IdentityFeatureTransferFinal
This is the current feature-transfer implementation. It operates on Klein's attention output, separates generated and reference image tokens using the runtime referenceimagenumtokens metadata, and builds a masked reference bank from the selected references. The transfer performs: 1. Per-image centering of generated and reference features. 2. Normalized similarity matching. 3. Similarity-floor filtering. 4. Temperature-controlled reference pooling. 5. Confidence-gated transfer at the scheduled double and single blocks. | Parameter | Default | Meaning | |---|---:|---| | preset | HARDLOCK | HARDLOCK, MIDLOCK, and SOFTLOCK replace the manual similarity, temperature, mask threshold, and block schedules. Use custom to edit them directly. | | enabled | true | Returns an unmodified model clone when disabled. | | referenceindices | all | Zero-based references used by the transfer. Accepts all, comma-separated indices such as 0,2,3, or ranges such as 0-3. | | referenceindex | 0 | Fallback reference when referenceindices resolves to no valid entries. | | similarityfloor | 0.040 | Minimum centered similarity allowed to contribute. Higher values reject more matches. | | softmaxtemperature | 0.0250 | Match sharpness. Lower values concentrate on fewer reference tokens; higher values blend more candidates. | | maskthreshold | 1.00 | Minimum pooled mask value required for a reference token. White is included; black is excluded. | | doubleblocks | 0-7:midimg=0.55 | Per-double-block transfer strengths. | | singleblocks | tested sparse schedule | Per-single-block transfer strengths. Empty tex…
Conditioning/Flux2klein · 10 entradas · 11 parámetros · 1 salida
