Skip to main content

Function: getNTCPaperProfile()

getNTCPaperProfile(textureResolution): object

Paper architecture and full training budget. This is intentionally opt-in: 250k batches of 524288 samples are much more expensive than browser defaults. Both grids use uint4 here; this is not a claim of a particular table-2 bitrate.

Parameters​

textureResolution​

number

Returns​

object

baseResolution​

baseResolution: number

batchSize​

batchSize: number

cosineAnnealingScale​

cosineAnnealingScale: number = 0

dualGrid​

dualGrid: boolean = true

gridChannels​

gridChannels: number = 8

hiddenActivation​

hiddenActivation: string = 'hgelu'

hiddenSizes​

hiddenSizes: number[]

iterations​

iterations: number = 250000

learningRate​

learningRate: number = 0.01

levels​

levels: number

lowResChannels​

lowResChannels: number = 12

mipFilter​

mipFilter: "lanczos"

mipsPerLevel​

mipsPerLevel: number = 2

positionalEncoding​

positionalEncoding: boolean = true

positionalEncodingPeriod​

positionalEncodingPeriod: number = 8

quantization​

quantization: object

quantization.mode​

mode: string = 'uint4'

retrainAfterQuantize​

retrainAfterQuantize: number = 0.05

textureResolution​

textureResolution: number

weightsLearningRate​

weightsLearningRate: number = 0.005