MGIEOptions

Options & Configuration · AiDotNet.VisionLanguage.Editing

Configuration options for MGIE: MLLM-guided image editing with LLaVA-based instruction understanding.

Inheritance: ModelOptions → GenerativeVLMOptions → EditingVLMOptions → MGIEOptions

For Beginners

These options configure the MGIE model. Default values follow the original paper settings.

Constructors

Constructor Summary
MGIEOptions() Initializes a new instance with default values.
MGIEOptions(MGIEOptions other) Initializes a new instance by copying from another instance.

Properties

Property Summary
LanguageModelTrainableScope TrainableLanguageModelScope { get; set; } Which parts of the instruction language model training updates.
bool EnableExpressiveInstructions { get; set; } Gets or sets whether string editing requests first generate an expressive continuation with the native MLLM.
double ConditionDropoutProbability { get; set; } Probability of dropping a training example’s conditioning, for classifier-free guidance.
double EditLossWeight { get; set; } Weight of the edit (diffusion noise-prediction) loss in the total objective.
double ImageGuidanceScale { get; set; } Image-only classifier-free guidance scale; the text scale is GuidanceScale.
double InstructionLossWeight { get; set; } Weight of the instruction (language-model) loss in the total objective.
int EditHiddenDim { get; set; } Hidden width of the edit mapper, 512 in the released MGIE implementation.
int EditNumHeads { get; set; } Attention heads in the mapper, independent of the language model’s heads.
int EditQueryCount { get; set; } Learned output queries, 77 in the released MGIE implementation.
int EditTokenCount { get; set; } Learned edit tokens appended to the joint image/instruction stream.
int VisionPatchSize { get; set; } Vision encoder patch size; independent of the diffusion output resolution.