Buckets:
| import"../chunks/DsnmJJEf.js";import{i as U,h as F,C as B,H as t,a as K,D as o,E as O,s as P}from"../chunks/BLbO27j4.js";import{p as C,o as q,s as e,f as X,a as S,b as Y,c as r,d as R,n as s,r as n}from"../chunks/jDjavuwI.js";const H='{"title":"Kandinsky 6 VAEs","local":"kandinsky-6-vaes","sections":[{"title":"Kandinsky6SRVAE","local":"diffusers.Kandinsky6SRVAE","sections":[],"depth":2},{"title":"MMAudioVAE","local":"diffusers.MMAudioVAE","sections":[],"depth":2},{"title":"MMAudioVocoder","local":"diffusers.MMAudioVocoder","sections":[],"depth":2}],"depth":1}';var Q=R('<meta name="hf:doc:metadata"/>'),$=R(`<p></p> <!> <!> <p>Kandinsky 6 uses a causal 3D K-VAE for video super-resolution and the MMAudio mel-spectrogram VAE, paired with a | |
| separate BigVGAN <a href="/docs/diffusers/pr_14521/en/api/models/kandinsky6_vae#diffusers.MMAudioVocoder">MMAudioVocoder</a>, for synchronized audio generation.</p> <!> <p>The causal 3D K-VAE used by <a href="/docs/diffusers/pr_14521/en/api/pipelines/kandinsky6#diffusers.Kandinsky6SRPipeline">Kandinsky6SRPipeline</a>. It processes arbitrarily long videos in bounded-memory | |
| segments while reproducing the exact output of a single, non-segmented pass.</p> <!> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>Causal 3D K-VAE used by <a href="/docs/diffusers/pr_14521/en/api/pipelines/kandinsky6#diffusers.Kandinsky6SRPipeline">Kandinsky6SRPipeline</a> to encode and decode video.</p> <p>Videos are processed in temporal segments of 16 pixel frames (plus the leading frame). The causal convolutions | |
| carry their padding state between segments, so the segmentation only bounds peak memory and does not change the | |
| result.</p> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>Encode a video into its latent distribution.</p></div> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>Decode latents into a video.</p></div> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!></div></div> <!> <p>The mel-spectrogram VAE used by <a href="/docs/diffusers/pr_14521/en/api/pipelines/kandinsky6#diffusers.Kandinsky6TI2VAPipeline">Kandinsky6TI2VAPipeline</a> when <code>sample_audio=True</code>. Its <code>decode</code> output is a mel | |
| spectrogram; pass it through <a href="/docs/diffusers/pr_14521/en/api/models/kandinsky6_vae#diffusers.MMAudioVocoder">MMAudioVocoder</a> to get a waveform.</p> <p>The reference implementation can be found at <a href="https://github.com/hkchengrex/MMAudio" rel="nofollow">hkchengrex/MMAudio</a> (MIT | |
| license).</p> <!> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>Audio VAE of <a href="/docs/diffusers/pr_14521/en/api/pipelines/kandinsky6#diffusers.Kandinsky6TI2VAPipeline">Kandinsky6TI2VAPipeline</a>: a magnitude-preserving autoencoder over log-mel spectrograms (MMAudio, <a href="https://arxiv.org/abs/2412.15322" rel="nofollow">https://arxiv.org/abs/2412.15322</a>).</p> <p><code>encode</code> turns a waveform into a latent distribution; <code>decode</code> turns latents back into a mel spectrogram, which <a href="/docs/diffusers/pr_14521/en/api/models/kandinsky6_vae#diffusers.MMAudioVocoder">MMAudioVocoder</a> then turns into a waveform. One latent frame covers <code>hop_length * 2</code> samples.</p> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>Encode a waveform into its latent distribution.