# Totally Spies S7 — model selection and training priorities ## Purpose This memo turns the benchmark-pack work into practical guidance for: 1. selecting future generation models / pipelines 2. prioritizing what to train or fine-tune first 3. deciding what to evaluate before claiming the system matches S7 style ## Evidence base This memo is based on: - downloaded official S7 corpus: - `materials/reference/totally-spies/youtube-official/inventory.json` - expanded corpus validation: - `docs/research/s7-animation-validation-expanded.md` - `materials/benchmark/youtube-s7-validation/expanded-validation-qwen2.5vl-7b.json` - benchmark packs: - `docs/research/s7-benchmark-packs.md` - earlier second-opinion pass: - `docs/research/s7-animation-style-second-opinion-gemma3-12b.md` - `materials/benchmark/official-vimeo-trailer/ts-vision-second-opinion-gemma3-12b.json` - pack-level evaluation on `qwen2.5vl:7b`: - `materials/benchmark/youtube-s7-validation/packs-eval-qwen2.5vl-7b.json` - pack-level evaluation on `gemma3:12b`: - `materials/benchmark/youtube-s7-validation/packs-eval-gemma3-12b.json` - `docs/research/s7-gemma-pack-confirmation.md` - pack-level evaluation on `gemma4:26b`: - `materials/benchmark/youtube-s7-validation/packs-eval-gemma4-26b-chat.json` - `docs/research/s7-gemma4-pack-confirmation.md` ## What the evidence consistently says Across trailer-only analysis, expanded corpus analysis, the benchmark packs, and the Gemma 3 second-opinion pass, the stable conclusion is: - S7 reads as **digital 2D cutout / rigged animation** - the visuals are **clean, vector-like, flat-filled, and production-consistent** - the motion discipline is **restrained**, with many held or near-held beats - even action remains **controlled**, not densely redrawn in a full-animation sense - therefore, the primary AI-video problem is **temporal consistency under a restrained cutout-style motion regime**, not realistic motion physics ## What matters most for model selection ### 1. The best model is not the one that makes the prettiest still For this project, still-image richness is not enough. A candidate model or pipeline should be preferred if it can do these things well: - keep faces stable across frames - keep costume details stable across frames - keep linework / edge behavior stable across frames - avoid adding unnecessary secondary motion - avoid adding texture / shading creep - preserve static or near-static backgrounds in held beats - keep gadgets readable and attached consistently during action ### 2. Temporal restraint matters more than realistic motion sophistication A strong candidate model for S7 should be biased toward: - low inter-frame drift - strong identity persistence - stable flat-colored surfaces - coherent motion during small pose changes - consistent cutout-like geometry under motion A weaker candidate model will typically fail by: - over-animating holds - adding body sway, hair physics, cloth dynamics, or camera drift - changing faces between frames - changing costume shape / color between frames - softening the flat 2D look into something more generic or more painterly - introducing **texture boiling / shimmering** on static lines and surfaces - making supposedly static backgrounds **breathe** or drift - causing **silhouette degradation** under motion or lighting change ### 3. Motion-sheet performance matters more than still-sheet performance One recurring lesson from the validation passes: - still-only sheets can cause a model to over-read the style as more fluid or more elaborate than it really is - motion sheets with sequential frames are much better at revealing whether a model understands the actual animation discipline So when comparing future candidate models, the ranking should put more weight on: - `dialogue-hold` motion-sheet performance - `action-gadget` motion-sheet performance than on single-frame aesthetic judgment alone. ## Training priorities ### Priority 1 — character / costume consistency Pack: - `materials/benchmark/youtube-s7-validation/packs/character-costume-consistency/` Why first: - if the characters are unstable, nothing else matters - this is the highest likely approval risk - identity drift is the fastest way to lose franchise trust What to optimize: - face consistency - silhouette consistency - costume / wardrobe stability - recurring supporting cast stability - trio coherence in shared shots Success looks like: - Sam looks like Sam across contexts - Clover looks like Clover across contexts - Alex looks like Alex across contexts - wardrobe and hero-mission outfits do not mutate between shots ### Priority 2 — dialogue / hold temporal consistency Pack: - `materials/benchmark/youtube-s7-validation/packs/dialogue-hold/` Why second: - this is where generic AI video most obviously breaks the S7 style - low-motion scenes are supposed to feel stable, not busy - if the model cannot hold still, it will not feel like the show What to optimize: - low inter-frame drift - stable heads / torsos during held beats - small controlled facial or mouth changes - static or near-static backgrounds - no invented secondary motion - no **texture boiling** on lines, collars, or facial features - no **background breathing** in screens, walls, consoles, or cockpit elements - no **motion bleed** where the system invents movement inside a supposed hold Success looks like: - held beats remain visually stable - changes happen only where intended - no flicker or breathing artifacts in the body / face - props and screens remain locked unless the shot explicitly animates them ### Priority 3 — action / gadget control Pack: - `materials/benchmark/youtube-s7-validation/packs/action-gadget/` Why third: - action is important, but the show's action is still controlled rather than hyper-fluid - gadget readability is commercially important for promo beats - this is where many models will add too much motion or lose attachment continuity What to optimize: - gadget readability - hand / gadget continuity - controlled motion under higher energy - no transition from crisp 2D cutout feel into generic AI action mush - restrained action timing rather than full-physics spectacle - no **gadget melting** into hands, sleeves, or nearby effects - no **color bleeding** across costume boundaries during motion - no edge-smear that destroys the cutout / layered look Success looks like: - gadgets stay recognizable and legible - action beats remain sparse and stylized - character identity remains stable under movement pressure - costume edges, gadget geometry, and silhouette anchors stay readable even in high-energy poses ## Recommended evaluation order for future generation tests When testing any future candidate generation workflow, evaluate in this order: 1. **character-costume-consistency** 2. **dialogue-hold** 3. **action-gadget** Reason: - if identity fails, stop - if holds fail, stop - only then test action pressure ## Practical scoring rubric Use a simple gate before any deeper investment: ### Gate 1 — identity Reject if: - faces drift across adjacent frames - costume details mutate - trio coherence breaks in shared shots - silhouettes degrade across angle / lighting changes ### Gate 2 — holds Reject if: - held shots shimmer, breathe, or jitter - backgrounds drift during supposed holds - facial changes propagate unnecessarily into the whole body - static linework shows texture boiling or crawling ### Gate 3 — action / gadgets Reject if: - gadgets become unreadable - motion becomes too fluid / realistic for S7 style - character shapes deform unpredictably under action - props melt into hands or costume edges bleed together ## Current recommendation ### Best current selection principle For this project, prefer models / pipelines that optimize for: - **temporal consistency** - **identity persistence** - **flat 2D cutout-style stability** - **restrained motion discipline** rather than models that simply score well on general video realism. ### Current local-evaluation takeaway Based on the available local validation runs: - `qwen2.5vl:7b` is good enough to identify the key failure modes when shown motion-oriented benchmark material - `gemma3:12b` confirmed the broader conclusion that S7 is a restrained cutout-style system and that temporal consistency matters more than realistic physics - `gemma4:26b` confirmed the same ordering and sharpened the failure-mode language around texture boiling, background breathing, gadget melting, color bleeding, and silhouette degradation - but none of these analysis models should be confused with the eventual generation model choice; they are evaluation aids ## Bottom line If we want a generation system that feels like Totally Spies S7, the training and selection priorities should be: 1. **lock identity** 2. **lock holds** 3. **lock controlled action and gadget readability** That is the shortest path to validating the earlier conclusion in operational terms.