@nodaro/shared 3.37.0 → 3.38.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/{chunk-4XI6B4RU.js → chunk-2BHOAYZW.js} +17 -4
- package/dist/chunk-2BHOAYZW.js.map +1 -0
- package/dist/{chunk-M6JHRWVY.cjs → chunk-7VDPBFER.cjs} +19 -3
- package/dist/chunk-7VDPBFER.cjs.map +1 -0
- package/dist/{chunk-6Y66YM4M.cjs → chunk-AQA6YXY5.cjs} +39 -10
- package/dist/chunk-AQA6YXY5.cjs.map +1 -0
- package/dist/{chunk-YVKUCBY2.js → chunk-BDSFMY4X.js} +5 -18
- package/dist/chunk-BDSFMY4X.js.map +1 -0
- package/dist/chunk-BLZBDPZJ.js +16 -0
- package/dist/chunk-BLZBDPZJ.js.map +1 -0
- package/dist/{chunk-NARVPCWW.cjs → chunk-D33YBZ3B.cjs} +170 -143
- package/dist/chunk-D33YBZ3B.cjs.map +1 -0
- package/dist/{chunk-7Q75OZBL.cjs → chunk-HZ46RDLD.cjs} +9 -24
- package/dist/chunk-HZ46RDLD.cjs.map +1 -0
- package/dist/{chunk-WXNQMJ67.js → chunk-LOWOJN65.js} +100 -6
- package/dist/chunk-LOWOJN65.js.map +1 -0
- package/dist/chunk-M7G66F6H.js +17 -0
- package/dist/chunk-M7G66F6H.js.map +1 -0
- package/dist/chunk-PSAWDKO3.cjs +19 -0
- package/dist/chunk-PSAWDKO3.cjs.map +1 -0
- package/dist/{chunk-L6HS52BV.js → chunk-QLCSS6DK.js} +165 -141
- package/dist/chunk-QLCSS6DK.js.map +1 -0
- package/dist/chunk-T3TOSWVH.cjs +20 -0
- package/dist/chunk-T3TOSWVH.cjs.map +1 -0
- package/dist/{chunk-J65CAOUV.cjs → chunk-TGZJXCLL.cjs} +105 -5
- package/dist/chunk-TGZJXCLL.cjs.map +1 -0
- package/dist/{chunk-WD6AJ533.js → chunk-THGNDVVH.js} +29 -4
- package/dist/chunk-THGNDVVH.js.map +1 -0
- package/dist/entries/credits.cjs +135 -134
- package/dist/entries/credits.js +3 -2
- package/dist/entries/llm-models.cjs +72 -52
- package/dist/entries/llm-models.d.cts +50 -13
- package/dist/entries/llm-models.d.ts +50 -13
- package/dist/entries/llm-models.js +2 -2
- package/dist/entries/media.cjs +66 -42
- package/dist/entries/media.d.cts +145 -4
- package/dist/entries/media.d.ts +145 -4
- package/dist/entries/media.js +1 -1
- package/dist/entries/models.cjs +43 -27
- package/dist/entries/models.d.cts +53 -3
- package/dist/entries/models.d.ts +53 -3
- package/dist/entries/models.js +2 -2
- package/dist/entries/pipeline.d.cts +2 -2
- package/dist/entries/pipeline.d.ts +2 -2
- package/dist/entries/social.cjs +187 -174
- package/dist/entries/social.d.cts +26 -1
- package/dist/entries/social.d.ts +26 -1
- package/dist/entries/social.js +3 -2
- package/dist/index.cjs +499 -426
- package/dist/index.d.cts +6 -5
- package/dist/index.d.ts +6 -5
- package/dist/index.js +7 -6
- package/dist/{scene-node-types-Cvbtf-0a.d.cts → model-constants-C43YS07I.d.cts} +1 -497
- package/dist/{scene-node-types-Cvbtf-0a.d.ts → model-constants-C43YS07I.d.ts} +1 -497
- package/dist/scene-node-types-BSrTKnV7.d.cts +499 -0
- package/dist/scene-node-types-BSrTKnV7.d.ts +499 -0
- package/package.json +1 -1
- package/src/__tests__/cast-swap-limits.test.ts +40 -0
- package/src/__tests__/fixtures/root-exports.baseline.json +6 -0
- package/src/__tests__/llm-models.test.ts +8 -8
- package/src/__tests__/llm-price-key.test.ts +55 -0
- package/src/__tests__/scrape-analysis-pricing.test.ts +41 -0
- package/src/__tests__/video-mask.test.ts +27 -0
- package/src/cast-swap-limits.ts +52 -0
- package/src/entries/llm-models.ts +4 -0
- package/src/entries/models.ts +7 -1
- package/src/entries/social.ts +1 -0
- package/src/llm-models.ts +61 -13
- package/src/mask-composite.ts +80 -3
- package/src/scrape-analysis-pricing.ts +73 -0
- package/src/script-model.ts +21 -0
- package/src/video-mask.ts +97 -0
- package/dist/chunk-4XI6B4RU.js.map +0 -1
- package/dist/chunk-6Y66YM4M.cjs.map +0 -1
- package/dist/chunk-7Q75OZBL.cjs.map +0 -1
- package/dist/chunk-FG4E7MTG.cjs +0 -4
- package/dist/chunk-FG4E7MTG.cjs.map +0 -1
- package/dist/chunk-J65CAOUV.cjs.map +0 -1
- package/dist/chunk-L6HS52BV.js.map +0 -1
- package/dist/chunk-M6JHRWVY.cjs.map +0 -1
- package/dist/chunk-NARVPCWW.cjs.map +0 -1
- package/dist/chunk-TZE673G6.js +0 -3
- package/dist/chunk-TZE673G6.js.map +0 -1
- package/dist/chunk-WD6AJ533.js.map +0 -1
- package/dist/chunk-WXNQMJ67.js.map +0 -1
- package/dist/chunk-YVKUCBY2.js.map +0 -1
|
@@ -0,0 +1,41 @@
|
|
|
1
|
+
import { describe, it, expect } from "vitest"
|
|
2
|
+
import { scrapeAnalysisPriceParts, scrapeAnalysisQuoteId, SCRAPE_ANALYSIS_FEATURE } from "../scrape-analysis-pricing.js"
|
|
3
|
+
import { buildMetaAdsScrapeCreditId, META_ADS_SCRAPE_MAX_COUNT } from "../meta-ads-scrape.js"
|
|
4
|
+
import { buildInstagramScrapeCreditId, INSTAGRAM_SCRAPE_MAX_COUNT } from "../instagram-scrape.js"
|
|
5
|
+
import { LLM_FEATURE_DEFAULTS } from "../llm-models.js"
|
|
6
|
+
import { resolveScriptModelId, SCRIPT_LEGACY_PROVIDER_MODELS } from "../script-model.js"
|
|
7
|
+
|
|
8
|
+
describe("scrapeAnalysisPriceParts — a scrape's analysis billed per call", () => {
|
|
9
|
+
it("no analysis, or another node: null", () => {
|
|
10
|
+
expect(scrapeAnalysisPriceParts({ type: "meta-ads-scrape", data: { count: 10 } })).toBeNull()
|
|
11
|
+
