You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: documentation/OPTIONS.es.md
+29-3Lines changed: 29 additions & 3 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -715,6 +715,7 @@ Muchas configuraciones se establecen a través del [dataloader config](DATALOADE
715
715
- Requiere que el modelo tenga un pipeline `IMG2IMG` o `IMG2VIDEO` registrado
716
716
- Puede combinarse con `--eval_dataset_id` para obtener imágenes de un dataset específico
717
717
- Para modelos i2v, permite usar un dataset de imágenes simple para validación sin la configuración compleja de emparejamiento de datasets de conditioning usada durante el entrenamiento
718
+
- Flux Kontext no usa este flag para validación; déjalo desactivado y usa `--eval_dataset_id` para elegir el dataset de edición mientras Kontext carga automáticamente su dataset de referencia emparejado
718
719
- La fuerza de des-ruido se controla con los ajustes normales de timestep de validación
719
720
720
721
### `--eval_dataset_id`
@@ -727,6 +728,7 @@ Muchas configuraciones se establecen a través del [dataloader config](DATALOADE
727
728
- El ID de dataset debe coincidir con un dataset configurado en tu config de dataloader
728
729
- Útil para mantener evaluación consistente usando un dataset de eval dedicado
729
730
- Para modelos de conditioning, los datos de conditioning del dataset (si existen) también se usarán
731
+
- Para Flux Kontext, esta es la forma correcta de seleccionar el dataset de validación; no habilites `--validation_using_datasets`
730
732
731
733
---
732
734
@@ -742,6 +744,7 @@ Algunos modelos no pueden funcionar sin entradas de conditioning:
742
744
-**Entrenamiento ControlNet**: Requiere imágenes de señal de control
743
745
744
746
Para estos modelos, un dataset de conditioning es obligatorio. La WebUI mostrará opciones de conditioning como requeridas, y el entrenamiento fallará sin ellas.
747
+
La validación de Flux Kontext también permanece en esta ruta basada en conditioning. Usa `--eval_dataset_id` para elegir el dataset de edición para validación y deja `--validation_using_datasets` desactivado.
745
748
746
749
### 2. Modelos que SOPORTAN Conditioning Opcional
747
750
@@ -774,6 +777,8 @@ Para estos modelos, PUEDES agregar datasets de conditioning pero no es obligator
774
777
775
778
**Modelos I2V con `--validation_using_datasets`**: Para modelos de video i2v (HunyuanVideo, WAN, Kandinsky5Video), habilitar este flag permite usar un dataset de imágenes simple para validación. Las imágenes se usan como entradas de conditioning de primer frame para generar videos de validación, sin necesidad de la configuración compleja de emparejamiento de datasets de conditioning.
776
779
780
+
**Flux Kontext con `--validation_using_datasets`**: No habilites este flag. Kontext es solo de edición y valida mediante su ruta normal de datasets emparejados de imagen + conditioning. Usa `--eval_dataset_id` para seleccionar el dataset de edición.
781
+
777
782
### Tipos de Datos de Conditioning
778
783
779
784
Diferentes modelos esperan diferentes datos de conditioning:
@@ -1098,11 +1103,32 @@ CREPA es una técnica de regularización para fine-tuning de modelos de difusió
1098
1103
-**Por qué**: Evita cargar DINOv2 cuando el backbone ya tiene una capa semántica más fuerte para supervisar.
1099
1104
-**Predeterminado**: `false`
1100
1105
1106
+
### `--crepa_feature_source`
1107
+
1108
+
-**Qué**: Selecciona de dónde obtiene CREPA su señal de maestro.
1109
+
-**Por qué**: Usa `encoder` para la ruta clásica con encoder externo, `backbone` para alineación interna bloque a bloque, o `self_flow` para el maestro EMA con vista más limpia usado por Self-Flow.
