User wants to generate a short animated 2D anime cartoon, music video, or animated scene with consistent characters, VLM verification, and narrative continuity.
A professional-grade 2D anime cartoon generator with VLM-verified assets,
video analysis via Gemini 3.1 Pro, sequential keyframes for continuity,
multi-model image/video generation, LLM-powered error recovery, and a
parallel asset + music pipeline — now with explicit quality modes, a
self-contained budget rail, honest partial delivery, and a two-VLM identity
judge panel used as a hard gate.
What's New in v3.0.0
1. Three quality modes
Mode
Video candidates
Continuity policy
Continuity regens
Judges
Scene cap
Est. USD / scene
low
1
off
0
0
4
$3.20 (extrapolated)
medium (default)
2
adjacent
1
1
8
$7.39 — MEASURED
max
3
all_recheck
2
2
24
$13.00 (extrapolated)
Cost honesty: medium is the ONLY measured figure — $7.39/scene, derived from
exactly one 2-scene run that cost $14.77. The low and max numbers are
extrapolations from that single data point (fewer/more candidates, judges, and
continuity passes), not facts. Treat every non-medium estimate as a planning
aid, never a billing statement.
2. Self-contained budget rail
This skill calls OpenRouter directly with a granted key, so core accounting
cannot see its provider spend — the host records the usage as
unknown/unmetered, and the core budget rail cannot protect a job. The skill
therefore carries its own rail:
A preflight estimate decomposes the job into unit costs (storyboard,
sheets, keyframes, video candidates, judges, cross-scene checks) before
anything is generated.
If the estimate exceeds the budget limit, the request is refused with
HTTP 402 naming the exact required amount — nothing generated, nothing
spent.
A running ledger of estimated spend is persisted into job.json
(progress.budget) as the job runs.
On limit breach, a hard stop stops starting new scenes while
preserving every asset already generated — the job assembles what it has
and ships partial, never discarding paid work.
3. partial status semantics
A job that finishes with missing or unverified scenes is still delivered — the
final video ships with the assembled scenes, and the job reports
status: "partial" with missing_scenes (indexes with no video clip),
unverified_scenes (indexes shipped without a passing verify), and
partial_reasons listed. A partial job never reports "Animation
complete!"; its final message names the delivered/expected scene counts, the
missing and unverified indexes, and the reasons.
The final-mix audio probe is now structural with three states: present,
absent, unknown. The previous substring-matching probe returned False on
ANY exception, and False routed the mix into the music-only branch — which
silently destroyed the scenes' native spoken dialogue. An unknown probe
now attempts the dialogue-preserving mix FIRST and falls back to music-only
only if that mix actually fails, with a warning either way.
Scene clips are gated by VLM judges drawn from DIFFERENT model families,
never the generator's family. The panel is two judges only at max; medium
runs ONE judge and low runs none — see "Correction: medium runs ONE judge"
below, which is the authoritative statement of this contract. A single judge
cannot deliver the cross-family disagreement signal, so at medium this gate is
a fail-closed single-judge check, not a panel. Every comparison runs in both
A/B orderings; an order-inconsistent verdict is downgraded to indeterminate.
Judges answer atomic per-attribute questions (hair side, accessory, eye
colour, character count) rather than an overall quality score, and
indeterminate is a first-class outcome that is never a pass — "could not
judge" cannot look like "judged and passed".
This hard VLM gate was the owner's explicit choice over objective metrics.
The published VLM-judge bias literature (position bias, self-preference bias,
verbosity bias) is exactly why these mitigations — family diversity, order
swap, atomic questions — are mandatory rather than optional.
identity_similarity_proxy is a Pillow colour-histogram cosine kept as a
diagnostic breadcrumb — it is NOT DINOv2, NOT calibrated, has no threshold,
and gates nothing.
Owner ruling (settled, not an open option): no local models. Identity is
judged ONLY by the OpenRouter VLM panel above. This payload carries no local
ML dependency — no torch, no local DINOv2, no local face/ReID embedding —
and must not acquire one; dependencies: is [Pillow] on purpose. A local
embedding cosine is therefore not a future upgrade path here, and no value in
this skill may be named identity_verified or used to gate a clip.
6. Raised caps: 24 scenes, 240 seconds
The global scene cap is now 24 and the duration cap 240 seconds. They had to
move together: scene count is bounded by duration_sec // 4, so a 20+
scene job was structurally unreachable while duration was capped at 60s —
raising the scene cap alone would have changed nothing.