expect(scrapeAnalysisPriceParts({ type: "llm-chat", data: { analyze: true } })).toBeNull()
|
|
12
|
+
})
|
|
13
|
+
it("Meta Ads: its fixed scrape (no analysis suffix), the analysis model, count × sources items (clamped)", () => {
|
|
14
|
+
expect(scrapeAnalysisPriceParts({ type: "meta-ads-scrape", data: { count: 10, analyze: true } })).toEqual({
|
|
15
|
+
fixedId: buildMetaAdsScrapeCreditId({ count: 10, sources: 1 }),
|
|
16
|
+
analysisModel: LLM_FEATURE_DEFAULTS[SCRAPE_ANALYSIS_FEATURE],
|
|
17
|
+
items: 10,
|
|
18
|
+
})
|
|
19
|
+
expect(scrapeAnalysisPriceParts({ type: "meta-ads-scrape", data: { count: 9999, analyze: true, analysisModel: "claude-opus-5.5" } })).toMatchObject({ analysisModel: "claude-opus-5.5", items: META_ADS_SCRAPE_MAX_COUNT })
|
|
20
|
+
})
|
|
21
|
+
it("Instagram: the count the route accepts × the targets", () => {
|
|
22
|
+
const parts = scrapeAnalysisPriceParts({ type: "instagram-scrape", data: { count: 500, analyze: true, mode: "profile", targets: "a\nb" } })
|
|
23
|
+
expect(parts).toEqual({ fixedId: buildInstagramScrapeCreditId({ count: INSTAGRAM_SCRAPE_MAX_COUNT, sources: 2 }), analysisModel: LLM_FEATURE_DEFAULTS[SCRAPE_ANALYSIS_FEATURE], items: INSTAGRAM_SCRAPE_MAX_COUNT * 2 })
|
|
24
|
+
})
|
|
25
|
+
})
|
|
26
|
+
|
|
27
|
+
describe("scrapeAnalysisQuoteId — the whole analysis as one price id", () => {
|
|
28
|
+
it("is the analysis model's metered key as a `<key>:<items>u` quote", () => {
|
|
29
|
+
const parts = scrapeAnalysisPriceParts({ type: "meta-ads-scrape", data: { count: 10, analyze: true, analysisModel: "claude-opus-5.5" } })!
|
|
30
|
+
expect(scrapeAnalysisQuoteId(parts)).toBe(`${SCRAPE_ANALYSIS_FEATURE}:claude-opus-5.5:10u`)
|
|
31
|
+
})
|
|
32
|
+
})
|
|
33
|
+
|
|
34
|
+
describe("resolveScriptModelId — the model a Generate Script call runs on", () => {
|
|
35
|
+
it("an explicit model, else a legacy provider's, else the default", () => {
|
|
36
|
+
expect(resolveScriptModelId("claude", "gpt-5.2")).toBe("gpt-5.2")
|
|
37
|
+
expect(resolveScriptModelId("claude")).toBe(SCRIPT_LEGACY_PROVIDER_MODELS.claude)
|
|
38
|
+
expect(resolveScriptModelId()).toBe(LLM_FEATURE_DEFAULTS["generate-script"])
|
|
39
|
+
expect(resolveScriptModelId("toString")).toBe(LLM_FEATURE_DEFAULTS["generate-script"])
|
|
40
|
+
})
|
|
41
|
+
})
|
|
@@ -2,9 +2,11 @@ import { describe, it, expect } from "vitest"
|
|
|
2
2
|
import { readFileSync } from "node:fs"
|
|
3
3
|
import { resolve } from "node:path"
|
|
4
4
|
import {
|
|
5
|
+
VIDEO_MASK_DEFAULTS,
|
|
5
6
|
VIDEO_MASK_LIMITS,
|
|
6
7
|
VIDEO_MASK_PROVIDERS,
|
|
7
8
|
resolveVideoMaskProvider,
|
|
9
|
+
resolveVideoMaskZone,
|
|
8
10
|
videoMaskBilledSeconds,
|
|
9
11
|
videoMaskCreditId,
|
|
10
12
|
videoMaskCreditIds,
|
|
@@ -65,6 +67,31 @@ describe("Video Mask pricing", () => {
|
|
|
65
67
|
})
|
|
66
68
|
})
|
|
67
69
|
|
|
70
|
+
describe("Video Mask zone", () => {
|
|
71
|
+
it("is off by default, and each margin is 0–49 % (the inner rectangle can never be empty)", () => {
|
|
72
|
+
expect(VIDEO_MASK_DEFAULTS.zone).toBeNull()
|
|
73
|
+
expect(VIDEO_MASK_LIMITS.ZONE_MARGIN_PCT_MIN).toBe(0)
|
|
74
|
+
expect(VIDEO_MASK_LIMITS.ZONE_MARGIN_PCT_MAX).toBe(49)
|
|
75
|
+
expect(VIDEO_MASK_LIMITS.ZONE_MARGIN_PCT_MAX * 2).toBeLessThan(100)
|
|
76
|
+
})
|
|
77
|
+
|
|
78
|
+
it("resolveVideoMaskZone reads a zone as the filter does: margins clamped to 0–49, anything else 0, all 0 = null", () => {
|
|
79
|
+
expect(resolveVideoMaskZone({ leftPct: 10, rightPct: 20, topPct: 5, bottomPct: 0 })).toEqual({ leftPct: 10, rightPct: 20, topPct: 5, bottomPct: 0 })
|
|
80
|
+
expect(resolveVideoMaskZone({ leftPct: 80, rightPct: -3 })).toEqual({ leftPct: 49, rightPct: 0, topPct: 0, bottomPct: 0 })
|
|
81
|
+
expect(resolveVideoMaskZone({ leftPct: "10", topPct: Number.NaN, bottomPct: Infinity })).toBeNull()
|
|
82
|
+
expect(resolveVideoMaskZone({ leftPct: 0, rightPct: 0, topPct: 0, bottomPct: 0 })).toBeNull()
|
|
83
|
+
expect(resolveVideoMaskZone(null)).toBeNull()
|
|
84
|
+
expect(resolveVideoMaskZone(undefined)).toBeNull()
|
|
85
|
+
expect(resolveVideoMaskZone("10,10,10,10")).toBeNull()
|
|
86
|
+
})
|
|
87
|
+
})
|
|
88
|
+
|
|
89
|
+
describe("Video Mask perShot", () => {
|
|
90
|
+
it("is off by default (one pass over the whole clip)", () => {
|
|
91
|
+
expect(VIDEO_MASK_DEFAULTS.perShot).toBe(false)
|
|
92
|
+
})
|
|
93
|
+
})
|
|
94
|
+
|
|
68
95
|
describe("Mask Composite units", () => {
|
|
69
96
|
// Repeated in docs/nodes/processing-video/mask-composite.md (at the 10-credit row).
|
|
70
97
|
it.each([[3, 1], [5, 1], [10, 2], [10.0000001, 2], [61, 13], [300, 60]])("%s s → %s unit(s)", (sec, units) => {
|
|
@@ -0,0 +1,52 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Cast Swap's input limits, and what a trend built on it may put in its
|
|
3
|
+
* prompts. One definition read by the server (which enforces them) and by
|
|
4
|
+
* every screen that edits them (which holds Save before a refusal), so the two
|
|
5
|
+
* cannot drift apart.