</p></div> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>Decode latents into a mel spectrogram.</p></div> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!></div></div> <!> <p>Adapted from the BigVGAN-v2 vocoder MMAudio bundles, itself from <a href="https://github.com/NVIDIA/BigVGAN" rel="nofollow">NVIDIA/BigVGAN</a> (MIT license), with the anti-aliased Snake activations of <a href="https://github.com/junjun3518/alias-free-torch" rel="nofollow">alias-free-torch</a> (Apache License 2.0).</p> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>BigVGAN-v2 vocoder (<a href="https://github.com/NVIDIA/BigVGAN" rel="nofollow">https://github.com/NVIDIA/BigVGAN</a>, MIT license) with the anti-aliased Snake activations of <a href="https://github.com/junjun3518/alias-free-torch" rel="nofollow">https://github.com/junjun3518/alias-free-torch</a> (Apache 2.0), turning the mel spectrograms <a href="/docs/diffusers/pr_14521/en/api/models/kandinsky6_vae#diffusers.MMAudioVAE">MMAudioVAE</a> decodes | |
| into waveforms for <a href="/docs/diffusers/pr_14521/en/api/pipelines/kandinsky6#diffusers.Kandinsky6TI2VAPipeline">Kandinsky6TI2VAPipeline</a>.</p> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!></div></div> <!> <p></p>`,1);function ne(z,I){C(I,!1),q(()=>{new URLSearchParams(window.location.search).get("fw")}),U();var f=$();F("18u55ck",E=>{var T=Q();P(T,"content",H),S(E,T)});var m=e(X(f),2);B(m,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"});var h=e(m,2);t(h,{title:"Kandinsky 6 VAEs",local:"kandinsky-6-vaes",headingTag:"h1"});var _=e(h,4);t(_,{title:"Kandinsky6SRVAE",local:"diffusers.Kandinsky6SRVAE",headingTag:"h2"});var g=e(_,4);K(g,{code:"aW1wb3J0JTIwdG9yY2glMEFmcm9tJTIwZGlmZnVzZXJzJTIwaW1wb3J0JTIwS2FuZGluc2t5NlNSVkFFJTBBJTBBdmFlJTIwJTNEJTIwS2FuZGluc2t5NlNSVkFFLmZyb21fcHJldHJhaW5lZCglMEElMjAlMjAlMjAlMjAlMjJrYW5kaW5za3lsYWIlMkZLYW5kaW5za3ktNi4wLVZTUi01cy1EaWZmdXNlcnMlMjIlMkMlMjBzdWJmb2xkZXIlM0QlMjJ2YWUlMjIlMkMlMjB0b3JjaF9kdHlwZSUzRHRvcmNoLmJmbG9hdDE2JTBBKQ==",highlighted:`<span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> Kandinsky6SRVAE | |
| vae = Kandinsky6SRVAE.from_pretrained( | |
| <span class="hljs-string">"kandinskylab/Kandinsky-6.0-VSR-5s-Diffusers"</span>, subfolder=<span class="hljs-string">"vae"</span>, torch_dtype=torch.bfloat16 | |
| )`,lang:"python",wrap:!1});var a=e(g,2),b=r(a);o(b,{name:"class diffusers.Kandinsky6SRVAE",anchor:"diffusers.Kandinsky6SRVAE",source:"https://github.com/huggingface/diffusers/blob/vr_14521/src/diffusers/models/autoencoders/autoencoder_kandinsky6_sr.py#L516",parameters:[{name:"in_channels",val:": int = 3"},{name:"out_channels",val:": int = 3"},{name:"latent_channels",val:": int = 64"},{name:"encoder_block_out_channels",val:": tuple[int, ...] = (16, 128, 256, 512, 1024)"},{name:"decoder_block_out_channels",val:": tuple[int, ...] = (16, 256, 512, 1024, 2048)"},{name:"layers_per_block",val:": int = 2"},{name:"temporal_compression_ratio",val:": int = 4"},{name:"temporal_compression_start_level",val:": int = 1"},{name:"scaling_factor",val:": float = 0.910344004631042"}],parametersDescription:[{anchor:"diffusers.Kandinsky6SRVAE.in_channels",description:`<strong>in_channels</strong> (<code>int</code>, defaults to <code>3</code>) — | |
| Number of pixel channels.`,name:"in_channels"},{anchor:"diffusers.Kandinsky6SRVAE.out_channels",description:`<strong>out_channels</strong> (<code>int</code>, defaults to <code>3</code>) — | |