7. Continuity regenerates on minor drift too
Cross-scene identity drift now triggers regeneration on minor verdicts as
well, not only major ones — minor drift is exactly how identity erosion
compounds across a chain of scenes. Regeneration is bounded by the per-mode
regeneration budget (0/1/2 for low/medium/max); once exhausted, the job ships
with a disclosed warning and partial status rather than silently.
Operational facts
These local edits are mortal. This payload lives under
data/skills/ouroboroshub/, and a marketplace update OVERWRITES it. Any
local fix that matters must be upstreamed to the hub or it will be lost on
the next install/update.
kill <pid> does NOT reload changed code. The generation worker is a
host-supervised companion process with a restart policy: killing it just
respawns the same process image, and an already-running worker keeps
executing the OLD code after payload edits. The skill must be reloaded
(disable + enable, or a host reconcile) for code changes to take effect. A
stale worker once held pre-revision code for 4.5 hours.
inject_chat is narrow by construction. The manifest declares the
permission ONLY for the structured PARTIAL/ERROR attention notification in
host_bridge.py: a fixed template with numeric/enumerated substitutions
(job id, mode, scene counts, whitelist-filtered reason keys), never
free-form content — not the theme, not the storyboard, not dialogue, not
warning strings. The skill functions fully without the grant; an ungranted
call is a logged 403 and generation proceeds unaffected.
Pipeline
Scenario Generation — LLM creates a structured storyboard with scene
transitions, dynamic action descriptions, and narrative continuity cues
Asset + Music Generation (parallel) — Character sheets (parallel), location
art (parallel), keyframes (SEQUENTIAL with inter-frame context), and music
all run concurrently via asyncio.gather
Image VLM Verification — Each critical asset (characters, keyframes) is
checked by Claude Sonnet for correctness and regenerated if needed (up to 2 retries)
Video Animation — Sequential scene generation with:
Full continuity chain (last frame → next scene reference)
ALL character sheets as references (not just first)
Explicit character identity descriptions in prompts
Negative prompting against text artifacts
Video VLM verification via Gemini 3.1 Pro (max 1 retry per scene)
Smart error recovery — LLM advisor recommends model switch on failure
Assembly — ffmpeg concatenation with audio mixing
Models Used (via OpenRouter)
Model
Purpose
openai/gpt-image-2
Character sheets, locations, keyframes (default, with fallback on transient failures)
google/gemini-3.1-flash-image-preview
Character sheets, locations, keyframes (nanobanana option and fallback)
anthropic/claude-sonnet-4.6
Scenario generation + image VLM verification
google/gemini-3.5-flash
Error advisor + prompt condensation — analyzes failures, condenses long prompts for models with character limits
google/gemini-3.1-pro-preview
Video VLM verification (multi-frame analysis)
bytedance/seedance-2.0
Video generation (default)
google/veo-3.1
Video generation (alternative — Google quality)
minimax/hailuo-2.3
Video generation (cheapest option)
kwaivgi/kling-v3.0-pro
Video generation (less strict copyright filter, auto-condensed prompts)
google/lyria-3-pro-preview
Original soundtrack clips
Requirements
OPENROUTER_API_KEY with access to the models above
Operator-provided ffmpegandffprobe on the host (video assembly +
frame extraction). Install via package manager (brew install ffmpeg,
apt install ffmpeg, etc.) or set FFMPEG_PATH / FFPROBE_PATH to
executable files. The skill does not download binaries.
v2.14.0 — the cross-scene identity gate actually gates
Diagnosed from job bb72b5fb, a 2-scene run whose delivered video visibly drifted
(a white hair streak swapped sides between scenes) even though the pipeline's own
reviewer had returned severity: "major". The gate existed — it was defeated by its
own bounds check.
A major verdict can no longer be silently discarded. The reviewer answered
with a "third scene" in a two-scene job, so 0 <= 2 < 2 was false and
_regenerate_single_scene was never called — no warning, no stat, nothing. An
out-of-contract index now falls back to the last compared scene and records a loud
cross_scene_index_fallback warning instead of cancelling the finding.
Frame indices map back to real scene indices. The reviewer only sees frames
that could be extracted, so its Nth answer meant "the Nth frame I was shown", not
"scene N". A scene whose clip was missing used to shift every later index silently.
The (scene, frame) pairing is now explicit and the prompt states the mapping and
the frame count, so the reviewer cannot name a scene it was never shown.
The reviewer's critique is fed back into the regeneration prompt. Re-running a
byte-identical prompt had no reason to drift less.