|
|
6
|
+
*/
|
|
7
|
+
|
|
8
|
+
/** The longest `postPrompt` Cast Swap takes. */
|
|
9
|
+
export const CAST_SWAP_POST_PROMPT_MAX = 4000
|
|
10
|
+
|
|
11
|
+
/**
|
|
12
|
+
* A trend's prompt limits.
|
|
13
|
+
*
|
|
14
|
+
* - `trendPromptMax` / `prePromptMax`: the longest prompt a trend version may
|
|
15
|
+
* store.
|
|
16
|
+
* - `detailsMax`: the longest details a user may add to a run. They follow the
|
|
17
|
+
* trend prompt in Cast Swap's `postPrompt`, after a blank line, so a new
|
|
18
|
+
* trend prompt is held to `CAST_SWAP_POST_PROMPT_MAX − 2 − detailsMax`.
|
|
19
|
+
* - `engineTextReserve`: the room kept for the text Cast Swap writes around a
|
|
20
|
+
* trend's prompts in the prompt it builds, at its longest (a block for each
|
|
21
|
+
* of the 13 roles a run allows with 200-character role names, five
|
|
22
|
+
* 500-character reference lines, the keep, mask and edit lines and the
|
|
23
|
+
* header: about 16,100 characters, rounded up so a small wording change does
|
|
24
|
+
* not need this moved). The pre prompt, the trend prompt and the details
|
|
25
|
+
* share what is left of Seedance 2.5's prompt cap.
|
|
26
|
+
*/
|
|
27
|
+
export const TREND_PROMPT_LIMITS = {
|
|
28
|
+
trendPromptMax: 3000,
|
|
29
|
+
prePromptMax: 3000,
|
|
30
|
+
detailsMax: 1000,
|
|
31
|
+
engineTextReserve: 17_000,
|
|
32
|
+
} as const
|
|
33
|
+
|
|
34
|
+
/** Cast Swap's `faceMasking` bounds, each `[min, max]` inclusive. */
|
|
35
|
+
export const CAST_SWAP_FACE_MASKING_BOUNDS = {
|
|
36
|
+
/** `minFaceHeightPct` and `maxFaceHeightPct`, % of the frame's height (max 0 = no limit). */
|
|
37
|
+
faceHeightPct: [0, 100],
|
|
38
|
+
/** Each `zone` margin, % of the frame. */
|
|
39
|
+
zonePct: [0, 49],
|
|
40
|
+
/** % of the frame's width. */
|
|
41
|
+
blurStrengthPct: [0.1, 5],
|
|
42
|
+
threshold: [0.05, 0.95],
|
|
43
|
+
} as const satisfies Record<string, readonly [number, number]>
|
|
44
|
+
|
|
45
|
+
/** What Cast Swap uses for a `faceMasking` setting left out. */
|
|
46
|
+
export const CAST_SWAP_FACE_MASKING_DEFAULTS = {
|
|
47
|
+
minFaceHeightPct: 15,
|
|
48
|
+
maxFaceHeightPct: 0,
|
|
49
|
+
zonePct: 0,
|
|
50
|
+
blurStrengthPct: 2,
|
|
51
|
+
threshold: 0.5,
|
|
52
|
+
} as const
|
|
@@ -50,6 +50,9 @@ export {
|
|
|
50
50
|
resolveLlmCreditId,
|
|
51
51
|
supportsAdvancedMode,
|
|
52
52
|
llmServesDirect,
|
|
53
|
+
llmPriceKey,
|
|
54
|
+
LLM_FEATURE_DEFAULTS_FROM,
|
|
55
|
+
llmDefaultsFeature,
|
|
53
56
|
LLM_CREDIT_RUNGS,
|
|
54
57
|
type LlmCreditRung,
|
|
55
58
|
llmCreditIdForRung,
|
|
@@ -69,3 +72,4 @@ export {
|
|
|
69
72
|
type LlmModelDef,
|
|
70
73
|
type LlmReasoningEffort,
|
|
71
74
|
} from "../llm-models.js"
|
|
75
|
+
export { SCRIPT_LEGACY_PROVIDER_MODELS, resolveScriptModelId } from "../script-model.js"
|
package/src/entries/models.ts
CHANGED
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
/**
|
|
2
|
-
* `@nodaro/shared/models` — Model constants: provider lists, prompt limits, per-model capability tables and defaults.
|
|
2
|
+
* `@nodaro/shared/models` — Model constants: provider lists, prompt limits, per-model capability tables and defaults, Cast Swap's input limits.
|
|
3
3
|
*
|
|
4
4
|
* One of the package's subpath entries (see `src/index.ts`). The root
|
|
5
5
|
* re-exports every entry, so each name here is also importable from
|
|
@@ -197,6 +197,12 @@ export {
|
|
|
197
197
|
maxVideoDurationSec,
|
|
198
198
|
PRICING_DEFAULT_RESOLUTION,
|
|
199
199
|
} from "../model-constants.js"
|
|
200
|
+
export {
|
|
201
|
+
CAST_SWAP_POST_PROMPT_MAX,
|
|
202
|
+
TREND_PROMPT_LIMITS,
|
|
203
|
+
CAST_SWAP_FACE_MASKING_BOUNDS,
|
|
204
|
+
CAST_SWAP_FACE_MASKING_DEFAULTS,
|
|
205
|
+
} from "../cast-swap-limits.js"
|
|
200
206
|
export type {
|
|
201
207
|
ImageAspectRatio,
|
|
202
208
|
ImageResolution,
|
package/src/entries/social.ts
CHANGED
|
@@ -6,6 +6,7 @@
|
|
|
6
6
|
* `@nodaro/shared`.
|
|
7
7
|
*/
|
|
8
8
|
export { normalizePinterestUrl } from "../pinterest-url.js"
|
|
9
|
+
export { SCRAPE_ANALYSIS_FEATURE, scrapeAnalysisPriceParts, scrapeAnalysisQuoteId, type ScrapeAnalysisPriceParts } from "../scrape-analysis-pricing.js"
|
|
9
10
|
export { SCRAPER_OUTPUT_FIELDS } from "../scraper-output-schemas.js"
|
|
10
11
|
export {
|
|
11
12
|
SOCIAL_POST_NODE_TYPES,
|
package/src/llm-models.ts
CHANGED
|
@@ -1064,27 +1064,30 @@ export const STRUCTURED_VISION_MODELS = LLM_MODELS.filter(
|
|
|
1064
1064
|
|
|
1065
1065
|
/**
|
|
1066
1066
|
* Models {@link STRUCTURED_VISION_MODELS} carries that describe-to-picker does
|
|
1067
|
-
* NOT offer
|
|
1068
|
-
*
|
|
1069
|
-
*
|
|
1070
|
-
*
|
|
1071
|
-
*
|
|
1072
|
-
*
|
|
1067
|
+
* NOT offer — now none.
|
|
1068
|
+
*
|
|
1069
|
+
* Claude Fable 5 was withheld here (production cost audit, 2026-10-20): a
|
|
1070
|
+
* legend-sized read cost far above every tier rung. Metered pricing charges
|
|
1071
|
+
* each read what it really costs (tokens × the model's rate on the lane that
|
|
1072
|
+
* served it), so nothing needs withholding and the set is empty.
