# Model Licensing — Verified Research *Last verified: April 2026. Sources: HuggingFace model cards, GitHub license files, BFL legal pages.* --- ## TL;DR for this project For a commercial project with a French rights holder (Banijay), only **Wan2.1** and **Wan2.2** are unambiguously usable without a legal negotiation. Everything else has a blocker. **Recommended model: Wan2.2 — Apache 2.0, open weights, genuinely hybrid (images + video), no caveats.** --- ## FLUX Family — Black Forest Labs ### FLUX.1 [schnell] - **License:** Apache 2.0 — verified from GitHub `model_licenses/LICENSE-FLUX1-schnell` - **Commercial use:** ✅ Free — no revenue threshold, no geography, no caveats - **Fine-tune:** ✅ Yes - **Output ownership:** ✅ Yours - **Catch:** Lowest quality tier in the FLUX family. Distilled 4-step model optimised for speed. Not the quality ceiling. - **Verdict for this project:** Technically clean but not the right model for production quality. ### FLUX.1 [dev] — and all [dev] variants *Covers: FLUX.1 dev, Fill dev, Depth dev, Canny dev, Redux dev, Kontext dev* - **License:** FLUX.1 [dev] Non-Commercial License v1.1.1 - **Commercial use:** ❌ **Explicitly non-commercial** - Direct quote: *"use for revenue-generating activity is NOT a Non-Commercial Purpose"* - Fine-tuning the weights is also non-commercial only - Commercial use requires a separate paid license from BFL → bfl.ai - **Fine-tune:** ✅ Yes, but for non-commercial purposes only - **Output ownership:** ✅ Outputs are yours (including commercial use of outputs) - **Verdict for this project:** ❌ Blocked. Weights cannot be used commercially without a BFL commercial agreement. ### FLUX.2 [dev] and FLUX.2-Klein - **License:** Custom "other" — full text not yet published as of April 2026 - **Commercial use:** ❌ Follows same non-commercial pattern as FLUX.1 [dev] - **BFL commercial path:** FLUX.2 [pro] and [max] are API-only (no weights). Self-hosted commercial use of [dev] requires BFL's Self-Hosted Commercial License Terms. - **Verdict for this project:** ❌ Blocked unless paid commercial license negotiated with BFL. ### FLUX.2 [pro] / [max] - **Weights:** None — API only via bfl.ai - **Commercial use:** ✅ Via API, pay-per-use - **Fine-tune:** ❌ Not possible - **Verdict for this project:** ❌ Cannot fine-tune; API-only with no franchise specificity. --- ## LTX-Video — Lightricks ### LTX-Video 13B v0.9.8 (current latest) *Covers all versions released since April 15, 2025 under LTXV Open Weights License 0.X* - **License:** LTXV Open Weights License 0.X (custom — not Apache, not MIT) - **Commercial use:** ⚠️ **Conditional on revenue** - Free for entities with annual revenue **under $10M** - Revenue ≥ $10M → **must obtain a paid commercial license** from Lightricks - Contact: `ltxv-licensing@lightricks.com` - Penalty for breach: double the license fees that would have been owed, plus legal costs - **Fine-tune:** ✅ Yes (LoRA included in license) - **Output ownership:** ✅ Outputs are yours - **Applies to:** All 2B and 13B variants, ICLoRA models, upscalers, distilled versions - **Verdict for this project:** ❌ Banijay Group annual revenue ~€3B. The $10M threshold is triggered. Even Cultshot as a vendor for Banijay likely pulls this in — requires legal review. Cannot use without a paid Lightricks agreement. --- ## HunyuanVideo — Tencent ### HunyuanVideo + HunyuanVideo-I2V (latest) - **License:** Tencent Hunyuan Community License - **Commercial use:** ✅ Free (under 100M MAU threshold, no revenue threshold) - **Fine-tune:** ✅ Yes - **Output ownership:** ✅ Yours - **EU/UK blocker:** ❌ **FATAL for this project** - Direct quote: *"THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA"* - Using this model in France (where Banijay and Laurent are based) = **unauthorized and unlicensed** - Governing law: Hong Kong SAR courts - **Additional