Same videos. Less money.
Everything here was run for real on our own account. If you are not an engineer, the first two parts are all you need.
What does it really cost?
The 8 showcase videos at 1080p, calculated at the rates billed in our console, against Genjutsu motion transfer for the same videos at 1080p.
| Video | Length | Ours (US$) | Genjutsu 1080p (US$) |
|---|---|---|---|
| Pink Loft | 8 s | 1.40 | 6.53 |
| Orange Jacket | 5 s | 0.80 | 4.08 |
| Subterranean | 15 s | 2.52 | 12.24 |
| Ceiling Fan | 15 s | 2.63 | 12.24 |
| Street Dance, Golden Gown, Anime Times Square, Handbag Slide | 5 s each | 0.80 each | 4.08 each |
| Total (8 videos) | 10.54 | 51.41 |
79.5% less
- Model
- Wan 3.0, 1080p
- Length
- 5 s + 5 s guide
- List price
- US$ 2.00
- Billed
- US$ 0.80
Billed at 40% of list on our account. Whether that discount lasts is not documented.
Against Genjutsu at 720p, the resolution of the official 9:16 showcases, the saving is about 50%; ours come out at 1080p. The 50% Genjutsu discount was measured at 480p; at 1080p it is our assumption. Our longer clips are computed at the 40% rate billed on 5 s clips. C36
Does it protect my money?
Two real runs on our account, with the log lines exactly as Kaneda printed them.
[eco] unitario≈$2.00 x 1 = $2.00 | dia=$0.00 cap=$0.5
[eco] slot 0: TETO — $2.00 > cap $0.50; para de submeter
[resumo] gasto=$0.00 no dia | candidatos ok=0A US$ 2.00 job against a US$ 0.50 cap: refused, 0 requests, US$ 0.00 spent. C2
[eco] slot 0: retoma 764cc291ada590eb (completed, ledger de 2026-10-18)
[resumo] gasto=$1.27 no dia | candidatos ok=1The second run found the finished job and sent nothing; the day total stays at the US$ 1.27 of the first run. C1
Does the video come out better?
Same inputs on both sides. Drag the line on each video.
The same job. Left, what Higgsfield returned; right, after Kaneda: 12 frozen frames, then 0. C35
Technical details: offline tests and a mixed result
Offline tests
141 offline tests passed on the clean package, network blocked, 2026-10-19 C37
A mixed result
Every claim on this site, with its proof.
The proofs live in our records; here we show the result and the date.
| ID | Claim | Status | Evidence |
|---|---|---|---|
C1 | Re-running an identical recipe returns the stored result with zero HTTP calls. On 2026-10-18 the second run of a recipe resumed job 764cc291ada590eb and sent nothing; the day total stayed at the US$ 1.27 of the first run. | Measured | run log, |
C2 | On 2026-10-18 a real run with a US$ 0.50 cap and a US$ 2.00 quote was refused before any request was sent: US$ 0.00 spent. An earlier synthetic dry run quoted at US$ 34.12 was refused the same way by a US$ 4.00 cap. | Measured | run log, |
C3 | Every call is quoted for free first. Uploads are cached. 30 offline tests cover the governor. | Measured | by design of the code, |
C4 | A 113-frame take came back with exactly 113 frames and the original audio. | Measured, one case | measured on the video file |
C5 | In an earlier run of that take, a voice the model made up did not reach the final video. | Measured, one case | measured on the video file |
C6 | Each rule of the method came from a failure we observed. | Measured | run log |
C7 | Our own identity metric, SFace, did not predict approval: AUC 0.33 on 7 labeled videos. Motion energy did better, AUC 0.80. A control we rejected scored 0.70 median. An approved take scored 0.72. So Kaneda logs the signals and keeps a human approval step. | Measured | measured on the video file |
C8 | Specs shown on this site are measured with ffprobe on the original files. The web previews are re-encoded and carry no audio. | Measured | measured on the video file |
C9 | Credits we paid per attempt: the Katana workflow took 70.12 for a 10 s reel. A Seedance 2.5 reel of three scenes with QA rerolls took 140.12. Genjutsu motion transfer at 720p took 105 and an object swap took 112. | Measured | console log |
C10 | Wan 3.0 pricing changes. Kaneda's quote command prints the live estimate for your account, free of charge. | By design | by design of the code |
C11 | In three days we ran 19 jobs: 16.04 USD in quotes plus 10 credits. These are quotes from the estimate endpoint at run time, not invoices. | Measured | run log, |
C13 | SFace median similarity against held-out photos, as telemetry only: Katana run 0.66, Seedance three-scene run 0.53, best Genjutsu run 0.69, Wan 3.0 with Kaneda 0.72 and 0.60. It measures similarity, not quality (see C7). | Measured | measured on the video file |
C14 | Eleven Wan 3.0 reference-to-video runs were quoted 1.00 USD each by the engine formula, which assumed a 50% discount. The provider console shows these runs billed at 40% of list: 0.80 USD for 5 s at 1080p with a 5 s guide. The 720p control was quoted 0.50 USD. | Measured | console log, |