|
|
1073
|
+
*
|
|
1074
|
+
* @deprecated Always empty since metered LLM pricing. Kept exported so existing
|
|
1075
|
+
* importers keep compiling; it will be removed in a future major.
|
|
1073
1076
|
*/
|
|
1074
|
-
export const DESCRIBE_TO_PICKER_WITHHELD_MODELS: ReadonlySet<string> = new Set(
|
|
1077
|
+
export const DESCRIBE_TO_PICKER_WITHHELD_MODELS: ReadonlySet<string> = new Set<string>()
|
|
1075
1078
|
|
|
1076
1079
|
/**
|
|
1077
|
-
* The same
|
|
1078
|
-
*
|
|
1079
|
-
*
|
|
1080
|
-
*
|
|
1081
|
-
*
|
|
1080
|
+
* The same (empty) set for the other legend-sized structured reads:
|
|
1081
|
+
* `POST /v1/text-to-picker` and `POST /v1/llm/structured`.
|
|
1082
|
+
*
|
|
1083
|
+
* @deprecated Always empty since metered LLM pricing; see
|
|
1084
|
+
* {@link DESCRIBE_TO_PICKER_WITHHELD_MODELS}.
|
|
1082
1085
|
*/
|
|
1083
1086
|
export const LLM_STRUCTURED_WITHHELD_MODELS: ReadonlySet<string> = DESCRIBE_TO_PICKER_WITHHELD_MODELS
|
|
1084
1087
|
|
|
1085
1088
|
/**
|
|
1086
1089
|
* The models describe-to-picker offers and accepts: the structured vision
|
|
1087
|
-
* models minus {@link DESCRIBE_TO_PICKER_WITHHELD_MODELS}. The ONE list the
|
|
1090
|
+
* models minus {@link DESCRIBE_TO_PICKER_WITHHELD_MODELS} (empty — every one). The ONE list the
|
|
1088
1091
|
* route gate, the config panel and the quick strip read, so a model the route
|
|
1089
1092
|
* refuses is never offered and a saved node naming one is snapped to the
|
|
1090
1093
|
* default (the route answers it with a `validation_error`).
|
|
@@ -1229,6 +1232,24 @@ export const LLM_FEATURE_DEFAULTS: Record<LlmFeature, string> = {
|
|
|
1229
1232
|
"content-ideas": "gemini-3.6-flash",
|
|
1230
1233
|
}
|
|
1231
1234
|
|
|
1235
|
+
/**
|
|
1236
|
+
* Billed features with no defaults of their own → the feature whose defaults
|
|
1237
|
+
* (model, {@link LLM_FEATURE_DEFAULT_EFFORTS} effort) their route borrows:
|
|
1238
|
+
* Generate Structured runs on Generate Text's, Text to Picker on Describe to
|
|
1239
|
+
* Picker's, the Choose Best judge on `pick-best-llm`'s. Read it through
|
|
1240
|
+
* {@link llmDefaultsFeature}. Structure only — no prices.
|
|
1241
|
+
*/
|
|
1242
|
+
export const LLM_FEATURE_DEFAULTS_FROM: Readonly<Record<string, LlmFeature>> = {
|
|
1243
|
+
"llm-structured": "llm-chat",
|
|
1244
|
+
"text-to-picker": "describe-to-picker",
|
|
1245
|
+
"reduce:pick-best-llm": "pick-best-llm",
|
|
1246
|
+
}
|
|
1247
|
+
|
|
1248
|
+
/** The feature whose defaults a call of `feature` runs with: its {@link LLM_FEATURE_DEFAULTS_FROM} entry, else itself. */
|
|
1249
|
+
export function llmDefaultsFeature(feature: string): string {
|
|
1250
|
+
return Object.hasOwn(LLM_FEATURE_DEFAULTS_FROM, feature) ? LLM_FEATURE_DEFAULTS_FROM[feature]! : feature
|
|
1251
|
+
}
|
|
1252
|
+
|
|
1232
1253
|
/**
|
|
1233
1254
|
* The model describe-to-picker re-asks ONCE when a Claude analyzer could not
|
|
1234
1255
|
* answer for a platform-side reason: its direct lane down or the platform's
|
|
@@ -1547,6 +1568,33 @@ export function llmServesDirect(
|
|
|
1547
1568
|
)
|
|
1548
1569
|
}
|
|
1549
1570
|
|
|
1571
|
+
/**
|
|
1572
|
+
* The price key of a metered LLM call — what a client asks
|
|
1573
|
+
* `POST /v1/credits/model-costs` (or `GET /v1/credits/model-cost?model=`) for
|
|
1574
|
+
* the call's estimated price: `<feature>:<modelId>`, plus `:direct` when the
|
|
1575
|
+
* call is served on the vendor's own API ({@link llmServesDirect}), which is
|
|
1576
|
+
* priced at the direct rate.
|
|
1577
|
+
*
|
|
1578
|
+
* With no `reasoningEffort` the default effort of the feature the call
|
|
1579
|
+
* borrows its defaults from applies ({@link llmDefaultsFeature},
|
|
1580
|
+
* {@link defaultReasoningEffort}) — the effort the call then runs at — so the
|
|
1581
|
+
* key names the lane the call is held and charged on.
|
|
1582
|
+
*
|
|
1583
|
+
* `modelId` must already be resolved: the model the call runs on, never empty
|
|
1584
|
+
* (a caller with no model passes the feature's default, `LLM_FEATURE_DEFAULTS`
|
|
1585
|
+
* of {@link llmDefaultsFeature}). An empty id throws a `TypeError`.
|
|
1586
|
+
*/
|
|
1587
|
+
export function llmPriceKey(
|
|
1588
|
+
feature: string,
|
|
1589
|
+
modelId: string,
|
|
1590
|
+
reasoningEffort?: string,
|
|
1591
|
+
advancedMode?: boolean,
|
|
1592
|
+
): string {
|
|
1593
|
+
if (!modelId) throw new TypeError(`llmPriceKey: modelId must be a resolved model id (feature "${feature}")`)
|
|
1594
|
+
const effort = reasoningEffort ?? defaultReasoningEffort(llmDefaultsFeature(feature), modelId)
|
|
1595
|
+
return llmServesDirect(modelId, effort, advancedMode) ? `${feature}:${modelId}:direct` : `${feature}:${modelId}`
|
|
1596
|
+
}
|
|
1597
|
+
|
|
1550
1598
|
/**
|
|
1551
1599
|
* The credit rungs an LLM feature bills on, cheapest first. `premium-direct` is
|
|
1552
1600
|
* reached only by a premium model served direct (decided 2026-10-08): the bump
|
package/src/mask-composite.ts
CHANGED
|
@@ -10,6 +10,15 @@
|
|
|
10
10
|
* background is fitted to that frame (`backgroundFit`) and loops when shorter;
|
|
11
11
|
* the mask is stretched to the frame and holds its last frame when shorter.