| Number of reconstructed pixel channels.`,name:"out_channels"},{anchor:"diffusers.Kandinsky6SRVAE.latent_channels",description:`<strong>latent_channels</strong> (<code>int</code>, defaults to <code>64</code>) — | |
| Number of latent channels.`,name:"latent_channels"},{anchor:"diffusers.Kandinsky6SRVAE.encoder_block_out_channels",description:`<strong>encoder_block_out_channels</strong> (<code>tuple[int, ...]</code>, defaults to <code>(16, 128, 256, 512, 1024)</code>) — | |
| Output width of the residual blocks at each encoder level; every level but the last halves the spatial size | |
| and doubles the width on the way to the next level.`,name:"encoder_block_out_channels"},{anchor:"diffusers.Kandinsky6SRVAE.decoder_block_out_channels",description:`<strong>decoder_block_out_channels</strong> (<code>tuple[int, ...]</code>, defaults to <code>(16, 256, 512, 1024, 2048)</code>) — | |
| Output width of the residual blocks at each decoder level.`,name:"decoder_block_out_channels"},{anchor:"diffusers.Kandinsky6SRVAE.layers_per_block",description:`<strong>layers_per_block</strong> (<code>int</code>, defaults to <code>2</code>) — | |
| Number of residual blocks per encoder level; the decoder uses one more per level.`,name:"layers_per_block"},{anchor:"diffusers.Kandinsky6SRVAE.temporal_compression_ratio",description:`<strong>temporal_compression_ratio</strong> (<code>int</code>, defaults to <code>4</code>) — | |
| Temporal compression factor; <code>log2</code> of it consecutive levels also compress time.`,name:"temporal_compression_ratio"},{anchor:"diffusers.Kandinsky6SRVAE.temporal_compression_start_level",description:`<strong>temporal_compression_start_level</strong> (<code>int</code>, defaults to <code>1</code>) — | |
| First level that compresses time.`,name:"temporal_compression_start_level"},{anchor:"diffusers.Kandinsky6SRVAE.scaling_factor",description:`<strong>scaling_factor</strong> (<code>float</code>, defaults to <code>0.910344</code>) — | |
| Scale applied to the latents before they enter the diffusion transformer.`,name:"scaling_factor"}]});var d=e(b,6),J=r(d);o(J,{name:"encode",anchor:"diffusers.Kandinsky6SRVAE.encode",source:"https://github.com/huggingface/diffusers/blob/vr_14521/src/diffusers/models/autoencoders/autoencoder_kandinsky6_sr.py#L585",parameters:[{name:"x",val:": torch.Tensor"},{name:"return_dict",val:": bool = True"}],parametersDescription:[{anchor:"diffusers.Kandinsky6SRVAE.encode.x",description:`<strong>x</strong> (<code>torch.Tensor</code> of shape <code>(batch_size, channels, num_frames, height, width)</code>) — | |
| Pixel video in <code>[-1, 1]</code>. <code>num_frames</code> should be <code>1 + k * temporal_compression_ratio</code>.`,name:"x"},{anchor:"diffusers.Kandinsky6SRVAE.encode.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, defaults to <code>True</code>) — | |
| Whether to return an <a href="/docs/diffusers/pr_14521/en/api/models/autoencoderkl_qwenimage21#diffusers.models.modeling_outputs.AutoencoderKLOutput">AutoencoderKLOutput</a> instead of a plain tuple.`,name:"return_dict"}]}),s(2),n(d);var i=e(d,2),D=r(i);o(D,{name:"decode",anchor:"diffusers.Kandinsky6SRVAE.decode",source:"https://github.com/huggingface/diffusers/blob/vr_14521/src/diffusers/models/autoencoders/autoencoder_kandinsky6_sr.py#L611",parameters:[{name:"z",val:": torch.Tensor"},{name:"return_dict",val:": bool = True"}],parametersDescription:[{anchor:"diffusers.Kandinsky6SRVAE.decode.z",description:`<strong>z</strong> (<code>torch.Tensor</code> of shape <code>(batch_size, latent_channels, num_latent_frames, height, width)</code>) — | |