The repair is re-verified once. Identity is re-checked after regeneration and
the outcome recorded as cross_scene_resolved; a drift that survives is warned
about explicitly rather than shipped looking clean. It does not loop.
resolution is selectable everywhere. The agent tool, the widget form and this
manifest's form had no resolution field at all, so _build_settings fell to its
"720p" default and 1080p was reachable only by hand-crafting an HTTP body.
Scene cap raised 8 → 12. A request for 10 scenes was silently clamped to 8.
(Superseded in v3.0.0: the cap is now per-mode — 4 / 8 / 24 — with a global
ceiling of 24 scenes and 240 s. modes.py is the single source; this line is
kept only as history.)
v2.13.0 — quality fixes (visual oracle, character consistency, honest degradation)
Diagnosed from a real 45s run whose keyframes drifted (long flowing hair in scene 0
vs short-cropped in scene 1), whose own cross-scene check reported that drift, and
whose final video shipped 30s instead of 45s while a valid scene clip sat on disk.
Image generation moved to the canonical POST /api/v1/images endpoint
(api_client.generate_image_unified). The previous path used
/chat/completions, which cannot carry reference images at all. The legacy
path is retained as the last rung of the fallback ladder.
Reference-image conditioning between scenes. Each keyframe is now
conditioned on up to 4 references with reserved slots: the previous keyframe
(only when it PASSED verification), up to two character sheets for the
characters in the shot, then the location art. Character identity was
previously carried by prose alone, which drifts.
A rejected frame is never fed forward as a reference, so verification
failure cannot compound into the rest of the run.
The visual oracle keeps the BEST candidate, selected by a normalized 0-10
score (_normalize_vlm_image_result — score is canonical, passed is derived,
so a self-contradicting verdict resolves one way). It stays deliberately
non-blocking (a keyframe is a required input to the video chain), but an
all-rejected outcome is now stated loudly with its score and counted as
keyframe_rejected_used, instead of silently shipping the last retry.
Every model degradation is LOUD. A fallback to nanobanana (flash tier) or
to the legacy path reaches job.progress.warnings and increments
image_model_fallbacks. A parameter-shaped 4xx (too many references) retries
the SAME model with fewer references rather than dropping to a weaker one.
Provider capabilities are read live (/images/models, /videos/models)
once per job. Optional request parameters are omitted when the catalog does not
list them, and video durations are clamped from the live snapshot instead of a
hardcoded per-model table.
A generated clip is never lost to a later failure. Candidates are ranked
(pass beats fail, then the existing multidim weighted score, ties keep the
earlier attempt) and the best is recovered as clips_recovered.
Diagnostics name the exception TYPE (_describe_exc). Many transport
exceptions have an empty str(), which previously produced warnings like
Video scene 3 failed: — no cause, no type.
Story causality. The storyboard schema requires a causal_link per scene in
BECAUSE OF THAT / BUT form, and rules forbid an AND-THEN chain and require a
setup -> complication -> escalation -> payoff arc.
SKILL_VERSION is stamped into every job (verification_stats.skill_version),
so a companion worker still running pre-edit cached code is detectable.
Verified offline with a fake client against the real pipeline methods (25 checks:
reference forwarding, same-model reference trim, loud fallback, best-of selection,
forward-reference trust gate, capability-driven duration clamp). No paid
end-to-end run was launched.
v3.1 — reference conditioning actually reaches the model
The defect this release fixes
The owner reported that a character matched the character sheet in the first
shot and then looked generated from the text description in the second. The
reference images were already being SENT (input_references was populated), but
the prompt never mentioned them while explicitly instructing the model to
follow the text description. Nothing bound image to instruction, so the model
was free to treat the attachments as decoration.
Seedance multi-reference conditioning — what is documented vs measured
Documented by the provider (Seedance 2.0 API reference, docs.byteplus.com;
OpenRouter "Reference to Video" guide, openrouter.ai/docs/features/video-generation):
Fact
Source
Reference images are accepted, 1–9 of them
BytePlus Seedance 2.0 API reference (image_urls, 1–9)
The prompt MUST address them as @Image1, @Image2, …
BytePlus Seedance 2.0 API reference
Over OpenRouter the transport field is input_references (inputReferences also documented)
OpenRouter reference-to-video guide
Reference support is per-model, not universal
OpenRouter reference-to-video guide
Hard first/last-frame anchoring is a SEPARATE field (frame_images + frame_type) from soft references
OpenRouter video docs + live /videos/models
Established empirically here, because no documentation states it. OpenRouter
does not document a per-model maximum, an item schema for input_references, or
any signal that references were ignored:
input_references is Zod-validated: each item must carry image_url,
audio_url or video_url, otherwise the request is rejected before routing.
data: URLs are accepted as reference items.