|
|
12
12
|
*
|
|
13
|
+
* `fill` picks what fills the masked area. `"video"` (the default) is the
|
|
14
|
+
* foreground, as above. `"pixelate"` and `"blur"` fill it with the BACKGROUND
|
|
15
|
+
* itself, pixelated or blurred: there is no foreground then (a sent
|
|
16
|
+
* `foregroundUrl` is ignored), so the render follows the BACKGROUND (its size,
|
|
17
|
+
* frame rate and length; `backgroundFit` has nothing to fit), and the run is
|
|
18
|
+
* priced on the background's length. With no foreground the result always
|
|
19
|
+
* carries the background's sound — `audioSource` is ignored there
|
|
20
|
+
* ({@link maskCompositeEffectiveAudioSource}); the canvas hides the control.
|
|
21
|
+
*
|
|
13
22
|
* Priced per started 5 seconds of the result (the foreground's length), like
|
|
14
23
|
* Trim / Loop / Combine Videos: the node's own `model_pricing` row is ONE unit
|
|
15
24
|
* (`maskCompositeUnits` counts them). Worked examples (repeated in
|
|
@@ -23,6 +32,10 @@
|
|
|
23
32
|
export const MASK_COMPOSITE_BACKGROUND_FITS = ["cover", "contain", "stretch"] as const
|
|
24
33
|
export type MaskCompositeBackgroundFit = (typeof MASK_COMPOSITE_BACKGROUND_FITS)[number]
|
|
25
34
|
|
|
35
|
+
/** What fills the masked area: the foreground video, or the background itself pixelated or blurred. */
|
|
36
|
+
export const MASK_COMPOSITE_FILLS = ["video", "pixelate", "blur"] as const
|
|
37
|
+
export type MaskCompositeFill = (typeof MASK_COMPOSITE_FILLS)[number]
|
|
38
|
+
|
|
26
39
|
/** Which sound the result carries. `mix` blends both tracks at equal level. */
|
|
27
40
|
export const MASK_COMPOSITE_AUDIO_SOURCES = ["foreground", "background", "mix", "none"] as const
|
|
28
41
|
export type MaskCompositeAudioSource = (typeof MASK_COMPOSITE_AUDIO_SOURCES)[number]
|
|
@@ -34,7 +47,19 @@ export const MASK_COMPOSITE_BOUNDS = {
|
|
|
34
47
|
/** Grow (+) or shrink (−) the white region, pixels. */
|
|
35
48
|
GROW_MIN: -10,
|
|
36
49
|
GROW_MAX: 10,
|
|
37
|
-
/**
|
|
50
|
+
/** How strongly the foreground covers the background inside the mask
|
|
51
|
+
* (1 = fully, today's output; 0 = not at all). */
|
|
52
|
+
OPACITY_MIN: 0,
|
|
53
|
+
OPACITY_MAX: 1,
|
|
54
|
+
/** Pixelate: the block size as a % of the frame's WIDTH (relative, so a
|
|
55
|
+
* setting looks the same at any resolution); the worker makes the blocks
|
|
56
|
+
* max(2, round(width × pct / 100)) px. 1.25 % = 16 px on a 1280-wide frame. */
|
|
57
|
+
PIXEL_SIZE_PCT_MIN: 0.2,
|
|
58
|
+
PIXEL_SIZE_PCT_MAX: 10,
|
|
59
|
+
/** Blur: the gaussian's sigma as a % of the frame's WIDTH (2 % ≈ sigma 25 px on a 1280-wide frame). */
|
|
60
|
+
BLUR_STRENGTH_PCT_MIN: 0.1,
|
|
61
|
+
BLUR_STRENGTH_PCT_MAX: 5,
|
|
62
|
+
/** The longest leading video (the foreground; the background when `fill` is not video) a run accepts, seconds. */
|
|
38
63
|
MAX_DURATION_SEC: 300,
|
|
39
64
|
} as const
|
|
40
65
|
|
|
@@ -60,6 +85,10 @@ export const MASK_COMPOSITE_DEFAULTS = {
|
|
|
60
85
|
invertMask: false,
|
|
61
86
|
feather: 2,
|
|
62
87
|
grow: 0,
|
|
88
|
+
opacity: 1,
|
|
89
|
+
fill: "video" as MaskCompositeFill,
|
|
90
|
+
pixelSizePct: 1.25,
|
|
91
|
+
blurStrengthPct: 2,
|
|
63
92
|
backgroundFit: "cover" as MaskCompositeBackgroundFit,
|
|
64
93
|
audioSource: "foreground" as MaskCompositeAudioSource,
|
|
65
94
|
} as const
|
|
@@ -68,6 +97,14 @@ export interface MaskCompositeSettings {
|
|
|
68
97
|
readonly invertMask: boolean
|
|
69
98
|
readonly feather: number
|
|
70
99
|
readonly grow: number
|
|
100
|
+
/** Inside the mask: opacity × foreground + (1 − opacity) × background. */
|
|
101
|
+
readonly opacity: number
|
|
102
|
+
/** What fills the masked area (`video` = the foreground). */
|
|
103
|
+
readonly fill: MaskCompositeFill
|
|
104
|
+
/** Pixelate block size, % of the frame width (used when `fill` is `pixelate`). */
|
|
105
|
+
readonly pixelSizePct: number
|
|
106
|
+
/** Blur strength: the gaussian's sigma as a % of the frame width (used when `fill` is `blur`). */
|
|
107
|
+
readonly blurStrengthPct: number
|
|
71
108
|
readonly backgroundFit: MaskCompositeBackgroundFit
|
|
72
109
|
readonly audioSource: MaskCompositeAudioSource
|
|
73
110
|
}
|
|
@@ -75,6 +112,9 @@ export interface MaskCompositeSettings {
|
|
|
75
112
|
const clampInt = (v: unknown, min: number, max: number, fallback: number): number =>
|
|
76
113
|
typeof v === "number" && Number.isFinite(v) ? Math.min(max, Math.max(min, Math.round(v))) : fallback
|
|
77
114
|
|
|
115
|
+
const clampNumber = (v: unknown, min: number, max: number, fallback: number): number =>
|
|
116
|
+
typeof v === "number" && Number.isFinite(v) ? Math.min(max, Math.max(min, v)) : fallback
|
|
117
|
+
|
|
78
118
|
/**
|
|
79
119
|
* Node data (or any loose object) → settings every engine renders with.