C15 | Output we measured: the Katana workflow and Genjutsu both delivered 720p at 24 fps. Wan 3.0 with Kaneda delivered 1080p, at the frame rate of the source clip. | Measured | measured on the video file |
C16 | 30 of 30 offline tests passed on a clean copy in 0.095 s, with an empty environment and no credentials. | Measured | run log, |
C17 | A spending-cap refusal, a replayed recipe and a retry with the same idempotency key are each covered by a named test, and all three pass. | Measured | by design of the code |
C18 | Control run: one old photo, the raw 5.04 s driver clip, a plain prompt, audio on, 720p, quoted 0.50 USD. It lost the hairstyle, picked up the dancer's watch and chain and a child in the pool, and we rejected it. | Measured, one case | run log |
C19 | A later run with Kaneda, on a different driver clip, so this pair does not isolate Kaneda: 1080p, quoted 1.00 USD. We approved it. | Measured, one case | run log, |
C20 | A wide shot that failed on the first run (the model re-rendered the original actor) was approved on the next run with Kaneda. | Measured, one case | run log |
C21 | Genjutsu motion transfer on our own inputs: runs at 480p scored 0.56 to 0.57 SFace median. The best run, at 720p, scored 0.69 median and 0.71 p90. | Measured | measured on the video file |
C22 | The Katana workflow took about 18 minutes end to end for the 10 s reel. | Measured, one case | run log |
C23 | On our seven videos, two replayed recipes would have cost about 3.70 USD and cost nothing. The synthetic cap test refused 34.12 USD. | Measured | run log |
C24 | Unit price did not drop. The quote is the price the API returns before the call: a Seedance 2.5 clip was quoted 1.8488 USD. Kaneda saves what you would otherwise spend twice. | Measured | run log |
C25 | Kaneda talks to two hosts: Higgsfield's API and, only if you add a Jev key, an optional decision API. There is no license server. Nothing is sent to us unless you switch the optional plane on. | By design | by design of the code |
C26 | Wan 3.0 reference-to-video, 5 s at 1080p with a 5 s guide: US$ 2.00 at list price. Our account was billed US$ 0.80, which is 40% of list. Whether that discount lasts is not documented. | Measured | console log, |
C27 | Real A/B on the Clone recast preset, same inputs, Wan 3.0 at 1080p: without Kaneda (raw clip, plain prompt) the later cuts went back to the original actor in a checked jacket and a cap. With Kaneda every cut shows the character from the card. | Measured, one case | measured on the video file, |
C28 | Real A/B on the Dance duo preset, a mixed result: without Kaneda the cargo pants and the face without sunglasses from the cards came out right. With Kaneda the run copied the driver's jeans and sunglasses. | Measured, one case | measured on the video file, |
C30 | A reel with three cuts and 219 frames went through Kaneda and came back with 219 frames, both cuts on the same frames (101 and 154). | Measured, one case | measured on the video file |
C31 | Genjutsu Restyle through the open API returned 59 frames for a 96 frame input. Remap returned exactly 96 frames at 24 fps. | Measured, one case | measured on the video file |
C32 | The showcase clips rebuild Higgsfield's Genjutsu showcase presets with Wan 3.0 through Kaneda, from each preset's own character card and driver clip. Each 5 s clip at 1080p comes to 0.80 USD at the rate billed in our console; the 8 s Pink Loft, 1.28 USD at the same rate. | Measured | console log, |
C33 | At list price, a 5 s clip at 1080p costs 5.69 USD on Seedance 2.5 (1.1372 USD per second of output, without input video) and 2.00 USD on Wan 3.0 reference-to-video with a 5 s guide: 65% less. Our account was billed 0.80 USD for that Wan clip. | External source | external source |
C34 | Kaneda calls Higgsfield through its pay-as-you-go API with your own key, so no monthly plan is needed to use it. A subscription can still be used through the CLI channel. | By design | by design of the code |
C35 | Pink Loft, 8 s at 1080p: 12 frozen frames in the raw output, 0 after Kaneda. Subterranean, 15 s vertical: 10 and 0. | Measured | measured on the video file |
C36 | Calculated, not billed: cost of the 8 showcase videos at the rates billed in our console. Ours: Wan 3.0 at 1080p at 40% of list, (output + guide seconds) × 0.20 × 0.40, plus the Qwen card and plate where used (0.115 or 0.225): Pink Loft 1.40, Orange Jacket 0.80, Subterranean 2.52, Ceiling Fan 2.63, four 5 s clips 0.80 each, total 10.54 USD. Genjutsu motion transfer at 1080p: 1.632 USD per second at list, 0.816 with the 50% discount we measured at 480p (assumed at 1080p): 6.53, 4.08, 12.24, 12.24 and 4.08 each, total 51.41 USD. 79.5% less. Against Genjutsu at 720p the saving is about 50%. | Measured | calculated with the console rate |
C37 | Kaneda 1.0.0 as shipped, unzipped into an empty folder with the network blocked: 141 offline tests passed and 1 was skipped because OpenCV was not installed. With OpenCV that one passes too. | Measured | run log, |