The Xet storage method uses de-duplication and chunked downloads to speed up transfers in some situations:
https://pypi.org/project/hf-xet/
But most importantly, installing the Xet support gets rid of some annoying HuggingFace CLI messages about missing the feature.
- Some users have been confused and were manually creating and activating Python venvs, which is not good since it can lead to the wrong Python version or dependency conflicts.
- Therefore, we add more detailed guidance to explain that `uv` manages the whole environment, the Python version, all dependencies and automatic environment activation.
- A few users were also confused about where `uv tool` installs binaries, but instead of explaining that in depth, we now add a link to the documentation page which explains how it works, and also instruct users to carefully read the `uv tool` output since it tells them how to add the installation to the system's path.
- We can't use the original "Show experimental features" checkbox implementation, because it *deeply* breaks Gradio.
- Gradio's `gr.Examples()` API binds itself to the original state of the user interface. Gradio crashes and causes various bugs if we try to change the available UI controls later.
- Instead, we must use `gr.Dataset()` which acts like a custom input/output control and doesn't directly bind itself to the target control. We must also provide a secret, hidden "all mode choices" component so that it knows the names of all "control modes" that are possible in examples.
- We now also have a very visible warning label in the user interface, to clearly mark the experimental features.
- Bugs fixed:
* The code was unable to toggle the visibility of Experimental demos in the Examples list. It was not possible with Examples (since it's a wrapper around Dataset, but Examples contains its own internal state/copy of all data). Instead, we use a Dataset and manipulate its list directly.
* Gradio crashes with a `gradio.exceptions.Error` exception if you try to load an example that tries to use an experimental feature if we have removed its UI element. This is because Examples binds to the original user interface and *remembers* the list of choices, and it *cannot* dynamically select something that did not exist when the `gr.Examples()` was initially created. This problem is fixed by switching to `gr.Dataset()`.
* Furthermore, Gradio's `gr.Examples()` handler actually remembers and caches the list of UI options. So every time we load an example, it rewrites the "Emotion Control Mode" selection menu to only show the options that were available when the Examples table was created. This means that even if we keep the "Show experimental features" checkbox, Gradio itself will erase the experimental mode from the Control Mode selection menu every time the user loads an example. There are no callbacks or "update" functions to allow us to override this automatic Gradio behavior. But by switching to `gr.Dataset()`, we completely avoid this deep binding.
* The "Show experimental features" checkbox is no longer tied to a column in the examples-table, to avoid fighting between Gradio's example table trying to set the mode, and the experimental checkbox being toggled and also trying to set the mode.
* Lastly, the "Show experimental features" checkbox now remembers and restores the user's current mode selection when toggling the checkbox, instead of constantly resetting to the default mode ("same as voice reference"), to make the UI more convenient for users.
- The emotion weight is always applied in every mode except "Same as voice reference", so we must make the slider visible so that users can control the value. Otherwise it would silently apply the last-set value without the user knowing, which is very confusing.
- Furthermore, having the slider even on the Emotion Vectors page is *very* useful, because it allows users to rapidly change the total strength of the current emotion vectors without having to manually/carefully move every individual emotion slider.
- The Advanced Settings contains some very advanced features which users shouldn't tweak, but it also contains important insight into segmentation generations, and the "max tokens per generation segment" feature which users must tweak if they have low VRAM.
- Therefore it's very important that users notice the "Advanced Settings" section so that they can read the VRAM help text and reduce the segment length if they have VRAM issues. So let's make the advanced category visible by default again until a better solution is determined.
- The WebUI was secretly squashing all emotion vectors and re-scaling them. It's a good idea for user friendliness, but it makes it harder to learn what values will work in Python when using the WebUI for testing.
- Instead, let's move the normalization code into IndexTTS2 as a helper function which is used by Gradio and can be used from other people's code too.
- The emotion bias (which reduces the influence of certain emotions) has also been converted into an optional feature, which can be turned off if such biasing isn't wanted. And all biasing values have been re-scaled to use 1.0 as the reference, to avoid scaling relative to 0.8 (which previously meant that it applied double scaling).
- Silently scaling the value internally is confusing for users. They may be tuning their settings via the Web UI before putting the same values into their Python code, and would then get a different result since the Web UI "lies" about the slider values.
- Instead, let's remove the silent scaling, and just change the default weight to a better recommendation.
- We were using ".select" to detect when tabs are changed, but Gradio has modified behavior in 5.45.0 to only trigger from user clicks. They now require that we use ".change" to detect tab changes from code. This fix makes the Examples work when loading on new Gradio versions.
- This is a major new feature, which now allows for much more natural speech generation by lowering the influence of the emotion vector/text control modes.
- It is particularly useful for the "emotion text description" control mode, where a strength of 0.6 or lower is useful to get much more natural speech.
- Added support for `emo_alpha` scaling of emotion vectors and emotion text inputs.
- This is a major new feature, which now allows for much more natural speech generation by lowering the influence of the emotion vector/text control modes.
- It is particularly useful for the "emotion text description" control mode, where a strength of 0.6 or lower is useful to get much more natural speech. Before this feature, it was not possible to make natural speech with that mode, because QwenEmotion assigns emotion scores to the text from 0.0-1.0, and that score was used directly as an emotion vector. This meant that the text mode always used very high strengths. Now, the user can adjust the strength of the emotions to get very natural results.
- Refactored `IndexTTS2.infer()` variable initialization logic to avoid repetition and ensure cleaner code paths.