|
|
80
120
|
* Unknown enum values fall back to the default and numbers are clamped, so a
|
|
@@ -88,6 +128,10 @@ export function resolveMaskCompositeSettings(data: Readonly<Record<string, unkno
|
|
|
88
128
|
invertMask: typeof data.invertMask === "boolean" ? data.invertMask : d.invertMask,
|
|
89
129
|
feather: clampInt(data.feather, b.FEATHER_MIN, b.FEATHER_MAX, d.feather),
|
|
90
130
|
grow: clampInt(data.grow, b.GROW_MIN, b.GROW_MAX, d.grow),
|
|
131
|
+
opacity: clampNumber(data.opacity, b.OPACITY_MIN, b.OPACITY_MAX, d.opacity),
|
|
132
|
+
fill: maskCompositeFillOf(data),
|
|
133
|
+
pixelSizePct: clampNumber(data.pixelSizePct, b.PIXEL_SIZE_PCT_MIN, b.PIXEL_SIZE_PCT_MAX, d.pixelSizePct),
|
|
134
|
+
blurStrengthPct: clampNumber(data.blurStrengthPct, b.BLUR_STRENGTH_PCT_MIN, b.BLUR_STRENGTH_PCT_MAX, d.blurStrengthPct),
|
|
91
135
|
backgroundFit: (MASK_COMPOSITE_BACKGROUND_FITS as readonly string[]).includes(data.backgroundFit as string)
|
|
92
136
|
? (data.backgroundFit as MaskCompositeBackgroundFit)
|
|
93
137
|
: d.backgroundFit,
|
|
@@ -97,10 +141,43 @@ export function resolveMaskCompositeSettings(data: Readonly<Record<string, unkno
|
|
|
97
141
|
}
|
|
98
142
|
}
|
|
99
143
|
|
|
144
|
+
/** The fill a node, payload or request body asks for: an unknown value is `video` (the default). */
|
|
145
|
+
export function maskCompositeFillOf(data: Readonly<Record<string, unknown>> | null | undefined): MaskCompositeFill {
|
|
146
|
+
const f = data?.fill
|
|
147
|
+
return (MASK_COMPOSITE_FILLS as readonly string[]).includes(f as string) ? (f as MaskCompositeFill) : MASK_COMPOSITE_DEFAULTS.fill
|
|
148
|
+
}
|
|
149
|
+
|
|
150
|
+
/** Does this fill take a foreground video? Only `video` does; pixelate and blur fill from the background. */
|
|
151
|
+
export function maskCompositeUsesForeground(fill: MaskCompositeFill): boolean {
|
|
152
|
+
return fill === "video"
|
|
153
|
+
}
|
|
154
|
+
|
|
155
|
+
/**
|
|
156
|
+
* The input whose size, rate and length the result has (and so the one a run
|
|
157
|
+
* is measured and priced on): the canvas handle — `video` (the foreground) for
|
|
158
|
+
* the video fill, `background-video` for pixelate and blur. The one rule the
|
|
159
|
+
* route, the workflow run, the run estimates and the length walks read.
|
|
160
|
+
*/
|
|
161
|
+
export function maskCompositeLeadHandle(data: Readonly<Record<string, unknown>> | null | undefined): "video" | "background-video" {
|
|
162
|
+
return maskCompositeUsesForeground(maskCompositeFillOf(data)) ? "video" : "background-video"
|
|
163
|
+
}
|
|
164
|
+
|
|
165
|
+
/**
|
|
166
|
+
* The sound a render actually carries. With no foreground (pixelate / blur) it
|
|
167
|
+
* is ALWAYS the background's — the only track there is — whatever
|
|
168
|
+
* `audioSource` says (the canvas hides that control for these fills, so a
|
|
169
|
+
* stored `none` from the Video fill can never silently mute the result).
|
|
170
|
+
*/
|
|
171
|
+
export function maskCompositeEffectiveAudioSource(s: Pick<MaskCompositeSettings, "fill" | "audioSource">): MaskCompositeAudioSource {
|
|
172
|
+
if (maskCompositeUsesForeground(s.fill)) return s.audioSource
|
|
173
|
+
return "background"
|
|
174
|
+
}
|
|
175
|
+
|
|
100
176
|
/** `POST /v1/mask-composite` request body. */
|
|
101
177
|
export interface MaskCompositeRequest extends Partial<MaskCompositeSettings> {
|
|
102
|
-
/** The video whose white-masked region is kept (its size, rate and length lead).
|
|
103
|
-
|
|
178
|
+
/** The video whose white-masked region is kept (its size, rate and length lead).
|
|
179
|
+
* Required when `fill` is `video` (the default); not used by `pixelate` / `blur`. */
|
|
180
|
+
readonly foregroundUrl?: string
|
|
104
181
|
/** The video shown where the mask is black. */
|
|
105
182
|
readonly backgroundUrl: string
|
|
106
183
|
/** The mask video — white = foreground (Video Mask's output). */
|
|
@@ -0,0 +1,73 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* The two price parts of a Meta Ads or Instagram scrape with AI analysis while
|
|
3
|
+
* metered LLM pricing is on: the scrape's own fixed id (no analysis suffix), and
|
|
4
|
+
* the requested items, each analysed by one call of the analysis model — the
|
|
5
|
+
* item count and model both routes hold through `reserveFixedPlusLlmCalls`
|
|
6
|
+
* (the analysis feature `meta-ads-analysis`, both scrapes). ONE rule for the
|
|
7
|
+
* server's estimate and the editor's, so the two can never price a scrape
|
|
8
|
+
* apart. Pure; null for a scrape without analysis and for any other node.
|
|
9
|
+
*/
|
|
10
|
+
import { LLM_FEATURE_DEFAULTS, llmPriceKey } from "./llm-models.js"
|
|
11
|
+
import { unitQuoteCreditId } from "./credit-estimators/unit-quote.js"
|
|
12
|
+
import {
|
|
13
|
+
META_ADS_SCRAPE_DEFAULT_COUNT,
|
|
14
|
+
META_ADS_SCRAPE_MAX_COUNT,
|
|
15
|
+
META_ADS_SCRAPE_MAX_SOURCES,
|
|
16
|
+
buildMetaAdsScrapeCreditId,
|
|
17
|
+
metaAdsAnalysisTierFrom,
|
|
18
|
+
metaAdsScrapeSources,
|
|
19
|
+
} from "./meta-ads-scrape.js"
|
|
20
|
+
import {
|
|
21
|
+
INSTAGRAM_SCRAPE_DEFAULT_COUNT,
|
|
22
|
+
INSTAGRAM_SCRAPE_MAX_COUNT,
|
|
23
|
+
buildInstagramScrapeCreditId,
|
|
24
|
+
instagramAnalysisTierFrom,
|
|
25
|
+
instagramRequestedCount,
|
|
26
|
+
instagramScrapeMode,
|
|
27
|
+
instagramScrapeSources,
|
|
28
|
+
} from "./instagram-scrape.js"
|
|
29
|
+
|
|
30
|
+
/** The analysis feature both scrapes bill their per-item analysis under. */
|
|
31
|
+
export const SCRAPE_ANALYSIS_FEATURE = "meta-ads-analysis"
|
|
32
|
+
|
|
33
|
+
export interface ScrapeAnalysisPriceParts {
|
|
34
|
+
/** The scrape's own fixed credit id (no analysis suffix). */
|
|
35
|
+
readonly fixedId: string
|
|
36
|
+
/** The model every item's analysis call runs on. */
|
|
37
|
+
readonly analysisModel: string
|
|
38
|
+
/** How many analysis calls the run holds: the items requested. */
|
|
39
|
+
readonly items: number
|
|
40
|
+
}
|
|
41
|
+
|
|
42
|
+
const clamp = (n: number, max: number) => Math.min(Math.max(Math.trunc(n) || 1, 1), max)
|
|
43
|
+
|
|
44
|
+
export function scrapeAnalysisPriceParts(node: { readonly type?: string; readonly data?: Record<string, unknown> }): ScrapeAnalysisPriceParts | null {
|
|
45
|
+
const data = node.data ?? {}
|
|
46
|
+
const analysisModel = typeof data.analysisModel === "string" && data.analysisModel ? data.analysisModel : LLM_FEATURE_DEFAULTS[SCRAPE_ANALYSIS_FEATURE]
|
|
47
|
+
if (node.type === "meta-ads-scrape" && metaAdsAnalysisTierFrom(data)) {
|
|
48
|
+
const count = typeof data.count === "number" ? data.count : META_ADS_SCRAPE_DEFAULT_COUNT
|
|
49
|
+
const sources = metaAdsScrapeSources(data)
|
|
50
|
+
return { fixedId: buildMetaAdsScrapeCreditId({ count, sources }), analysisModel, items: clamp(count, META_ADS_SCRAPE_MAX_COUNT) * clamp(sources, META_ADS_SCRAPE_MAX_SOURCES) }
|
|
51
|
+
}
|
|
52
|
+
if (node.type === "instagram-scrape" && instagramAnalysisTierFrom(data)) {
|
|
53
|
+
// The count the route accepts (its Zod: 1..INSTAGRAM_SCRAPE_MAX_COUNT), as the run sends it.