A misspelled field is silently ignored. Sending image_urls (Seedance's
own native name) through OpenRouter produced a normal accepted job with the
references simply absent — no warning, no error. That silent-drop behaviour is
why the redacted request audit below exists.
@ImageN resolves to input_references submission order. Verified by a
paid colour-discrimination probe: references [solid RED, solid BLUE] with a
prompt commanding @Image2; the produced frame measured mean RGB
R=0 G=25 B=229 (blue). One-based, positional, and controllable.
OpenRouter accepted 12 reference items without complaint, so its validator is
NOT the enforcer of Seedance's documented 9 — the model is. The cap here is 9,
from the vendor documentation.
No provider mechanism reports that references were ignored. There is no
field, warning, or status for it. That absence is the reason this release logs a
structural self-check instead of trusting the provider.
How references are now assembled
Pipeline._assemble_scene_references is the SINGLE ordered source for both the
prompt labels and the input_references payload. Both are derived from the same
list after truncation, so the @ImageN number in the prompt cannot disagree with
submission order. Order is fixed: character sheets first, then the location
plate, then the previous approved frame — so the reference cap eats the
environment or the continuity anchor before it eats an identity anchor.
Labels are index-redundant on purpose (@Image2 — the CHARACTER SHEET of Mochi: …), so even if numbering were ever mis-resolved the label still says WHO the
image shows. Authority is scoped by kind: a location plate governs
environment only, a continuity frame governs appearance across the cut, and the
scene text still decides action, camera and dialogue.
Reference state is explicit, never a silent text fallback:
complete — every scene character had a sheet and nothing was dropped.
partial — something is missing or was dropped by the cap; the job records a
warning naming exactly what.
none — nothing reached the request; the job records
verification_stats["scenes_without_references"] and warns that identity is
unanchored for that shot. Generation still proceeds; the state is disclosed.
Request audit (host-verifiable, redacted)
Every /videos POST logs VIDEO_REQUEST_BODY and appends a redacted summary to
verification_stats["video_requests"] (a list, capped at 60, with
video_requests_omitted counting anything beyond — never a silent drop). The
audit sink is passed per call and fires BEFORE the POST, so a request that
raises still leaves evidence and retries cannot overwrite each other's record.
It carries counts, per-entry mime/byte-length/sha256 prefix, frame_type, and
prompt_names_all_references — a structural check that the prompt really named
every reference it attached. It deliberately carries no base64 payload and
no prompt text (the prompt contains owner-authored theme and dialogue).
Reference cap raised 4 → 9
VIDEO_MAX_INPUT_REFERENCES was 4. With characters ordered first, a three-character
scene consumed the whole budget and the cap silently discarded the cross-scene
continuity frame — the exact anchor that keeps identity stable between shots. It
is now 9, the vendor-documented Seedance maximum. A model accepting fewer replies
4xx and the existing reference-rejection path trims and retries the same model.
minimax/hailuo-3 added
From the live /videos/models catalog (not guessed):
minimax/hailuo-3
bytedance/seedance-2.0
catalog id
minimax/hailuo-3 ("MiniMax: H3")
bytedance/seedance-2.0
resolutions
["2K"] — 2K ONLY
480p / 720p / 1080p / 4K
durations
5–15s
4–15s
frame images
first_frame, last_frame
first_frame, last_frame
native audio
yes
yes
pricing SKUs
duration_seconds: $0.13, reference_images: $0.04
video_tokens: $0.000007
Does it support reference conditioning? Yes — and this is catalog evidence, not
an assumption: it publishes a reference_images price SKU, so references are a
billed, first-class input. It has NOT been exercised live here, so its
reference-following QUALITY relative to Seedance is unmeasured.
Is it cheaper? Yes, on published prices. A 10-second shot costs
10 × $0.13 = $1.30 plus $0.04 per reference image — roughly $1.46 with four
references, against a measured Seedance cost near $3.3 per candidate at 720p/10s.
About 2× cheaper.
Two honest caveats.
It is 2K-only. Selecting it with any other resolution is reconciled
upward to 2K with a warning by the existing capability gate; 2K was added to
that gate's ladder, without which this model raised "supports none of the
known resolutions" and was impossible to select at all.
Cost estimation is calibrated for Seedance's token pricing. MiniMax bills
per second plus per reference image, which this estimator does not model, so
its estimate for MiniMax is too HIGH. That errs toward stopping early rather
than overspending, but it is an estimate mismatch and not a measurement.
