> **⚠️ SUPERSEDED** — This document is a historical receipt. See `s7-current-understanding.md` for the authoritative current position. # Totally Spies S7 animation style vs. video diffusion model relevance ## Purpose This note qualifies whether the architectural differences between image diffusion and video diffusion models are material for the Totally Spies marketing video pipeline. The analysis was performed using local Ollama `qwen2.5vl:7b` vision inference against actual Season 7 frames from the official Banijay trailer, combined with published video diffusion architecture research. ## Production provenance Season 7 (26 × 22′) is produced by Zodiak Kids & Family (part of Banijay Kids & Family) with animation by Ollenom Studio. The Totally Spies Wiki and Animation Magazine both confirm the series is animated using **Toon Boom Harmony**, which is the industry-standard software for digital cutout / puppet- rig 2D animation. The show uses Harmony's cutout animation workflow: character models are broken down into separate drawing layers for each movable part (head, torso, arms, legs, hands, mouth shapes), arranged in a peg-based hierarchy, and animated by setting keyframes on transforms (position, rotation, scale) with Harmony interpolating the in-betweens automatically. ## Sources ### Frame sources - Scene frames: `materials/benchmark/official-vimeo-trailer/scene_frames/` - Motion analysis sequences: extracted at native framerate from the official S7 trailer at ~20s (action) and ~40s (dialogue/hold), 12 consecutive frames each - Raw Qwen2.5-VL results: `materials/benchmark/official-vimeo-trailer/ts-vision-results.json` ### Architecture research - LTX-Video (arXiv:2501.00103) — 3D VAE with 32×32×8 spatiotemporal compression - Wan2.1 (arXiv:2503.20314) — Wan-VAE with unified spatiotemporal latent space - IV-VAE (arXiv:2411.06449, CVPR 2025) — split keyframe + temporal branch architecture - Stable Video Diffusion (arXiv:2311.15127) — temporal convolution and attention layers inserted after spatial layers - Lumiere (arXiv:2401.12945) — Space-Time UNet for full-duration generation - Sora technical report (OpenAI) — spacetime patch tokenization + full self-attention - Video LDM (arXiv:2304.08818) — temporal layer insertion into pretrained 2D UNet - Empirical survey of temporal layer design (arXiv:2502.07001) ## What Qwen2.5-VL sees in the S7 frames ### Visual style (6 diverse scene frames) | Property | Finding | |---|---| | Line art | Clean vector lines, uniform stroke weight, crisp geometric edges. No hand-drawn wobble or variable line texture. Consistent with Toon Boom Harmony vector output. | | Coloring | Flat color fills (the traditional cel-painting look). Solid uniform areas with minimal gradient use (skies, water only). No painted textures, no soft airbrush rendering. Note: this is not "cel shading" in the 3D NPR sense (toon shaders on 3D geometry). The show is native 2D; the flat fill approach is inherited from classical cel-painting conventions applied in a digital ink-and-paint pipeline. | | Character rendering | Digital cutout rigs (Toon Boom Harmony puppet-style). Character models are broken into hierarchical drawing layers per body part, animated via transform keyframes with software-interpolated in-betweens. This is the standard Harmony cutout workflow, distinct from traditional frame-by-frame hand-drawn animation. Consistent limb proportions and clean joint articulation across the entire season. | | Backgrounds | Flat vector or lightly gradient-shaded. Static across all motion pairs. Treated as reusable environment plates, not per-frame painted artwork. | ### Motion behavior #### Seq A — action sequence (~20s mark, 12 frames at 25 fps) Frames compared in pairs spaced 3 frames apart (~120 ms gap). - Characters shift stance between frames with all body parts moving in coordinated synchrony. The vision model interpreted this as the whole figure being "repositioned" or "redrawn" as a unit. In practice, Harmony cutout rigs animate each body-part layer on its own interpolation curve, but when the animator keys multiple parts on the same frames, the rendered output appears as a coordinated whole-body shift. The vision model cannot distinguish "rig transforms applied" from "character fully redrawn" based on rendered pixels alone. - Background is **completely static** across all pairs. - Movement is coordinated but simplified — low drawing count between key poses, consistent with cutout-rig interpolation rather than frame-by-frame redraw. - Vision model identified this as **limited animation** in 2 of 3 comparisons, full