|
|
54
|
+
const stored = typeof data.count === "number" ? data.count : INSTAGRAM_SCRAPE_DEFAULT_COUNT
|
|
55
|
+
const count = instagramRequestedCount(instagramScrapeMode(data.mode as string | undefined), clamp(stored, INSTAGRAM_SCRAPE_MAX_COUNT))
|
|
56
|
+
const sources = instagramScrapeSources(data)
|
|
57
|
+
return { fixedId: buildInstagramScrapeCreditId({ count, sources }), analysisModel, items: Math.max(1, count * sources) }
|
|
58
|
+
}
|
|
59
|
+
return null
|
|
60
|
+
}
|
|
61
|
+
|
|
62
|
+
/**
|
|
63
|
+
* The whole analysis of a scrape as ONE price id: `<analysis key>:<items>u`
|
|
64
|
+
* (`unitQuoteCreditId` on the analysis model's metered price key). The server
|
|
65
|
+
* prices it as the run is held: the key's estimate × the items, rounded up
|
|
66
|
+
* once, marked up once at the call's charge id — so a client quotes the
|
|
67
|
+
* scrape's fixed id plus this, never the served one-call estimate × the items
|
|
68
|
+
* (rounded up per call, and marked up per call). Answered as an estimate while
|
|
69
|
+
* metered pricing is on; `missing` under the rollback.
|
|
70
|
+
*/
|
|
71
|
+
export function scrapeAnalysisQuoteId(parts: ScrapeAnalysisPriceParts): string {
|
|
72
|
+
return unitQuoteCreditId(llmPriceKey(SCRAPE_ANALYSIS_FEATURE, parts.analysisModel), parts.items)
|
|
73
|
+
}
|
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* The model a Generate Script call runs on, in one place: an explicit
|
|
3
|
+
* `llmModel`, else a legacy script `provider` alias (nodes saved before the
|
|
4
|
+
* per-model picker), else the feature's default. The route validates against
|
|
5
|
+
* it, the worker runs it, and the server's estimate and the editor's price it —
|
|
6
|
+
* one rule, so none of them can name a different model.
|
|
7
|
+
*/
|
|
8
|
+
import { LLM_FEATURE_DEFAULTS } from "./llm-models.js"
|
|
9
|
+
import type { ScriptProvider } from "./model-constants.js"
|
|
10
|
+
|
|
11
|
+
/** Legacy provider names → the LLM model ids they ran. */
|
|
12
|
+
export const SCRIPT_LEGACY_PROVIDER_MODELS: Readonly<Record<ScriptProvider, string>> = {
|
|
13
|
+
gemini: "gemini-3.6-flash",
|
|
14
|
+
claude: "claude-sonnet-4.6",
|
|
15
|
+
gpt: "gpt-5.2",
|
|
16
|
+
}
|
|
17
|
+
|
|
18
|
+
export function resolveScriptModelId(provider?: string, llmModel?: string): string {
|
|
19
|
+
const legacy = provider && Object.hasOwn(SCRIPT_LEGACY_PROVIDER_MODELS, provider) ? SCRIPT_LEGACY_PROVIDER_MODELS[provider as ScriptProvider] : undefined
|
|
20
|
+
return llmModel || legacy || LLM_FEATURE_DEFAULTS["generate-script"]
|
|
21
|
+
}
|
package/src/video-mask.ts
CHANGED
|
@@ -23,10 +23,39 @@ export const VIDEO_MASK_PROVIDERS = ["sam-3.1", "sam-3"] as const
|
|
|
23
23
|
export type VideoMaskProvider = (typeof VIDEO_MASK_PROVIDERS)[number]
|
|
24
24
|
export const DEFAULT_VIDEO_MASK_PROVIDER: VideoMaskProvider = "sam-3.1"
|
|
25
25
|
|
|
26
|
+
/**
|
|
27
|
+
* `zone`: keep only the subjects whose bounding-box CENTRE lies inside an
|
|
28
|
+
* inner rectangle, given as margins in % of the frame (each
|
|
29
|
+
* `VIDEO_MASK_LIMITS.ZONE_MARGIN_PCT_MIN`–`ZONE_MARGIN_PCT_MAX`). All zero, or
|
|
30
|
+
* no zone, = off.
|
|
31
|
+
*/
|
|
32
|
+
export interface VideoMaskZone {
|
|
33
|
+
readonly leftPct: number
|
|
34
|
+
readonly rightPct: number
|
|
35
|
+
readonly topPct: number
|
|
36
|
+
readonly bottomPct: number
|
|
37
|
+
}
|
|
38
|
+
|
|
26
39
|
/** A new Video Mask node's settings — spread into its node definition. */
|
|
27
40
|
export const VIDEO_MASK_DEFAULTS = {
|
|
28
41
|
threshold: 0.5,
|
|
29
42
|
provider: "sam-3.1" as VideoMaskProvider,
|
|
43
|
+
/** Off: the mask is the model's, frame for frame (see `HOLD_EMPTY_BELOW_FRACTION`). */
|
|
44
|
+
holdLastMask: false,
|
|
45
|
+
/** 0 = off: every region the model finds is kept (see `MIN_SUBJECT_HEIGHT_PCT_*`). */
|
|
46
|
+
minSubjectHeightPct: 0,
|
|
47
|
+
/** 0 = off: no ceiling on a region's height (see `MAX_SUBJECT_HEIGHT_PCT_*`). */
|
|
48
|
+
maxSubjectHeightPct: 0,
|
|
49
|
+
/** Off: subjects anywhere in the frame are kept (see `ZONE_MARGIN_PCT_*`). */
|
|
50
|
+
zone: null as VideoMaskZone | null,
|
|
51
|
+
/**
|
|
52
|
+
* Off: one pass over the whole clip. On: the clip's scene cuts are found
|
|
53
|
+
* first and each shot is segmented on its own — the model starts afresh
|
|
54
|
+
* after every cut (finding the new shot's subjects), and `holdLastMask`
|
|
55
|
+
* never carries a mask across a cut. A clip with no cut (a flash is not
|
|
56
|
+
* one) is one shot: the same mask, fitted to the clip's exact frame count.
