The thesis

One paragraph

A video model can produce motion, light and life that no renderer can. It cannot be told who is on screen and where they are standing. armature stages a canonical character mesh in headless Blender and turns the render into a per-frame control sequence the video model must obey. You block the shot; the model shoots it. Structure comes from geometry we own; life comes from the model; identity is a named, versioned thing that rides in the prompt and the reference stack, never an accident of a lucky frame. The deliverable is footage — film, cutscenes, character performance, anything image-to-video can make, made from a scene you own. A game is one consumer, not the boundary.

That thesis is now measured. The mechanism first — as of 2026-08-11, across 22 generations in five closed experiments:

• a rendered control sequence governs where the figure is, at what scale, and when it moves;
• it governs authored subject motion — an 85.0° arm sweep against 0.062° when the same control is held still;
• control owns the outline and the reference owns surface, material and costume, with the reference able to extend a silhouette only where the control is silent;
• and inverting the depth control end-to-end does not break tracking, so the model is reading geometry rather than tone.

The two questions that were open then have since been answered at product level. The rigged character mesh exists — E07 rebuilt the rig and the Director approved the skeleton by eye — and identity has held through paint at the judging eye three times: driven from the rig (E08), unanchored with no reference and no driving signal at all (E11 wave 1), and through the camera tier, where a handed world also holds to the last frame on two seeds (E12). What remains open is route-shaped, not thesis-shaped — and E13 answered it on 2026-08-13, its whole arc (dispatch, a zero-spend halt on two structural premise failures, a repair arc that rebuilt the reference kit with true alpha, re-arm, four generations) inside one date: identity survives a hosted, human-trained tier fed nothing but authored references, both arms, both seeds, at the Director’s eye — and the reference’s own ground steers the model-decided world. Full detail in docs/experiments/; every claim above is traceable to a numbered ruling.

What follows is the published evidence that made the thesis worth testing, gathered by a founding research swarm on 2026-08-10 and recorded in docs/research-grounding.md. Findings are cited by their number there.

How much to trust each finding

The grounding pass has a stated ceiling, and it matters more than any individual number.

Existence was checked for all 34 arXiv identifiers by a deterministic retrieval oracle rather than by model memory: 34 of 34 resolved, zero fabricated. First authors and years were checked against the oracle’s returned metadata, and one attribution error was caught and corrected that way (F9 was reported as 2026; the oracle returns 2025-04).

Groundedness — whether a paper’s abstract actually supports the claim attributed to it — was run by two decorrelated non-Claude model families, reasoning-stripped, on the two claims that decide the first experiment’s architecture. Not on all 34. Findings marked unverified-groundedness below have confirmed existence and correct attribution but no second check; they may inform a design and may not carry it alone.

That distinction did real work immediately. F2 was downgraded from load-bearing to diagnostic because both checkers returned NOT_IN_ABSTRACT — its numbers live in a table, not the abstract, and the table has not been retrieved. It is not used as a design constraint here.

The core claim: dense guidance rendered from a body model beats a 2D skeleton

F1 — the only finding confirmed by both groundedness checkers. Champ (Zhu et al. 2024, arXiv:2403.14781) compared dense 3D-parametric guidance — depth, normal and semantic maps rendered from a posed parametric body — against a 2D skeleton alone:

	Skeleton-only	Dense 3D-parametric
FVD (lower better)	192.34	170.20
SSIM (higher better)	0.672	0.773
LPIPS (lower better)	0.296	0.235

And the ablation that matters most: dropping the skeleton entirely still beat skeleton-only on FVD, at 184.24. The skeleton is a small additive term, not the backbone.

This is armature’s thesis, measured by someone else. It also names a starting channel set — depth, normal and semantic, with a skeleton optional — rather than leaving it to taste.

F21, from the same paper, is the closest published precedent to armature’s exact move: Champ renders those maps from a 3D parametric body and explicitly bypasses video-based pose detection. No 2D estimator in the loop. That is the same architectural decision, taken for the same reason, and it is why armature is not a wrapper around a pose extractor.

What none of this establishes: Champ’s subject matter is not stylized game characters, and its generator is not the one armature will probe. Whether the result transfers is the question the repo’s first generation experiment exists to ask, not something this page can answer.

Where the risk actually sits

F6 (unverified-groundedness). VACE (Jiang et al. 2025, arXiv:2503.07598) accepts depth, pose, scribble, gray, layout, flow, spatiotemporal masks and reference images, and beats task-specific baselines on depth and pose control — but loses reference-to-video to Vidu 2.0 (3.40 vs 3.84).

Read the two halves together: structure control is the solved leg; identity is the weak one. That is armature’s risk profile precisely, and it is why identity gets phases of its own rather than a paragraph.

F8 (unverified-groundedness). Even the strongest current reference-to-video methods reach only modest face similarity — FaceSim-Arc 0.571 for the proposed method in Xu et al. 2026 (arXiv:2607.20247), with VACE-14B at 0.531 and Phantom-14B at 0.495. Visible identity drift is the default, not the exception. Any prediction written here starts from that.

F13 (unverified-groundedness). Multi-shot benchmarks measure character-recurrence gaps across as many as 48 shots and report that naive multi-shot systems introduce extra characters, with semantic drift and identity inconsistency (He et al. 2026, arXiv:2605.15199; Zhang et al. 2025, arXiv:2512.12372). Identity is not free across a cut. It needs an explicit mechanism, and finding out which one is a phase in the arc.

