ToolScore
Image Gen 5 min read

The License Traps Hiding in 'Open' Image Models

9 traps, 1 US ban the number that matters here

By Stan K.

Qwen-Image-2.1 quietly switched to a non-commercial Research License while the original stayed Apache. MiniMax-H3-Image bans the USA verbatim. FLUX.1-dev is NC as a model but commercial in its outputs. Nine traps verified from LICENSE files and terms of service — including the Midjourney ban rumor that is wrong.

Open-weight image models ship with a README that says “open” and a LICENSE that says something else. Nine traps we verified by reading the actual license files and terms of service — each one is a live clause, not a rumor. Two of them can end a US-based project on the spot.

1. The version bump that removes your rights

Qwen-Image-2.1 ships under the Qwen Research License — non-commercial. The original Qwen-Image and Qwen-Image-Edit are Apache 2.0. Nothing about the naming warns you: “just updated the version” silently swaps a permissive license for a research-only one. Any pipeline that pulled the newer checkpoint without re-reading the LICENSE changed its legal footing overnight. Pin your model versions and re-read the license on every bump.

2. The model that is illegal to use in the United States

MiniMax-H3-Image is banned in the USA — in the license text itself

The MiniMax Community License (dated 2026-11-10, Nanonoble Pte. Ltd.) defines its excluded territories verbatim: “Excluded Territories means the European Union, the United Kingdom, the Republic of Korea and the United States of America.” A 33.1B DiT spread over 130GB+ across multiple GPUs — and unusable for commercial work in the US, full stop. Check territory clauses before you spend a week on weights.

3. NC model, commercial outputs: the FLUX.1-dev reversal

FLUX.1-dev inverts the usual trap. The model is non-commercial — but the outputs “may be use[d] for any purpose (including for commercial purposes)”. Free to run, free to sell what it makes; serving third parties with it still requires a paid BFL license. Most licenses trap outputs; this one traps the deployment. If you run it on a 3060 like we do, our production notes are in FLUX vs Qwen-Image.

4. The ban list that doesn’t include the US

HunyuanImage looks like the same trap as MiniMax and isn’t. The Tencent Hunyuan Community License excludes the EU, the UK and Korea — the USA is not on the list. HunyuanImage-2.1 (17B, 24GB VRAM with FP8 and offload for 2048²) and the 83B HunyuanImage-3.0 both carry it. “China-based vendor” is not a legal category; only the named territories are.

5. The $1M revenue cap

Stability’s SD 3.5 Community License is free while your annual revenue stays under $1M — over the cap, you need a paid deal. The cap is the whole clause: no territories, no output restrictions, just a revenue tripwire that fires the year your business works. (The community-reported history of an earlier license rollback circulates as a cautionary tale; the current 05.07.2024 license text is the only thing that binds you.)

6. Midjourney: ownership has a revenue clause — and the ban rumor is wrong

Midjourney’s ToS grants real ownership — with two teeth. First: if you (or your employer) are a company with over $1M in annual revenue, Pro or Mega is mandatory to own your Assets at all — and it reaches employees’ personal accounts. Second: Midjourney can “suspend or ban… at any time, and for any reason”, without refund. But the rumor that a ban costs you your rights is wrong: ownership of already-created Assets survives termination — the license “survives termination”. Also note there is no public banned-words list to comply against; the moderation line is invisible until you cross it.

7. Firefly credits evaporate monthly

Adobe’s own FAQ: generative credits “do not roll over to the next month”. Standard generations are unlimited in the new plans; the premium pool is 2,000 (Standard), 7,000 (Pro) and 50,000 (Premium) premium generations per month. Bursty workloads — the kind image teams actually have — forfeit the difference every cycle.

8. The upscale that bills twice

Midjourney’s docs say it plainly: “Upscaling can cost twice as many GPU minutes as generating your initial images.” On metered GPU-minute plans, the finalize step can cost more than the exploration. Budget keepers at 2× their sticker in GPU time — the per-keeper math compounds this on top of the rejection rate.

9. The watermark with no opt-out

Google stamps SynthID on every generated image — official policy, no option to remove it, on Nano Banana and Nano Banana Pro alike. For provenance that’s arguably good citizenship; for compositing into client deliverables it’s a term of service, not a footnote. Watermark claims we couldn’t verify in live official docs stay off this page.

The 60-second license audit

  • Territory clause: are you in an “Excluded Territory”? (MiniMax: US banned; Hunyuan: US fine)
  • Version pin: does the LICENSE you read match the checkpoint you pulled? (Qwen 2.1 vs Apache original)
  • Commercial axis: model, outputs, or both? (FLUX.1-dev: outputs yes, serving no)
  • Revenue caps: yours, or your employer’s? (SD 3.5 $1M; Midjourney Pro/Mega over $1M)
  • Rollover and metering: do credits survive the month? Does the finalize step bill double?

The full model-by-model table — every license, VRAM figure and price in one place — is in the Image Gen hub, and the stack that keeps you off this whole list is in When Local Image Gen Beats Every API.

Numbers change — the math doesn't
Token prices and plan limits shift monthly; the structural arguments in this guide age slowly. For current per-tool numbers see the live rankings.