|
|
57
|
+
*/
|
|
58
|
+
perShot: false,
|
|
30
59
|
} as const
|
|
31
60
|
|
|
32
61
|
export const VIDEO_MASK_LIMITS = {
|
|
@@ -42,6 +71,41 @@ export const VIDEO_MASK_LIMITS = {
|
|
|
42
71
|
THRESHOLD_STEP: 0.05,
|
|
43
72
|
/** Default detection confidence. Lower finds more (and less precise) regions. */
|
|
44
73
|
THRESHOLD_DEFAULT: VIDEO_MASK_DEFAULTS.threshold,
|
|
74
|
+
/**
|
|
75
|
+
* `holdLastMask`: a frame whose mask covers less than this fraction of the
|
|
76
|
+
* picture (0.002 % — about 41 px at 1080p: a stray speck, never a subject) counts as empty — the subject lost (hidden behind a
|
|
77
|
+
* hand, turned away) — and shows the most recent non-empty frame's mask
|
|
78
|
+
* instead. Frames before the first detection stay as they are.
|
|
79
|
+
*/
|
|
80
|
+
HOLD_EMPTY_BELOW_FRACTION: 0.00002,
|
|
81
|
+
/**
|
|
82
|
+
* `minSubjectHeightPct`: drop every connected white region (8-connected)
|
|
83
|
+
* whose bounding box is shorter than this % of the frame height — extras'
|
|
84
|
+
* small faces behind the main subject. 0 = off. Applied before
|
|
85
|
+
* `holdLastMask`, so a frame left with only small regions counts as empty
|
|
86
|
+
* and is held from the last frame with a big subject.
|
|
87
|
+
*/
|
|
88
|
+
MIN_SUBJECT_HEIGHT_PCT_MIN: 0,
|
|
89
|
+
MIN_SUBJECT_HEIGHT_PCT_MAX: 100,
|
|
90
|
+
/**
|
|
91
|
+
* `maxSubjectHeightPct`, the mirror: drop every region whose bounding box is
|
|
92
|
+
* TALLER than this % of the frame height (keep only the small subjects — a
|
|
93
|
+
* paste-back mask of the background faces). 0 = off. With both set, a
|
|
94
|
+
* region is kept when min ≤ height ≤ max (min > max is refused). Same pass,
|
|
95
|
+
* same order (before `holdLastMask`).
|
|
96
|
+
*/
|
|
97
|
+
MAX_SUBJECT_HEIGHT_PCT_MIN: 0,
|
|
98
|
+
MAX_SUBJECT_HEIGHT_PCT_MAX: 100,
|
|
99
|
+
/**
|
|
100
|
+
* `zone` margins (left / right / top / bottom, % of the frame width or
|
|
101
|
+
* height): drop every connected region whose bounding-box centre lies
|
|
102
|
+
* outside `[w·left/100, w·(1−right/100)] × [h·top/100, h·(1−bottom/100)]`
|
|
103
|
+
* — the faces at the edges of the picture. At most 49 % a side, so the
|
|
104
|
+
* inner rectangle is never empty. Applied after the size cut and before
|
|
105
|
+
* `holdLastMask`, so a frame left with nothing inside counts as empty.
|
|
106
|
+
*/
|
|
107
|
+
ZONE_MARGIN_PCT_MIN: 0,
|
|
108
|
+
ZONE_MARGIN_PCT_MAX: 49,
|
|
45
109
|
} as const
|
|
46
110
|
|
|
47
111
|
const VIDEO_MASK_BUCKETS_SEC = [5, 10, 15, 30, 60, 90, 120, 180, 240, 300] as const
|
|
@@ -84,6 +148,23 @@ export function resolveVideoMaskProvider(value: unknown): VideoMaskProvider {
|
|
|
84
148
|
: DEFAULT_VIDEO_MASK_PROVIDER
|
|
85
149
|
}
|
|
86
150
|
|
|
151
|
+
/**
|
|
152
|
+
* A zone as the region filter reads it: each margin a finite number clamped
|
|
153
|
+
* to `ZONE_MARGIN_PCT_MIN`–`ZONE_MARGIN_PCT_MAX` (anything else is 0), and
|
|
154
|
+
* null when there is no zone or every margin is 0 (off). The one reading
|
|
155
|
+
* shared by the worker, a workflow run's payload and the node's own run.
|
|
156
|
+
*/
|
|
157
|
+
export function resolveVideoMaskZone(value: unknown): VideoMaskZone | null {
|
|
158
|
+
if (!value || typeof value !== "object") return null
|
|
159
|
+
const v = value as Record<string, unknown>
|
|
160
|
+
const margin = (x: unknown) =>
|
|
161
|
+
typeof x === "number" && Number.isFinite(x)
|
|
162
|
+
? Math.min(VIDEO_MASK_LIMITS.ZONE_MARGIN_PCT_MAX, Math.max(VIDEO_MASK_LIMITS.ZONE_MARGIN_PCT_MIN, x))
|
|
163
|
+
: 0
|
|
164
|
+
const zone = { leftPct: margin(v.leftPct), rightPct: margin(v.rightPct), topPct: margin(v.topPct), bottomPct: margin(v.bottomPct) }
|
|
165
|
+
return zone.leftPct > 0 || zone.rightPct > 0 || zone.topPct > 0 || zone.bottomPct > 0 ? zone : null
|
|
166
|
+
}
|
|
167
|
+
|
|
87
168
|
/** `POST /v1/video-mask` request body. */
|
|
88
169
|
export interface VideoMaskRequest {
|
|
89
170
|
/** The clip to segment. */
|
|
@@ -93,4 +174,20 @@ export interface VideoMaskRequest {
|
|
|
93
174
|
/** Detection confidence, `VIDEO_MASK_LIMITS.THRESHOLD_MIN`–`THRESHOLD_MAX` (default 0.5). */
|
|
94
175
|
readonly threshold?: number
|
|
95
176
|
readonly provider?: VideoMaskProvider
|
|
177
|
+
/** Hold the last mask where the subject is lost (default false): an empty frame
|
|
178
|
+
* after the first detection repeats the most recent non-empty one. */
|
|
179
|
+
readonly holdLastMask?: boolean
|
|
180
|
+
/** Ignore subjects smaller than this % of the frame height, 0–100 (default 0 = off):
|
|
181
|
+
* connected regions whose bounding box is shorter are dropped from every frame. */
|
|
182
|
+
readonly minSubjectHeightPct?: number
|
|
183
|
+
/** Ignore subjects larger than this % of the frame height, 0–100 (default 0 = off):
|
|
184
|
+
* connected regions whose bounding box is taller are dropped from every frame. */
|
|
185
|
+
readonly maxSubjectHeightPct?: number
|
|
186
|
+
/** Keep only the subjects whose bounding-box centre is inside these margins, each
|
|
187
|
+
* 0–49 % of the frame (a margin left out is 0). null / omitted / all 0 = off. */
|
|
188
|
+
readonly zone?: Partial<VideoMaskZone> | null
|
|
189
|
+
/** Segment each shot on its own (default false): the clip's scene cuts are found
|
|
190
|
+
* first and the subjects are detected again in every shot; `holdLastMask` never
|
|
191
|
+
* carries a mask across a cut. The output then carries `shots`. */
|
|
192
|
+
readonly perShot?: boolean
|
|
96
193
|
}
|