It is not made the default for scenes with characters: Seedance's
reference-following is the behaviour actually measured here.
Independent corroboration from the model pages (read 2026-08-04)
The catalog table above was derived from /videos/models. The two OpenRouter
model pages were then read separately and agree with it, which is why these
numbers are stated as facts rather than as a single source's claim.
minimax/hailuo-3 ("MiniMax: H3", released 2026-07-29) — its own FAQ:
"H3 generates video at 2K, supports 5–15 second clips, covers the 21:9, 16:9,
4:3, 1:1, 3:4 and 9:16 aspect ratios, can be steered with a first frame and
last frame image and produces a matching audio track."
That independently confirms 2K-only, the 5–15s window, first/last-frame
steering and native audio. Its overview also states "native audiovisual output
for reference-driven generation".
bytedance/seedance-2.0 — its own overview:
"It supports text-to-video, image-to-video with first and last frame control,
and multimodal reference-to-video. It is particularly strong at preserving
character consistency, visual style, and camera movement from reference
material. The number of tokens is given by
(height of output video * width of output video * duration * 24) / 1024"
Correction to the "cheaper" claim — the headline prices invert it. The pages
show Seedance "from $0.06726/second" against H3 "from $0.13/second", so
per advertised second H3 is the more expensive of the two. The comparison above
is still the right one, but only because it is stated at comparable quality:
Seedance bills by tokens that scale with height × width, so its "from" price is
the floor at its cheapest resolution, while H3's per-second price is flat and
already includes 2K. Against the measured Seedance cost near $3.3 per
720p/10s candidate, a 10-second H3 shot at 2K (10 × $0.13 plus $0.04 per
reference image ≈ $1.46) is cheaper at a higher resolution. Anyone comparing
the two headline $/second figures alone would reach the opposite conclusion,
which is exactly why this is written down.
Minimum duration is 5s, not 4s._clamp_duration falls back to the global
4–15s range only when the capability snapshot is empty; with H3 selected and an
empty snapshot a 4-second scene would be sent to a model whose documented floor
is 5. The snapshot normally supplies durations and clamps correctly, so this is
a narrow unknown-capability window, disclosed rather than fixed.
Correction: medium runs ONE judge, not a panel of two
Earlier documentation implied a two-family judge panel at medium. That is true
only at max. At medium the identity gate runs a single judge
(google/gemini-3.1-pro-preview), which is a diagnostic gate: it can still fail
a shot and still refuses to green an unanswered checklist, but it provides no
cross-family disagreement signal. Two judges from different families remain a
max feature. This is a documentation correction, not a behaviour change.
Identity checklist keys are now short ids
The judge checklist previously used the entire question text — which embedded
a character's whole visual_traits string — as the JSON key the model had to
reproduce verbatim, twice, for the forward and reversed pass to be matched. No
model reproduces hundreds of characters byte-identically, so coverage never
matched and every scene of the last live run returned indeterminate with an
empty attributes map. The gate could not pass by construction.
Keys are now short ordinal ids (char_01_hair, char_02_eyes,
character_count), collision-proof by construction rather than by slugifying
names that are frequently non-ASCII. Full descriptions stay in the question body.
Refusal remains fail-closed — partial coverage, order disagreement, a judge
error, or too few judges are all still indeterminate, never a pass — a green
verdict merely became reachable.
A character whose sheet did not reach the judge is not asked about (there
would be no ground truth) but its ids stay in the expected set, so an
unverifiable character forces indeterminate instead of a free pass.
v3.2: ONE authority for the native-audio flag
generate_audio was recomputed independently in the main scene loop and again in
continuity regeneration, while the preflight capability check only warned — so
a model reporting generate_audio=false still received generate_audio=true on
all four paid dispatch paths, and a gate later applied to one path would have
silently missed the others. _effective_generate_audio(job) is now the single
authority: the owner's dialogue setting is the intent, the live capability
snapshot is the veto, and a suppression is recorded as
verification_stats["native_audio_suppressed"] so a mute clip states its cause
instead of looking like a generation failure. An unknown model keeps the owner's
intent, because a permissive flag can at worst be ignored by the provider, while
a restrictive one would mute a model that does support native speech.
Known, not fixed in v3.1
LEASE_TTL_SEC (900s) is still shorter than TIMEOUT_VIDEO (1320s): a single
long video call can outlive its lease and be requeued, paying twice. Low risk
at short durations; real at long ones.
Budget is charged at the logical unit, not the physical POST; api_client can
issue up to three POSTs per charge, and condense/simplify/advisor calls do not
charge at all.