1110
+
-**Opciones**: `encoder`, `backbone`, `self_flow`
1111
+
-**Predeterminado**: `encoder`
1112
+
1113
+
### `--crepa_self_flow`
1114
+
1115
+
-**Qué**: Alias booleano legado que activa el modo Self-Flow.
1116
+
-**Por qué**: Las configuraciones antiguas aún pueden usarlo, pero las nuevas deberían preferir `crepa_feature_source=self_flow`.
1117
+
-**Predeterminado**: `false`
1118
+
-**Nota**: Entra en conflicto con `crepa_use_backbone_features` y con `crepa_feature_source` si apunta a otro modo.
1119
+
1120
+
### `--crepa_self_flow_mask_ratio`
1121
+
1122
+
-**Qué**: Fracción de tokens que reciben el timestep alternativo en Self-Flow.
1123
+
-**Por qué**: Controla cuánta asimetría de información se introduce entre tokens más limpios y más ruidosos. Valores altos fortalecen la señal auto-supervisada pero pueden desestabilizar el entrenamiento.
1124
+
-**Predeterminado**: `0.1`
1125
+
-**Rango**: `0.0` a `0.5`
1126
+
1101
1127
### `--crepa_teacher_block_index`
1102
1128
1103
-
-**Qué**: Índice del bloque maestro al usar características del backbone.
1104
-
-**Por qué**: Te permite alinear un bloque estudiante temprano con un bloque maestro más profundo sin un encoder externo. Si no se establece, cae en el bloque estudiante.
1105
-
-**Predeterminado**: Usa `crepa_block_index` si no se proporciona.
1129
+
-**Qué**: Índice del bloque maestro al usar características del backbone o Self-Flow.
1130
+
-**Por qué**: Te permite alinear un bloque estudiante temprano con un bloque maestro más profundo sin un encoder externo. Self-Flow lo requiere explícitamente para que el profesor EMA lea una capa semántica más profunda.
1131
+
-**Predeterminado**: Usa `crepa_block_index` si no se proporciona en modo backbone; es obligatorio en modo Self-Flow.
Copy file name to clipboardExpand all lines: documentation/OPTIONS.hi.md
+29-3Lines changed: 29 additions & 3 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -713,6 +713,7 @@ Alternative attention mechanisms समर्थित हैं, जिनक
713
713
- Model में `IMG2IMG` या `IMG2VIDEO` pipeline registered होना चाहिए
714
714
-`--eval_dataset_id` के साथ combine कर सकते हैं specific dataset से images लेने के लिए
715
715
- i2v models के लिए, यह training में use होने वाली complex conditioning dataset pairing setup के बिना simple image dataset validation के लिए use करने देता है
716
+
- Flux Kontext validation के लिए इस flag का उपयोग नहीं करता; इसे disabled रखें और edit dataset चुनने के लिए `--eval_dataset_id` उपयोग करें, जबकि Kontext paired reference dataset अपने आप load कर लेता है
716
717
- Denoising strength normal validation timestep settings से control होती है
717
718
718
719
### `--eval_dataset_id`
@@ -725,6 +726,7 @@ Alternative attention mechanisms समर्थित हैं, जिनक
725
726
- Dataset ID आपके dataloader config में configured dataset से match होना चाहिए
726
727
- Dedicated eval dataset use करके consistent evaluation maintain करने के लिए useful
727
728
- Conditioning models के लिए, dataset का conditioning data (यदि हो) भी use होगा
729
+
- Flux Kontext के लिए validation dataset चुनने का सही तरीका यही है; `--validation_using_datasets` enable न करें
728
730
729
731
---
730
732
@@ -740,6 +742,7 @@ SimpleTuner conditioning inputs (reference images, control signals, आदि) u
740
742
-**ControlNet training**: Control signal images require करता है
741
743
742
744
