# Totally Spies AI Pipeline — Meeting Talking Points *Quick-scan format — follow the flow top to bottom* *Model licensing verified April 2026 — see `model-licensing.md` for full source citations* --- ## 1. Why generic AI tools won't work - Foundation models (Sora, Kling, Midjourney, Wan2.1 base) can make *some* animation — not *this* one - **Sam, Clover, Alex share the same body proportions** — the only differentiators are hair colour and suit colour - Without explicit anchoring, a 235B VLM confuses Sam and Clover in ~30% of shots - A generation model will do the same → brand error on every frame - **S7 is a visual reboot** — different art style from S1–6 (cleaner linework, flat colour, anime-influenced) - A model trained on the back catalogue produces the *wrong* S7 look - **New S7 characters exist nowhere in public training data** - Zerlina Lewis, Toby, Cyberchac, WOOHP World Singapore, the WOOHP-e device - You cannot prompt a foundation model into knowing what these look like - Only path: fine-tuning on franchise material --- ## 2. The method — Bible-First Fine-Tuning **Three phases, deliberately sequenced:** ### Phase 1 — Build the visual bible ✅ Done - Catalogued every shot across 13 S7 episodes before touching any generation model - Cross-validated everything against the official Totally Spies Wiki - Corrected villain identities for all 13 episodes - Anchored Sam = orange hair + green suit, Clover = blonde + red suit, Alex = black hair + yellow suit - Named ~30 gadgets by their exact wiki names (e.g. WOOHP-e, Ultra-fixative Structural Foam) - Built 18-category canonical location hierarchy (Singapore city, WOOHP HQ, AIYA Academy, etc.) - The bible is the **training signal** — caption quality determines output quality - "Woman in green suit" → model learns vague style - "Sam (red-orange hair, green catsuit) in WOOHP HQ, gadget insert, WOOHP-e device" → model learns the franchise grammar ### Phase 2 — Caption-supervised training clips ✅ Done - Extracted **1,551 clips** from 13 episodes — 106 minutes of 720p footage - Generated a VLM caption for every clip using the bible as context - Character visual anchors injected into every prompt → VLM cannot swap Sam and Clover - Episode-specific villains and gadgets injected per clip - Speaker-attributed dialogue (51% of 29,731 transcript words) used as grounding - Result: 96% of clips correctly identify characters, 49% name gadgets by wiki name ### Phase 3 — LoRA fine-tuning ⏳ Ready to run - One model targeted: **Wan2.2** (Wan-AI / Apache 2.0) - Genuinely hybrid: generates images AND video from the same weights - T2V: text prompt → video clip - I2V: reference frame → animated clip (character identity locked by input) - T2I / image mode: text → still (720p, marketing-ready quality) - Runs on a single RTX 4090 (5B model) or multi-GPU (14B MoE model) - **LoRA** = Low-Rank Adaptation — a thin franchise-specific layer on top of existing model - Base model keeps its knowledge of motion, physics, animation - LoRA layer learns the S7 style, characters, locations - Training time: ~8–12 hours on one GPU (not weeks, not six figures) - Output: small portable weights (~few hundred MB), versioned, updatable - **Single remaining dependency: licensed episode files from Banijay** --- ## 3. The specific problems we've already solved | Problem | What we did | |---|---| | Character confusion (Sam/Clover/Alex) | Wiki-anchored every caption; hair + suit colour explicit in every VLM prompt | | S7 art style vs earlier seasons | S7-only analysis; style tagged as "2D digital, flat colour, anime-influenced" in every caption | | New S7 characters not in any dataset | Built descriptions from scratch via bible; fine-tuning teaches the model their appearance | | Singapore setting | 18-category location hierarchy with full scene descriptions per location | | WOOHP-e device (was misidentified as a car/tablet) | Wiki-corrected in the bible; correct in every caption where it appears | | 13 different villain designs | Per-episode villain database; injected into captions for that episode's clips | | Marketing shot grammar | Shot taxonomy built from the reference material; encoded into every caption | --- ## 4. What the pipeline produces in practice - Writer/director provides a **brief** → "trio confronts villain on Singapore rooftop, Clover leads, mid-morning" - Brief is translated into a structured prompt using our prompt bible - **Wan2.2 T2V + LoRA** generates a 5–8 second clip directly from the prompt - **Wan2.2 I2V + LoRA** animates from a storyboard frame or key art reference — character identity locked by the input image - Clips are assembled into a marketing sequence - The model knows the franchise — it doesn't need to be told "make it look like Totally Spies" --- ## 5. Pipeline state right now | What | Status | |---|---| | Visual bible — 2,852 shots, 13 episodes | ✅ Done | | Wiki cross-validation — villains, gadgets, characters | ✅ Done | | Training captions — 1,551 clips, wiki-anchored | ✅ Done | | Dataset packages — Wan2.2 format | ✅ Ready | | Training infrastructure — GPU builder provisioned | ✅ Ready | | Fine-tuning run | ⏳ Waiting for licensed episode files | | Episodes 14–26 | ⏳ Behind paywall — monitored for YouTube release | **When licensed files arrive → pipeline runs in 12–16 hours, end to end.