Spaces:

AIBRUH
/

LAYER

Paused

Claude

Mark which stages are actually responsible for perceived realism

6b5c34f unverified 4 months ago

2.5 kB

export type ImplementationMode = "local" | "nim" | "unavailable";

export type Implementation = {
  mode: ImplementationMode;
  model: string;
  runtime?: string;
  precision?: string;
  latencyMs?: number;
  vramGb?: number;
  throughput?: string;
  notes?: string;
};

export type FailureModeId = 1 | 2 | 3 | 4 | 5;

export const FAILURE_MODES: Record<FailureModeId, { name: string; description: string }> = {
  1: {
    name: "The Uncanny Valley",
    description:
      "Stiff upper facial regions, unnatural lack of global head movement, severe temporal jitter across frames.",
  },
  2: {
    name: "The End-of-Speech Pop",
    description:
      "The generated mouth region visibly clashes in color/saturation with the original static image on reverting to silence.",
  },
  3: {
    name: "Cascaded Latency",
    description:
      "ASR → LLM → TTS → THG run fully serially, each adding sequential latency that caps responsiveness.",
  },
  4: {
    name: "VRAM Death",
    description: "Requires 32GB+ with no quantization path — measured, not theoretical.",
  },
  5: {
    name: "The VASA-1 Problem",
    description: "A brilliant claim with no repo and no reproduction. If there's no code, it doesn't count.",
  },
};

export type FailureVerdict = "pass" | "fail" | "at-risk" | "n/a";

export type FailureAudit = {
  mode: FailureModeId;
  verdict: FailureVerdict;
  note: string;
};

export type Tier = "A" | "B" | "C";

export type Confidence = "confirmed" | "inferred";

export type PipelineNode = {
  id: string;
  label: string;
  category:
    | "vision"
    | "asr"
    | "llm"
    | "tts"
    | "avatar"
    | "video"
    | "keyframer"
    | "orchestration";
  summary: string;
  handoffOut?: string;
  expertNote: string;
  implementations: Implementation[];
  children?: PipelineNode[];
  tier?: Tier;
  failureModeAudit?: FailureAudit[];
  confidence?: Confidence;
  // True for stages whose own output quality directly determines how human
  // this looks/sounds (Avatar, TTS, Video) — as opposed to stages that affect
  // correctness or latency but aren't themselves a realism bottleneck.
  realismCritical?: boolean;
};

export type EdgeKind = "data" | "hosts" | "loop";

export type PipelineEdge = {
  source: string;
  target: string;
  label?: string;
  kind: EdgeKind;
};

export type Pipeline = {
  id: string;
  name: string;
  description: string;
  sourceInput?: string;
  nodes: PipelineNode[];
  edges: PipelineEdge[];
  pipelineAudit?: FailureAudit[];
};

	export type ImplementationMode = "local" \| "nim" \| "unavailable";

	export type Implementation = {
	mode: ImplementationMode;
	model: string;
	runtime?: string;
	precision?: string;
	latencyMs?: number;
	vramGb?: number;
	throughput?: string;
	notes?: string;
	};

	export type FailureModeId = 1 \| 2 \| 3 \| 4 \| 5;

	export const FAILURE_MODES: Record<FailureModeId, { name: string; description: string }> = {
	1: {
	name: "The Uncanny Valley",
	description:
	"Stiff upper facial regions, unnatural lack of global head movement, severe temporal jitter across frames.",
	},
	2: {
	name: "The End-of-Speech Pop",
	description:
	"The generated mouth region visibly clashes in color/saturation with the original static image on reverting to silence.",
	},
	3: {
	name: "Cascaded Latency",
	description:
	"ASR → LLM → TTS → THG run fully serially, each adding sequential latency that caps responsiveness.",
	},
	4: {
	name: "VRAM Death",
	description: "Requires 32GB+ with no quantization path — measured, not theoretical.",
	},
	5: {
	name: "The VASA-1 Problem",
	description: "A brilliant claim with no repo and no reproduction. If there's no code, it doesn't count.",
	},
	};

	export type FailureVerdict = "pass" \| "fail" \| "at-risk" \| "n/a";

	export type FailureAudit = {
	mode: FailureModeId;
	verdict: FailureVerdict;
	note: string;
	};

	export type Tier = "A" \| "B" \| "C";

	export type Confidence = "confirmed" \| "inferred";

	export type PipelineNode = {
	id: string;
	label: string;
	category:
	\| "vision"
	\| "asr"
	\| "llm"
	\| "tts"
	\| "avatar"
	\| "video"
	\| "keyframer"
	\| "orchestration";
	summary: string;
	handoffOut?: string;
	expertNote: string;
	implementations: Implementation[];
	children?: PipelineNode[];
	tier?: Tier;
	failureModeAudit?: FailureAudit[];
	confidence?: Confidence;
	// True for stages whose own output quality directly determines how human
	// this looks/sounds (Avatar, TTS, Video) — as opposed to stages that affect
	// correctness or latency but aren't themselves a realism bottleneck.
	realismCritical?: boolean;
	};

	export type EdgeKind = "data" \| "hosts" \| "loop";

	export type PipelineEdge = {
	source: string;
	target: string;
	label?: string;
	kind: EdgeKind;
	};

	export type Pipeline = {
	id: string;
	name: string;
	description: string;
	sourceInput?: string;
	nodes: PipelineNode[];
	edges: PipelineEdge[];
	pipelineAudit?: FailureAudit[];
	};