For two years, "which open image model" had one obvious answer. Stable Diffusion was the ecosystem — the LoRAs, the UIs, the tutorials, the fine-tunes. Then Black Forest Labs — founded by people who had worked on widely used open image models — started shipping FLUX, and the answer stopped being obvious.
In 2026 the comparison is no longer about which produces prettier pictures. It is about licence terms, model size versus your GPU, and which ecosystem you are willing to be locked into.
Sourcing note: FLUX specifications come from Black Forest Labs' official model pages, GitHub repository, Hugging Face model cards, and pricing page. Stable Diffusion specifications come from Stability AI's official Stable Image page. No cross-family benchmark is quoted, because neither vendor has published a neutral one. Last verified July 28, 2026.
What this comparison solves
The pain point: most FLUX-vs-SD articles compare sample images and stop. That decides nothing. What actually breaks projects is discovering the licence after launch, or downloading a 32B checkpoint onto a 12 GB card.
The differentiator: this compares the two families on the four axes that carry real consequences — licence, hardware, resolution, and ecosystem — and is explicit about where a vendor's claim about a rival should be treated as marketing rather than evidence.
The line-ups as they stand
Black Forest Labs — FLUX
| Model | Size | Licence | VRAM | Notes |
|---|---|---|---|---|
| FLUX.2 [klein] 4B | 4B | Apache 2.0 | 8.4 GB | ~1.2s on RTX 5090 |
| FLUX.2 [klein] 4B Base | 4B | Apache 2.0 | 9.2 GB | Undistilled, for fine-tuning |
| FLUX.2 [klein] 9B | 9B | FLUX Non-Commercial | 19.6 GB | ~2s |
| FLUX.2 [dev] | 32B | Non-Commercial + licence | fp8 path for consumer GPUs | BFL: most capable open-weight image model |
| FLUX.1 [schnell] | 12B | Apache 2.0 | — | Previous generation, permissive |
| FLUX.1 [dev] | 12B | Non-Commercial | — | Largest LoRA ecosystem |
Hosted tiers — FLUX.2 [pro], [flex], [max] — run from $0.03 to $0.07 for the first megapixel, with editing up to 4MP and up to 10 reference images.
Stability AI — Stable Diffusion 3.5
Per Stability's official page, SD 3.5 comes in three variants:
- Large — the most powerful in the family, positioned for professional use at 1 megapixel resolution
- Turbo — faster, generating high-quality images with strong prompt adherence in four steps
- Medium — the quality/customisation balance, designed to run on consumer hardware
Stability offers a self-hosted licence, a platform API, and cloud partner deployment. Their stated strengths: versatile styles, market-leading prompt adherence rivalling much larger models, and diverse outputs.
Axis 1 — Licence (decide here first)
This is the axis that eliminates options fastest.
FLUX: a genuinely permissive tier exists. FLUX.1 [schnell], FLUX.2 [klein] 4B and 4B Base, and the FLUX.2 VAE are Apache 2.0 — commercial use included, no licence conversation. Everything else in the open line-up is non-commercial by default, with commercial rights sold through tiers: Builder (klein models, 10K images/month, 1 domain), Platform (klein Base 9B + FLUX.2 [dev], 100K images/month), Professional (FLUX.2 [dev], 100K/month, up to 3 domains, aimed at agencies), and Enterprise. A synthetic-data licence with rights to train on outputs exists as a separate track.
Stable Diffusion: Stability routes commercial self-hosting through their own licence, alongside API and cloud-partner options.
Practical rule: if you need free commercial rights on downloadable weights and do not want a licence negotiation, the FLUX Apache 2.0 tier is the shortest path. If you are licensing commercially either way, this axis stops being decisive and hardware takes over.
Axis 2 — Hardware
FLUX publishes precise numbers, which makes planning easy: 8.4 GB VRAM for klein 4B, 9.2 GB for 4B Base, 19.6 GB for klein 9B, 21.7 GB for 9B Base. FLUX.2 [dev] at 32B is a different class — BFL points consumer-GPU users to an fp8 reference implementation built with NVIDIA and ComfyUI.
Stability positions SD 3.5 Medium explicitly for consumer hardware, with Turbo as the speed option at four steps.
Practical rule: under ~10 GB VRAM, you are choosing between FLUX.2 klein 4B and SD 3.5 Medium/Turbo. Above ~20 GB, FLUX's larger open models come into range and the quality gap widens in FLUX's favour by parameter count alone.
Or skip the GPU question entirely. Flux 3 AI runs generation, editing, reference workflows, and upscaling in the browser — no checkpoints, no VRAM, no licence homework. Open the image generator or see the credit plans.