** The clips change. The captions, the pipeline, the infrastructure — all stay the same. --- ## 6. Why not the alternatives | Alternative | Why no | |---|---| | Prompt-only (Midjourney / Sora / ChatGPT) | Character consistency impossible; new S7 characters unknown to any model; no franchise control | | Off-the-shelf animation tool | No AI motion generation; labour-intensive; not scalable | | Full model training from scratch | Weeks of compute, hundreds of thousands of dollars; unnecessary — base models already know motion | | Licensing a competitor's animation AI | No franchise specificity; still confuses Sam and Clover; still needs the same fine-tuning | | Fine-tuning without a bible | Generic inaccurate captions → lower output quality → brand errors | | FLUX [dev] variants | ❌ Non-commercial license on the weights — commercial use requires a paid BFL agreement | | LTX-Video (Lightricks) | ❌ $10M revenue threshold in the license — Banijay triggers it; double-damage penalty for breach | | HunyuanVideo (Tencent) | ❌ License explicitly excludes EU, UK, South Korea — cannot be used legally in France | --- ## 7. Model choice — why Wan2.2 - **Only two models** are commercially clean for a French production, with no revenue threshold, no geographic restriction, and open fine-tuneable weights: **Wan2.1** and **Wan2.2** - Every other candidate has a verified legal blocker: - **FLUX.1 [dev] and FLUX.2 [dev]** — non-commercial license on the weights. Direct quote: *"use for revenue-generating activity is NOT a Non-Commercial Purpose."* Commercial use requires a paid BFL agreement. - **LTX-Video (Lightricks)** — $10M annual revenue threshold in the license. Banijay at ~€3B/year triggers it. Penalty for breach: double damages. Cannot use without a Lightricks commercial agreement. - **HunyuanVideo (Tencent)** — license directly states: *"this agreement does not apply in the European Union, United Kingdom and South Korea."* Using it in France is unauthorized. Full stop. - **Wan2.7** — no open weights. API-only via Atlas Cloud. Cannot be fine-tuned by anyone. - **Wan2.2 over Wan2.1** because it is a strict upgrade: - MoE architecture: more model capacity at the same compute cost - +65% more image training data → better stills quality - +83% more video training data → better motion quality - Cinematic-level aesthetic training with lighting, composition, colour tone labels - Same Apache 2.0 license, same fine-tuning approach, drop-in replacement - **Wan2.2 is a genuine hybrid model** — same weights produce both images and video - `TI2V-5B`: Text-and-Image-to-Video — images now, video later, one training run - `T2V-A14B`: higher quality video generation (MoE 14B) - `I2V-A14B`: animate from a reference frame or storyboard --- ## 8. Budget fallback — image-only path **If video budget is cut, the image path costs nothing extra to enable — the data work is already done.** - Use Wan2.2's image mode: same fine-tuned LoRA, generate stills instead of clips - Or use **FLUX.1 [schnell]** (Apache 2.0, no caveats) as a dedicated image model - Note: [schnell] is the *only* FLUX variant that is commercially clean — all [dev] variants are non-commercial - Lower quality ceiling than [dev] but fully usable - Training data is identical for both paths: 6,645 frames + 1,551 VLM captions already built - **Wan2.2 image mode is the better budget fallback** — one model, one training run, upgrade to video later without starting over ### What image-only produces - Key art and promotional stills - Character reference sheets — any character, outfit, location - Social media posts (static) - Storyboard frames for human animators - Episode thumbnails, product / merchandise mockups ### What it doesn't produce - Motion / video — Reels, TikTok, YouTube Shorts, animated promos ### The upgrade path > *Fine-tune Wan2.2 once. Use image mode now. Switch to video mode when budget allows. Same weights, no restart.* --- ## 9. The two-tier offer | | Image tier | Video tier | |---|---|---| | Model | Wan2.2 (image mode) | Wan2.2 (video mode) | | Training | Same LoRA fine-tune | Same LoRA fine-tune | | Training time | ~4–6 h (image data only) | ~8–12 h (clips + images) | | Output | Stills, key art, social images | Clips, promos, animated content | | Data prep | ✅ Done | ✅ Done | | Licensed files needed | Yes | Yes | | License | Apache 2.0 — no caveats | Apache 2.0 — no caveats | | Upgrade path | → video mode when ready | — | --- ## 10. Key numbers - **13 episodes** analysed (S7 first half, 14–26 behind paywall) - **2,852 shots** catalogued - **1,551 training clips** — 106 minutes at 720p - **6,645 frames** extracted for image training - **29,731 words** transcribed, 51% speaker-attributed - **1,551 / 1,551 VLM captions** — 100%, wiki-anchored - **~30 gadgets** named by exact wiki name - **18 canonical locations** defined - **1 model** targeted: Wan2.2 (Apache 2.0) - **LoRA training**: ~8–12 h on one GPU once licensed material arrives - **Full licensing research**: `docs/research/model-licensing.md`