Generation models

Host and distribute a model

murakumo.cloud now includes a public model repository API, separate from the fleet's runnable-model catalogue below. Publishers upload immutable content-addressed blocks, then publish a DAG-CBOR revision manifest with model card, license, file hashes and provenance. Public reads and byte-range downloads need no token; repository, block and revision writes require a DID-bound murakumo token with the model-hosting scope.

• GET /api/v1/model-hosting — capability and limits
• GET /api/v1/model-hosting/repositories — searchable hosted repositories
• Publishing protocol — repository, block and revision request shapes

This is the hosting plane: durable weights and versioned metadata. The tables below are the execution plane: models currently wired into the Murakumo generation fleet. Hosting a revision does not claim that the fleet can execute it.

Video

hunyuanvideo-1.5

Best OSS output quality here: monochrome-manga by default with no prompting for it. Also the slowest by a wide margin.

ltx-2.3

Was the default until 2026-08-03, when image-to-video was measured returning status=done with every frame black (luma mean 0, std 0). Text-to -video passed an end-to-end run on 2026-07-16 and is still selectable; image-to-video is refused rather than silently returning an empty clip.

seedance-2.0

Hosted reference-to-video, up to 4K, up to 9 reference images. Verified against the live API on 2026-08-03.

seedance-2.0-fast

The production path for comic work as of 2026-08-03. Multi-shot actually works: one 15-second clip came back containing the four requested shots in order — a whole page from a single generation, which is why per-page cost is a fair unit here.

wan-dancer-14b

Dance / cosplay image-to-video. Weights are still being fetched by the disk-gated watcher and the graph is not wired, so selecting it fails closed instead of falling back to something else.

wan2.2-ti2v-5b

Default. Pen-and-ink look down to the screentone grain; character consistency holds from a single reference. Output is a fixed 768x448 landscape, so tall panels get cropped — measured 2026-08-03.

Cost is per manuscript page rather than per clip because the two paths divide the work differently: the fleet models take one clip per shot, while Seedance's multi-shot returns a page's worth in a single generation. Per-clip pricing would flatter the hosted path by 4x.

Image

3D, rig and VRM

Audio, voice, motion and effects

The measurement behind the video defaults

Six engines, one scenario, one session — the data the default model is chosen from, published rather than summarised.

Style fit / character consistency / prop intelligence are 1–5 scores from one measurement session on one scenario (monochrome manga + screentone, node: gad (AMD Ryzen AI Max+ 395 / gfx1151 / ROCm / ComfyUI 0.22.0)). They rank these six engines against each other on comic work; they are not general model benchmarks. Two ids in this table are not in the catalogue above: seedance-2.0-mini is the id this bench recorded for the fast hosted tier that the catalogue now serves as seedance-2.0-fast (same $0.22 per page, same 90s per clip), and animatediff-sdxl was evaluated and not adopted. Row names are left exactly as they were recorded.

Re-measured 2026-08-03

読切 P.12(4コマ) と P.21(4コマ) を genko の B4 原稿へ実際に組んだ — the run that moved the default.

• ltx-2.3 — broken i2v ジョブが status=done を返すのに出力が全フレーム真っ黒(輝度 mean 0 / std 0、49frames 45KB)。 2026-07-25 の bench では :oss-speed だったが、2026-08-03 時点では成果物にならない。 無料経路を試すなら wan2.2 を使う。
• wan2.2-ti2v-5b — works but unfit 動く。キャラ一致も良い(参照が単体なら)。ただし縦コマの原稿には合わない: ①出力が 768x448 の横長固定で、縦長のコマに入れると顔や主体が切れる ②i2v なので参照画像が1フレーム目になり、構図が寄り絵に固定される — 引きの絵や、参照と違う構図を作れない ③1クリップ1ショットなので 4コマ=4クリップ=約12分/頁 比率をコマに合わせて生成する改善は未検証(効く見込みはある)。
• seedance-2.0-fast — production multi-shot が実際に効いた: duration 10 で 15秒クリップが返り、その中に 指示した4ショットが順に入っていた = 1クリップで原稿1頁ぶん。 1コマ1クリップで焼くと4倍払うことになる(実際に1頁 $0.88 払った)。

What is not here

Models people ask for that this fleet does not serve, and why.

Reading this catalogue from code

The same data, unstyled and always current: GET /api/v1/generation/catalog . No token required — it is a menu, not a capability. The document reports execute as menu-not-configured unless GET /api/v1 says generation-configured is true.

curl -s https://murakumo.cloud/api/v1/generation/catalog \
  | jq '{execute, speech, voice: .modalities.voice}'

Running a generation is gated and, on the public Worker observed 2026-08-27, not configured: POST /api/v1/generation needs a murakumo access token with the generation scope and a bound generation upstream. POST /api/v1/audio/speech is the OpenAI-shaped speech contract and returns 501 with configured: false until a local TTS backend can actually return audio. Voice ids (cosyvoice2, kokoro) stay on the menu as unverified. See the connection guide and the storefront.

Status vocabulary