|---|---|---|---|
| Flux.1 Schnell | $0.003 (Replicate/Fal), free self-hosted | Apache 2.0 | High-volume, commercial, real-time |
| Flux.1 Dev | ~$0.025 hosted, free self-hosted (non-commercial) | Non-commercial | Personal projects, fine-tuning, research |
| Flux.1 Dev (commercial license) | Quote-only from BFL [verify pricing] | Commercial (paid) | Startups shipping Dev weights in production |
| Flux 1.1 Pro | ~$0.04 per image | API-only, closed | Product teams wanting quality without ops |
| Flux 1.1 Pro Ultra | ~$0.06 per image | API-only, closed | 4MP high-detail work |
| Flux Kontext Pro | $0.04 per image (Fal confirmed) | API-only | In-place instruction editing |
| Flux Kontext Max | ~$0.08 per image | API-only | Highest-fidelity editing |
| BFL API — video (FLUX 3) | $0.17 per second | API-only | Video generation |
Compared to alternatives: Midjourney's cheapest plan is $10/month for ~200 fast images (about $0.05 each, but subscription-locked). DALL-E 3 via the OpenAI API is $0.040 per standard image and $0.080 for HD. Ideogram's paid plans start at $8/month. Adobe Firefly bundles into a $9.99/mo credit plan. For a developer generating 10,000 images a month, Flux Schnell on Replicate costs about $30 vs. Midjourney's $60 Pro tier vs. DALL-E 3's roughly $400 — that spread is why Flux dominates programmatic workloads.
Pros
Cons
Midjourney — from $10/month. Pick Midjourney if you are a solo creator or small team who wants gorgeous images with zero setup, prompt in Discord or the web app, and never touch an API. Pick Flux.1 if you need programmatic generation, custom LoRAs, or per-image costs below $0.01.
DALL-E 3 (via ChatGPT / OpenAI API) — $0.040-0.080 per image. Pick DALL-E 3 if prompt understanding for casual, conversational requests matters more than fine control — it is still the best at interpreting "make it feel cozy" without a hundred descriptors. Pick Flux.1 if you need reproducibility (same seed, same output), fine-tuning, or self-hosting.
Ideogram — from $8/month. Pick Ideogram if your work is typography-forward: logos, posters, layouts where the text has to be perfect. Pick Flux.1 if you need broader image types and your text does not need to be flawless past 20 characters.
Stable Diffusion 3.5 (Stability AI) — free open weights, API from ~$0.065/image. Pick SD 3.5 if you need the largest existing ecosystem of extensions, ControlNets, and community models (still deeper than Flux's, though narrowing). Pick Flux.1 if you want stronger baseline prompt adherence and better default quality without stacking twelve ControlNets.
Adobe Firefly — $9.99/mo standard, $29.99/mo pro. Pick Firefly if you are an enterprise team that needs legal indemnification and Creative Cloud integration. Pick Flux.1 if you do not need Adobe's commercial-safety paperwork and want an order of magnitude more control.
Buy Flux.1 if you are a developer shipping image generation inside a product, a technical creator running ComfyUI with custom LoRAs, or a startup that needs per-image costs below Midjourney's subscription math — Schnell on Replicate for volume, 1.1 Pro on Fal when quality matters, Kontext Pro when you are editing. Skip Flux.1 if you are a solo creator who just wants pretty images in a browser — Midjourney at $10/month is the correct answer for that use case, and no amount of open-weights ideology changes it. Pick Flux Schnell first (it is basically free), promote to Flux 1.1 Pro only when Schnell's output cannot cross your quality bar, and read the Dev license before you ship anything using those weights.
Q: Is Flux.1 free? A: Flux.1 Schnell is free under Apache 2.0 — you can self-host and use it commercially at zero cost. Flux.1 Dev is free to download but non-commercial only; commercial use requires a paid license from Black Forest Labs. Flux.1 Pro and 1.1 Pro are API-only and cost roughly $0.04-0.06 per image depending on the host.
Q: How does Flux.1 compare to Midjourney for image quality? A: Midjourney still wins on default aesthetic — its images look more artistically tuned out of the box. Flux 1.1 Pro Ultra closes most of the technical quality gap (resolution, sharpness, coherence) but loses on the "feels like art" dimension. For photorealistic work Flux is competitive; for stylized art Midjourney is ahead.
Q: Can I use Flux.1 Dev in a commercial product? A: Not with the default weights. Flux.1 Dev ships under a non-commercial license. To use it commercially you either buy a commercial license from Black Forest Labs (quote-only, contact their sales), use Flux.1 Schnell instead (Apache 2.0, no restrictions), or use the Pro tiers via API where the hosting provider's terms cover commercial use.
Q: Which host should I use — Replicate, Fal, or Together? A: Replicate has the best documentation and easiest onboarding for prototypes. Fal has the lowest latency and often the sharpest pricing for high-volume Kontext workloads. Together AI is competitive for Schnell at scale. All three run the same weights, so start with whichever your stack already touches and switch if cost or latency becomes an issue.
Q: Can Flux.1 generate video? A: Not directly — Flux.1 is text-to-image only. Black Forest Labs' video model is FLUX 3, priced around $0.17 per second on the BFL API and released as part of the FLUX.2/FLUX 3 lineup. For Flux.1-specific video you would use a separate model like Runway Gen-4 or a workflow that chains Flux stills through an image-to-video model.
Q: How much VRAM does self-hosting Flux.1 Dev need? A: Flux.1 Dev at full precision (FP16) needs about 24GB of VRAM — that puts it on a 3090, 4090, or A100. Quantized to 8-bit or 4-bit via bitsandbytes or GGUF, it runs on a 12GB 3060 with some quality loss. Schnell has the same footprint but generates in 4 steps instead of 25, so it is dramatically faster per image.
Q: Is Flux.1 or FLUX.2 the right choice in 2026? A: FLUX.2 is Black Forest Labs' newer flagship and has better quality on paper, but Flux.1 has the mature LoRA ecosystem, deeper ComfyUI workflow support, and stable, well-documented hosted endpoints. For production work today Flux.1 (specifically 1.1 Pro or Schnell) is the safer pick; FLUX.2 is worth testing if you are starting a new project and do not need existing fine-tunes.
Q: Does Flux.1 have built-in content filtering? A: The hosted Pro APIs on BFL, Replicate, and Fal apply content filters — NSFW, likeness of public figures, and violent content are blocked at the provider level. Self-hosted Schnell and Dev have no filtering; you are responsible for whatever the model outputs. That is a feature for research use and a legal risk for consumer-facing products, so most teams shipping to end users add a moderation layer on top.