animation in 1. The "full animation" classification in one pair likely reflects the vision model's inability to distinguish cutout-rig repositioning from traditional full redraw. #### Seq B — dialogue / hold sequence (~40s mark, 12 frames at 25 fps) - **Near-identical frames** across all 3 comparison pairs. - Minimal changes: subtle expression shifts, minor hand and posture adjustments. - Background, limbs, and faces **held completely still**. - All 3 pairs identified as **limited animation** with emphasis on expression over motion. ### Production technique summary - **Digital cutout animation** pipeline built in Toon Boom Harmony. Characters are rigged as cutout puppets with hierarchical peg structures; motion is created by setting keyframes on transforms and letting Harmony interpolate the in-betweens (often called "tweens" in this context, equivalent to the traditional "inbetweening" process but automated by software). - **Limited animation**: heavy use of held drawings (also called "holds" or "extended exposures" — a single drawing is displayed across many consecutive frames). The show frequently animates on threes or fours during dialogue and holds, switching to twos during action beats. At the show's 25 fps PAL frame rate, on twos yields ~12.5 unique drawings per second; on threes yields ~8.3; on fours yields ~6.25. This is consistent with standard TV-budget limited animation practice since Hanna-Barbera. - **Modular compositing structure**: static background plates layered separately from animated character layers. Background art does not change frame to frame within a shot. This layer separation is fundamental to the cutout pipeline and visually obvious in the frame analysis. ### Methodological limitations - Qwen2.5-VL analyzes rendered composite frames and **cannot see the underlying rig structure**. It cannot distinguish between a cutout rig whose parts were transformed (Harmony's actual workflow) and a character that was fully redrawn frame to frame. All conclusions about "limited" vs "full" animation are inferred from pixel-level similarity between frames, not from production metadata. - The Qwen2.5-VL "production technique summary" (analyzing three still frames in isolation) characterized the show as "high-framerate fluid" with "no visible holds" — **directly contradicting** the motion-pair analysis where all sequence B pairs showed near-identical held frames. This discrepancy demonstrates that single-frame style assessment is unreliable for temporal behavior; the motion-pair comparisons are the more trustworthy signal. - Frames were extracted from H.264-compressed video. Compression artifacts mean that even truly held drawings will not be literally pixel-identical in the decoded frames. ## Relevance to video diffusion architecture ### Why temporal architecture matters for S7 The prior analysis established that video diffusion models (LTX-Video, Wan2.1, Sora) use 3D spatiotemporal latent spaces where temporal relationships are baked into the representation — fundamentally different from generating frames independently. For Totally Spies S7, this architecture is well-matched, but for a counterintuitive reason: 1. **The temporal patterns are simple and learnable.** S7 uses limited animation: most frames are near-identical holds with occasional coordinated pose shifts. A video model's temporal compression (e.g., LTX-Video's 8-frame temporal tokens) naturally captures this — hold frames compress to near-zero temporal variation in latent space, and the model learns that this is the expected behavior. It does not need complex physics simulation. It needs to learn that flat-colored shapes hold still for extended exposures and then transition via smooth software-interpolated in-betweens. 2. **The cutout-rig motion language is structurally regular.** Unlike traditional full animation (animating "on ones" with unique drawings per frame, as in classic Disney features), S7's character motion follows the predictable patterns of digital cutout animation: individual body-part layers rotate or translate at their peg pivot points, figures slide on static backgrounds, and expression changes happen via drawing substitution in the mouth or eye layer while the rest of the body holds. This regularity — geometric transforms on flat vector shapes — is exactly what temporal attention layers can capture efficiently. 3. **The risk is misfire, not inability.