| Latents, already divided by <code>scaling_factor</code>.`,name:"z"},{anchor:"diffusers.Kandinsky6SRVAE.decode.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, defaults to <code>True</code>) — | |
| Whether to return a <code>~models.autoencoder_kl.DecoderOutput</code> instead of a plain tuple.`,name:"return_dict"}]}),s(2),n(i);var v=e(i,2),N=r(v);o(N,{name:"forward",anchor:"diffusers.Kandinsky6SRVAE.forward",source:"https://github.com/huggingface/diffusers/blob/vr_14521/src/diffusers/models/autoencoders/autoencoder_kandinsky6_sr.py#L642",parameters:[{name:"sample",val:": torch.Tensor"},{name:"sample_posterior",val:": bool = False"},{name:"return_dict",val:": bool = True"},{name:"generator",val:": torch.Generator | None = None"}],parametersDescription:[{anchor:"diffusers.Kandinsky6SRVAE.forward.sample",description:`<strong>sample</strong> (<code>torch.Tensor</code> of shape <code>(batch_size, channels, num_frames, height, width)</code>) — | |
| Pixel video in <code>[-1, 1]</code>. <code>num_frames</code> should be <code>1 + k * temporal_compression_ratio</code>.`,name:"sample"},{anchor:"diffusers.Kandinsky6SRVAE.forward.sample_posterior",description:`<strong>sample_posterior</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) — | |
| Whether to sample from the latent posterior instead of using its mode.`,name:"sample_posterior"},{anchor:"diffusers.Kandinsky6SRVAE.forward.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) — | |
| Whether to return a <code>~models.autoencoder_kl.DecoderOutput</code> instead of a plain tuple.`,name:"return_dict"},{anchor:"diffusers.Kandinsky6SRVAE.forward.generator",description:`<strong>generator</strong> (<code>torch.Generator</code>, <em>optional</em>) — | |
| A <a href="https://pytorch.org/docs/stable/generated/torch.Generator.html" rel="nofollow"><code>torch.Generator</code></a> to make sampling | |
| deterministic.`,name:"generator"}],returnDescription:`<script context="module">export const metadata = 'undefined';<\/script> | |
| <p>If <code>return_dict</code> is True, a <code>~models.autoencoder_kl.DecoderOutput</code> is returned, otherwise a plain | |
| <code>tuple</code> is returned. Its <code>sample</code> is the reconstructed video.</p> | |
| `,returnType:`<script context="module">export const metadata = 'undefined';<\/script> | |
| <p><code>~models.autoencoder_kl.DecoderOutput</code> or <code>tuple</code></p> | |
| `}),n(v),n(a);var k=e(a,2);t(k,{title:"MMAudioVAE",local:"diffusers.MMAudioVAE",headingTag:"h2"});var M=e(k,6);K(M,{code:"aW1wb3J0JTIwdG9yY2glMEFmcm9tJTIwZGlmZnVzZXJzJTIwaW1wb3J0JTIwTU1BdWRpb1ZBRSUwQSUwQWF1ZGlvX3ZhZSUyMCUzRCUyME1NQXVkaW9WQUUuZnJvbV9wcmV0cmFpbmVkKCUwQSUyMCUyMCUyMCUyMCUyMmthbmRpbnNreWxhYiUyRkthbmRpbnNreS02LjAtUHJvLWRpc3RpbGwtNXMtRGlmZnVzZXJzJTIyJTJDJTIwc3ViZm9sZGVyJTNEJTIyYXVkaW9fdmFlJTIyJTJDJTIwdG9yY2hfZHR5cGUlM0R0b3JjaC5iZmxvYXQxNiUwQSk=",highlighted:`<span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> MMAudioVAE | |
| audio_vae = MMAudioVAE.from_pretrained( | |