इन models के लिए, conditioning dataset mandatory है। WebUI conditioning options को required दिखाएगी, और training इनके बिना fail होगी।
745
+
Flux Kontext validation भी इसी conditioning-based path पर रहती है। Validation के लिए edit dataset चुनने हेतु `--eval_dataset_id` उपयोग करें और `--validation_using_datasets` disabled रखें।
743
746
744
747
### 2. Models जो Optional Conditioning SUPPORT करते हैं
745
748
@@ -772,6 +775,8 @@ SimpleTuner conditioning inputs (reference images, control signals, आदि) u
772
775
773
776
**I2V models के साथ `--validation_using_datasets`**: i2v video models (HunyuanVideo, WAN, Kandinsky5Video) के लिए, यह flag enable करने पर validation के लिए simple image dataset use कर सकते हैं। Images validation videos generate करने के लिए first-frame conditioning inputs के रूप में use होती हैं, complex conditioning dataset pairing setup की जरूरत के बिना।
774
777
778
+
**Flux Kontext के साथ `--validation_using_datasets`**: यह flag enable न करें। Kontext edit-only है और validation अपनी सामान्य paired image + conditioning dataset path से चलाता है। इसके बजाय edit dataset चुनने के लिए `--eval_dataset_id` उपयोग करें।
779
+
775
780
### Conditioning Data Types
776
781
777
782
Different models different conditioning data expect करते हैं:
@@ -1096,11 +1101,32 @@ CREPA एक regularization तकनीक है जो video diffusion models
1096
1101
-**Why**: जब backbone के पास पहले से मजबूत semantic layer हो, तब DINOv2 लोड करने से बचता है।
1097
1102
-**Default**: `false`
1098
1103
1104
+
### `--crepa_feature_source`
1105
+
1106
+
-**What**: चुनता है कि CREPA अपना teacher signal कहां से ले।
1107
+
-**Why**: classic external encoder path के लिए `encoder`, internal block-to-block alignment के लिए `backbone`, और Self-Flow के cleaner EMA teacher view के लिए `self_flow` उपयोग करें।
1108
+
-**Choices**: `encoder`, `backbone`, `self_flow`
1109
+
-**Default**: `encoder`
1110
+
1111
+
### `--crepa_self_flow`
1112
+
1113
+
-**What**: Self-Flow mode enable करने वाला legacy boolean alias।
1114
+
-**Why**: पुराने configs अब भी इसका उपयोग कर सकते हैं, लेकिन नए configs को `crepa_feature_source=self_flow` उपयोग करना चाहिए।
1115
+
-**Default**: `false`
1116
+
-**Note**: यह `crepa_use_backbone_features` और किसी दूसरे mode वाले `crepa_feature_source` के साथ conflict करता है।
1117
+
1118
+
### `--crepa_self_flow_mask_ratio`
1119
+
1120
+
-**What**: Self-Flow में alternate timestep पाने वाले tokens का fraction।
1121
+
-**Why**: यह cleaner और noisier tokens के बीच information asymmetry नियंत्रित करता है। ऊंचे values self-supervised signal को मजबूत करते हैं, लेकिन training unstable भी कर सकते हैं।
1122
+
-**Default**: `0.1`
1123
+
-**Range**: `0.0` से `0.5`
1124
+
1099
1125
### `--crepa_teacher_block_index`
1100
1126
1101
-
-**What**: backbone features उपयोग करते समय teacher block index।
1102
-
-**Why**: external encoder के बिना, earlier student block को later teacher block से align करने देता है। unset होने पर student block पर fallback होता है।
1103
-
-**Default**: यदि नहीं दिया गया, तो `crepa_block_index` उपयोग होगा।
1127
+
-**What**: backbone features या Self-Flow उपयोग करते समय teacher block index।
1128
+
-**Why**: external encoder के बिना earlier student block को later teacher block से align करने देता है। Self-Flow में यह explicitly जरूरी है ताकि EMA teacher deeper semantic layer से पढ़े।
1129
+
-**Default**: backbone mode में नहीं दिया गया तो `crepa_block_index` उपयोग होगा; Self-Flow mode में यह required है।
0 commit comments