** A general-purpose video model trained on live-action and 3D-rendered footage will produce excessive inter-frame variation by default — adding physics-driven secondary motion (hair sway, cloth dynamics), parallax shifts, and sub-pixel noise where the show deliberately uses flatness and extended holds. The custom pipeline needs to teach the model that stillness is correct and that the animation density (drawings per second) is intentionally low — the show operates on threes, fours, or higher for much of its runtime. 4. **Frame-by-frame image generation would fail here** — not because the motion is complex, but because it would introduce per-frame stochastic variation (flickering line weights, shifting color fills, drifting proportions) that S7's style does not exhibit. The show's visual language demands that consecutive held frames be visually identical, and smoothly interpolated between key poses during movement. Only a model with temporal awareness can enforce this level of inter-frame consistency. (Note: "identical" here means derived from the same held drawing in the production pipeline. In the final H.264-encoded video, minor compression artifacts may introduce sub-pixel differences even within true holds.) ## Summary assessment | Factor | Assessment | |---|---| | Does S7's style need a custom video model? | Yes — but to learn restraint (limited animation, extended holds, smooth cutout interpolation), not physical complexity | | Is temporal architecture relevant? | Yes — hold-frame consistency (extended exposure) and smooth interpolated transitions are temporal phenomena that require cross-frame awareness | | Would image-only generation plus stitching work? | No — per-frame variation would break the show's characteristic stillness | | What is the training signal? | The model needs to learn: static backgrounds, cutout-rig pivot-based motion, deliberate extended holds, flat color fills, and modular composited shot structure | | Primary risk | Excessive inter-frame variation. A pretrained video model's prior toward fluid realistic motion will add secondary dynamics, sub-pixel drift, and frame-to-frame noise that S7 deliberately avoids. The LoRA / fine-tune must suppress this bias toward high drawing counts and continuous motion. | ## Terminology note This document uses animation industry terminology as follows: - **Limited animation**: an animation process that reuses drawings, uses fewer unique drawings per second, and relies on held cels / extended exposures. Distinct from full animation where every frame is uniquely drawn. Standard reference: Wikipedia "Limited animation"; originated with Hanna-Barbera in the 1950s. - **Cutout animation** / **puppet rig**: Toon Boom Harmony's standard character animation workflow where models are broken into separate drawing layers and animated via transform keyframes. Officially documented by Toon Boom as "cut-out animation." Not the same as physical paper cutout (Terry Gilliam style). - **Hold** / **extended exposure**: a single drawing displayed across multiple consecutive frames. In dope-sheet / x-sheet terminology, this is "extending the exposure" of a drawing. - **On ones / on twos / on threes**: industry-standard terms for how many frames each unique drawing is held. "On twos" = each drawing shown for 2 frames = 12 unique drawings per second at 24 fps (or 12.5 at S7's 25 fps PAL rate). S7 frequently operates on threes or fours during dialogue. - **Tween / in-between**: in digital cutout animation, the software- interpolated frames between keyframes. In traditional animation, "inbetweening" is the manual drawing process; in Harmony, it is automated. - **Flat color fill**: solid uniform color areas without gradient, as in classical cel painting. This is not "cel shading" in the modern 3D NPR sense (toon shaders applied to 3D geometry). - **Drawing count** (our shorthand: "animation density"): the number of unique drawings or significant pose changes per second. This is not a standard industry term; the industry describes this via "on twos / on threes" notation or simply as the drawing count per second. We use "animation density" as informal shorthand in this document. ## Open questions - How many training clips of deliberate hold sequences (extended exposures) are needed to teach the model that near-zero inter-frame variation is the expected default? - Does the temporal compression ratio of the chosen video VAE align well with S7's typical hold duration (often 8–12+ frames at 25 fps, i.e., animating on fours to sixes during dialogue holds)? - Can LoRA training on a limited-animation dataset effectively suppress the base model's bias toward high animation density, or does the base model's prior toward fluid motion overwhelm the fine-tune?