restriction:** Cannot use model outputs to improve any other AI model - **Verdict for this project:** ❌ Hard disqualification. Banijay is French. EU territory is explicitly excluded from the license. --- ## Wan Family — Wan-AI / Alibaba ### Wan2.1 — all variants *T2V-14B, T2V-1.3B, I2V-14B-720P, I2V-14B-480P, VACE-14B* - **License:** Apache 2.0 — verified from HuggingFace model cards - **Commercial use:** ✅ Free — no revenue threshold, no geography, no MAU limit - **Fine-tune:** ✅ Yes - **Output ownership:** ✅ — direct quote: *"We claim no rights over your generated contents"* - **Open weights:** ✅ Yes — no gate, no approval required - **Verdict for this project:** ✅ Clean. No blockers. ### Wan2.2 — all variants *T2V-A14B (MoE), I2V-A14B (MoE), TI2V-5B, Animate-14B* - **License:** Apache 2.0 — verified from HuggingFace model cards - **Commercial use:** ✅ Free — same as Wan2.1, pure Apache 2.0 - **Fine-tune:** ✅ Yes - **Output ownership:** ✅ Same quote: *"We claim no rights over your generated contents"* - **Open weights:** ✅ Yes — no gate - **Upgrade over Wan2.1:** MoE architecture, +65% more image training data, +83% more video training data, cinematic-level aesthetic training, 720p/24fps on single 4090 (5B) or multi-GPU (14B) - **Verdict for this project:** ✅ **Clean. Recommended.** ### Wan2.7 - **License:** Proprietary — no open weights released as of April 2026 - **Availability:** API only via Atlas Cloud (`atlascloud.ai`) - **Pricing:** $0.026–$0.085/use depending on model and output - **Fine-tune:** ❌ Not possible — closed model, no weights - **Note:** Full suite (T2I, I2I, T2V, I2V, V2V, Reference-to-Video) available via API. Reference-to-Video (up to 5 character references → new video) is the most relevant capability for this project as a proof-of-concept tool — not a fine-tuning base. - **Verdict for this project:** ❌ Cannot fine-tune. Useful for fast API-based proof-of-concept testing only. --- ## Summary Table | Model | Commercial | EU/UK | Revenue bar | Fine-tune | Open weights | |---|---|---|---|---|---| | FLUX.1 [schnell] | ✅ free | ✅ | none | ✅ | ✅ | | FLUX.1 [dev] | ❌ NC only | ✅ | N/A (NC) | NC only | ✅ | | FLUX.2 [dev] | ❌ NC only | ✅ | N/A (NC) | NC only | ✅ | | LTX-Video 13B | ⚠ <$10M free | ✅ | $10M → paid | ✅ | ✅ | | HunyuanVideo | ✅ (<100M MAU) | ❌ **EU excluded** | none | ✅ | ✅ | | Wan2.1 | ✅ free | ✅ | none | ✅ | ✅ | | **Wan2.2** | **✅ free** | **✅** | **none** | **✅** | **✅** | | Wan2.7 | ✅ API | ✅ | per-use | ❌ | ❌ | --- ## Practical implications ### For the meeting with Banijay/Laurent - **Do not propose LTX** without a prior Lightricks commercial license conversation. The $10M clause is enforced with double-damage penalties. - **Do not propose HunyuanVideo** at all — Tencent has explicitly excluded EU from the license. This is not a grey area. - **FLUX [dev] variants** are not commercially usable as open weights. If BFL is referenced, it needs to be as an API cost, not as a "free open source model." - **Wan2.2 is the correct answer** — commercially clean, open weights, fine-tuneable, genuinely hybrid (images + video), Apache 2.0. ### Pipeline update required The current `devenv.nix` targets `Wan2.1-I2V-14B-720P`. This should be updated to `Wan2.2` variants: - `Wan2.2-TI2V-5B` for the hybrid single-model path (images + video, runs on one 4090) - `Wan2.2-T2V-A14B` for higher quality video generation - `Wan2.2-I2V-A14B` for image-to-video Same Apache 2.0 license, same fine-tuning approach, better model. ### LTX as a future option If this project grows to a point where a commercial license conversation with Lightricks makes sense (and Banijay/Cultshot want LTX's specific capabilities), it's a straightforward negotiation — Lightricks has a standard commercial agreement process. Not a blocker in principle, just a step that must happen before use.