F10 (unverified-groundedness). Reference sets spanning diverse subject orientations produced substantially stronger identity coherence than a single canonical view (Zeng et al. 2026, arXiv:2604.07823). If that holds here, the right input is a full turnaround rather than one hero frame — and a turnaround is exactly what the sibling tool upstream already produces. Whether it holds here is a measurement nobody has taken.

F5 (unverified-groundedness). Control-scale curves are non-monotonic: past a threshold, dynamics decrease as motion degenerates into abrupt cuts. Rotation is far harder to control than translation — leakage 1.08 versus 6.04 (Hou & Rupprecht 2026, arXiv:2605.14815). The uncomfortable implication: an orbiting camera is the obvious first shot to try and may be the hardest regime there is.

F4 (unverified-groundedness). Control held at full strength is a named failure mode, not a theoretical one — a ControlNet scale pinned at 1.0 produces a walk described as rigid and translational rather than a natural gait; annealing the scale over early denoising steps restores it (Zhou et al. 2026, arXiv:2603.15614). Whatever the right control strength turns out to be, it is likely a schedule rather than a constant.

Why a shot, and not a continuous roll

F14 (unverified-groundedness). Open-weight native lengths are short and set by the training horizon — HunyuanVideo 129 frames, Wan 2.2 optimal at or below 120 frames, LTX-2 10 seconds standard. The shot is already the native unit. Staging per shot matches the engines rather than fighting them.

F15 (unverified-groundedness). This is the strongest quantitative support for armature’s overall shape. Anchor-bounded generation measurably suppresses drift: plain autoregressive rollout swings 5.94–6.54 VBench Imaging Quality points at each chunk boundary versus 1.23–1.50 for anchor-bounded, and scores 6.2–6.6 points higher in aggregate over a thirty-minute horizon (Bendel et al. 2026, arXiv:2605.20476). A shot bounded by fixed staged state is an easier problem than free-running continuation.

F16 and F18 (unverified-groundedness). Autoregressive degradation has a formal account — cumulative per-step error, with more conditioning frames monotonically reducing forgetting without removing degradation. And every long-form system found decomposes into shots or keyframe anchors plus assembly; none is one continuous roll. Shots plus cuts is the working grammar, which is convenient, because shots plus cuts is also how films are made.

Format is not a detail

Getting the output conventions wrong would silently invalidate every later experiment, so they are grounded rather than guessed.

F19 (unverified-groundedness). ControlNet-family depth conditioning is inverse relative depth — near is bright — normalized per frame, and hard-capped at 8-bit because the conditioning image is an RGB PNG with R=G=B. A proposal to pack more precision across channels was never adopted upstream. Sixteen-bit buys these consumers nothing.

F20 — verified by direct source retrieval, not by model memory. The OpenPose-18 skeleton is fully code-specified: 18 keypoints, a fixed 19-pair limb sequence, and an 18-colour per-limb palette. The pair list was fetched and read at ruling time, and it is 1-indexed — a live trap for anyone writing a renderer from scratch, along with the fact that the target is COCO-18, not COCO-17 and not Body25.

F24 (unverified-groundedness). Wan and VACE require width and height divisible by 16 and frame counts of the form 4n+1. A frame size that is not generator-legal breaks every downstream pairing, and it fails quietly — which is the worst way for a constraint to fail.

F22 (unverified-groundedness). A rendered geometric edge pass sidesteps Canny’s per-image dual-threshold tuning, which does not generalize across shots and is contrast-sensitive. So rendered edges should be more temporally consistent than Canny run over rendered RGB — stated here as a testable claim, not an assumption.

The gaps, named so silence is not mistaken for evidence

The research swarm was explicit about what it could not answer:

• No quantitative strength-versus-identity-drift curve exists anywhere. There is no prior to predict against; that curve has to be measured here.
• No clean single-architecture ablation isolating the number of reference views against an identity metric. The closest (F10) comes from a different application.
• No study compares per-character adapters against zero-shot conditioning on stylized game art specifically — every quantified comparison found used photoreal subjects.
• No head-to-head of depth versus pose versus segmentation on a single identity metric.
• No controlled study isolating bounded versus unbounded drift on the same model.
• No primary technical reports for Kling, Seedance or Veo 3 — closed systems, no papers.

One thing no metric will decide

Whether the figure on screen is the same character is canon, and no metric approximates it. The repo learned this secondhand from facet, which established it twice — once with high-pass statistics for material identity, once with silhouette IoU for character identity.

There is a sharper version for this domain. F12 (unverified-groundedness): standard face-recognition embeddings hold roughly 0.76–0.78 true-positive rate in-distribution but collapse to 0.372 on unseen stylization, mistaking a style shift for an identity change (Yun et al. 2026, arXiv:2604.21689), corroborated by a 5,013-identity cartoon benchmark showing photo-trained recognizers transfer poorly (Zheng et al. 2019, arXiv:1907.13394).

So ArcFace-family metrics cannot gate identity on stylized game art, and a diagnostic that returns a number on a face its detector never found is noise wearing a unit. Identity and quality diagnostics ride reports as diagnostics and gate nothing. The Director’s eye is the judge.