Make version 4.0.x latest

Author: DarthSim

…ed_docs/version-4-pre/cache/external.mdx docs/cache/external.mdxversioned_docs/version-4-pre/cache/external.mdx renamed to docs/cache/external.mdx versioned_docs/version-4-pre/cache/external.mdx renamed to docs/cache/external.mdx

@@ -88,7 +88,11 @@ IMGPROXY_AUTOQUALITY_FORMAT_MAX="avif=65"
 
 ## Autoquality with ML
 
-This method is almost the same as autoquality with [DSSIM](#autoquality-by-dssim) but imgproxy will try to predict the initial quality using neural networks. This requires neural networks to be configured (see the config examlpe or the config documentation). If a neural network for the resulting format is not provided, the [DSSIM](#autoquality-by-dssim) method will be used instead.
+This method is almost the same as autoquality with [DSSIM](#autoquality-by-dssim), but imgproxy will try to predict the initial quality using neural network models. This requires configuring neural network models (see the config example or the config documentation). If a model for the resulting format is not provided, the [DSSIM](#autoquality-by-dssim) method will be used instead.
+
+:::tip
+You don't need to configure autoquality neural network models if you're using an imgproxy Pro Docker image with a tag suffixed with `-ml`. Such images already have the models for JPEG, WebP, AVIF, and JPEG XL included and configured.
+:::
 
 :::info
 When this method is used, imgproxy will save JPEG images with the most optimal [advanced JPEG compression](../configuration/options.mdx#advanced-jpeg-compression) settings, ignoring config and processing options.
@@ -111,15 +115,15 @@ IMGPROXY_AUTOQUALITY_MAX=80
 # Quality 63 for AVIF is pretty the same as 80 for JPEG
 IMGPROXY_AUTOQUALITY_FORMAT_MIN="avif=60"
 IMGPROXY_AUTOQUALITY_FORMAT_MAX="avif=65"
-# Neural networks paths for JPEG, WebP, AVIF, and JPEG XL
-IMGPROXY_AUTOQUALITY_JPEG_NET="/networks/autoquality-jpeg.pb"
-IMGPROXY_AUTOQUALITY_WEBP_NET="/networks/autoquality-webp.pb"
-IMGPROXY_AUTOQUALITY_AVIF_NET="/networks/autoquality-avif.pb"
-IMGPROXY_AUTOQUALITY_JXL_NET="/networks/autoquality-jxl.pb"
+# Neural network models paths for JPEG, WebP, AVIF, and JPEG XL
+IMGPROXY_AUTOQUALITY_JPEG_NET="/networks/autoquality-jpeg.onnx"
+IMGPROXY_AUTOQUALITY_WEBP_NET="/networks/autoquality-webp.onnx"
+IMGPROXY_AUTOQUALITY_AVIF_NET="/networks/autoquality-avif.onnx"
+IMGPROXY_AUTOQUALITY_JXL_NET="/networks/autoquality-jxl.onnx"
 ```
 
 :::info
-If you trust your neural network's autoquality, you may want to set `IMGPROXY_AUTOQUALITY_ALLOWED_ERROR` to 1 (the maximum possible DSSIM value). In this case, imgproxy will always use the quality predicted by the neural network.
+If you trust your neural network model's autoquality, you may want to set `IMGPROXY_AUTOQUALITY_ALLOWED_ERROR` to 1 (the maximum possible DSSIM value). In this case, imgproxy will always use the quality predicted by the neural network model.
 :::
 
 ### Processing options example
@@ -128,12 +132,11 @@ If you trust your neural network's autoquality, you may want to set `IMGPROXY_AU
 .../autoquality:ml:0.02:70:80:0.001/...
 ```
 
-## Neural networks format
+## Neural network models format
 
-Neural networks should fit the following requirements:
-* Tensorflow frozen graph format
-* Input layer size is 416x416
-* Output layer size is 1x100
-* Output layer values are logits of quality probabilities
+Neural network models should fit the following requirements:
+* ONNX format
+* Input layer size is 256x256x3 (RGB pixels normalized to [0..1])
+* Output layer size is 1 (the predicted quality value in the [0..1] range)
 
-If you're an imgproxy Pro user and you want to train your own network but you don't know how, feel free to contact the imgproxy team for instructions.
+If you're an imgproxy Pro user and you want to train your own model but you don't know how, feel free to contact the imgproxy team for instructions.