Axis 3 — Resolution and editing
This is where the families genuinely diverge.
| FLUX.2 | SD 3.5 | |
|---|---|---|
| Output resolution | Up to 4MP (editing at up to 4MP) | Large positioned at 1MP |
| Reference images | Up to 10 simultaneously | — |
| Editing in the same model | Yes, single checkpoint | Separate tooling |
| Text rendering | A stated focus; documented hex and typography control | Listed among general improvements |
FLUX.2 combining generation and multi-reference editing in one checkpoint is the structural advantage. If your work is product photography, brand-consistent campaign sets, or anything where the same character or product recurs, that capability is the reason to choose FLUX — independent of image quality.
Axis 4 — Ecosystem
Stable Diffusion's advantage is accumulated: years of LoRAs, fine-tunes, ControlNets, UIs, and community knowledge. For niche styles and specific characters, someone has probably already trained what you need.
FLUX's ecosystem is younger but growing fast, and BFL is deliberately feeding it — Apache 2.0 variants, reference inference code on GitHub, an fp8 path built with NVIDIA and ComfyUI, klein models shipping on-device with ASUS ProArt laptops, and licence tiers that explicitly include fine-tuning and LoRA rights.
Practical rule: if your project depends on existing community fine-tunes, SD's back catalogue is a real asset. If you are training your own, FLUX's newer base models and licence structure are the better foundation.
A note on vendor comparisons
BFL's own licensing page positions their open weights against "closed APIs" (naming Google, OpenAI, Midjourney) and "open source" (naming Stable Diffusion, Qwen, Wan). That is a vendor's framing of its rivals, and it belongs in this article as positioning, not as evidence.
The same caution applies in reverse. Neither BFL nor Stability has published a neutral head-to-head, so anyone showing you a decisive FLUX-vs-SD quality score built one themselves. Decide on licence, hardware, resolution, and ecosystem — those are verifiable.
FAQ
Is FLUX better than Stable Diffusion? No neutral benchmark exists. FLUX offers larger open models (32B), higher resolution (4MP), and multi-reference editing; SD offers a mature ecosystem and models tuned for consumer hardware.
Which is free for commercial use? FLUX.1 [schnell] and FLUX.2 [klein] 4B / 4B Base are Apache 2.0. Stability routes commercial self-hosting through its own licence.
Which needs less VRAM? FLUX.2 klein 4B at 8.4 GB is the lightest documented FLUX option; SD 3.5 Medium is explicitly positioned for consumer hardware.
Which is better for text in images? FLUX documents specific typography techniques including quoted strings and hex colour binding, and BFL cites text rendering as a headline FLUX.2 improvement.
Which has more LoRAs? Stable Diffusion, by a wide margin — a function of time in the field.
Can either generate video? Within the scope of this comparison, no. BFL's video model is FLUX 3, announced July 23, 2026 and currently in gated early access.
Bottom line
Choose FLUX if: you need permissive commercial weights without a licence deal (klein 4B, schnell), you want maximum open-weight quality (FLUX.2 [dev], 32B), you need 4MP output, or your work depends on multi-reference consistency.
Choose Stable Diffusion if: your pipeline depends on existing community fine-tunes, or you are optimising for modest consumer hardware within Stability's licensing.
Choose neither yet if: what you actually need is output this week rather than a stack. Generate in the Flux 3 AI browser workspace and revisit the local-inference question when the work justifies it.
Sources
- Stable Diffusion 3.5 (Stability AI official page) — SD 3.5 Large / Turbo / Medium positioning, 1MP, four-step Turbo, deployment options
- FLUX.2 klein model page (BFL) — variant table, licences, VRAM, inference times
- FLUX.2: Frontier Visual Intelligence (BFL, November 25, 2025) — FLUX.2 [dev] 32B, fp8 implementation with NVIDIA and ComfyUI, 4MP editing, multi-reference
- black-forest-labs/flux on GitHub — open model list, licences, reference inference code
- FLUX.1-schnell model card (Hugging Face) — 12B parameters, Apache 2.0
- FLUX.1-dev model card (Hugging Face) — non-commercial licence terms
- FLUX open weights licensing (BFL) — commercial tiers, synthetic-data licence, vendor positioning against open and closed rivals
- BFL API pricing — hosted tier pricing and billing rules
- FLUX.2 pro / flex model page — multi-reference control and production positioning
- FLUX 3 — Real World Models (BFL, July 23, 2026) — current status of BFL's video-capable model
Scope note: licence summaries are informational, not legal advice. Flux 3 AI is an independent creator workspace, not affiliated with Black Forest Labs or Stability AI. Verify current terms at bfl.ai and stability.ai.