| <span class="hljs-string">"kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers"</span>, subfolder=<span class="hljs-string">"audio_vae"</span>, torch_dtype=torch.bfloat16 | |
| )`,lang:"python",wrap:!1});var c=e(M,2),A=r(c);o(A,{name:"class diffusers.MMAudioVAE",anchor:"diffusers.MMAudioVAE",source:"https://github.com/huggingface/diffusers/blob/vr_14521/src/diffusers/models/autoencoders/autoencoder_mmaudio.py#L350",parameters:[{name:"mel_bins",val:": int = 128"},{name:"latent_channels",val:": int = 40"},{name:"hidden_channels",val:": int = 512"},{name:"channel_multipliers",val:": tuple[int, ...] = (1, 2, 4)"},{name:"layers_per_block",val:": int = 2"},{name:"sample_rate",val:": int = 44100"},{name:"n_fft",val:": int = 2048"},{name:"hop_length",val:": int = 512"},{name:"scaling_factor",val:": float = 0.417"}],parametersDescription:[{anchor:"diffusers.MMAudioVAE.mel_bins",description:`<strong>mel_bins</strong> (<code>int</code>, defaults to <code>128</code>) — | |
| Number of mel bins.`,name:"mel_bins"},{anchor:"diffusers.MMAudioVAE.latent_channels",description:`<strong>latent_channels</strong> (<code>int</code>, defaults to <code>40</code>) — | |
| Number of latent channels.`,name:"latent_channels"},{anchor:"diffusers.MMAudioVAE.hidden_channels",description:`<strong>hidden_channels</strong> (<code>int</code>, defaults to <code>512</code>) — | |
| Base width of the autoencoder.`,name:"hidden_channels"},{anchor:"diffusers.MMAudioVAE.channel_multipliers",description:`<strong>channel_multipliers</strong> (<code>tuple[int, ...]</code>, defaults to <code>(1, 2, 4)</code>) — | |
| Width multipliers of the autoencoder levels.`,name:"channel_multipliers"},{anchor:"diffusers.MMAudioVAE.layers_per_block",description:`<strong>layers_per_block</strong> (<code>int</code>, defaults to <code>2</code>) — | |
| Residual blocks per encoder level; the decoder uses one more per level.`,name:"layers_per_block"},{anchor:"diffusers.MMAudioVAE.sample_rate",description:`<strong>sample_rate</strong> (<code>int</code>, defaults to <code>44100</code>) — | |
| Waveform sample rate.`,name:"sample_rate"},{anchor:"diffusers.MMAudioVAE.n_fft",description:`<strong>n_fft</strong> (<code>int</code>, defaults to <code>2048</code>) — | |
| FFT size of the mel front end.`,name:"n_fft"},{anchor:"diffusers.MMAudioVAE.hop_length",description:`<strong>hop_length</strong> (<code>int</code>, defaults to <code>512</code>) — | |
| Hop length of the mel front end. Must match the total upsampling factor of the <a href="/docs/diffusers/pr_14521/en/api/models/kandinsky6_vae#diffusers.MMAudioVocoder">MMAudioVocoder</a> this VAE | |
| is paired with.`,name:"hop_length"},{anchor:"diffusers.MMAudioVAE.scaling_factor",description:`<strong>scaling_factor</strong> (<code>float</code>, defaults to <code>0.417</code>) — | |
| Scale applied to the latents before they enter the diffusion transformer.`,name:"scaling_factor"}]});var l=e(A,6),W=r(l);o(W,{name:"encode",anchor:"diffusers.MMAudioVAE.encode",source:"https://github.com/huggingface/diffusers/blob/vr_14521/src/diffusers/models/autoencoders/autoencoder_mmaudio.py#L403",parameters:[{name:"audio",val:": torch.Tensor"},{name:"return_dict",val:": bool = True"}],parametersDescription:[{anchor:"diffusers.MMAudioVAE.encode.audio",description:`<strong>audio</strong> (<code>torch.Tensor</code> of shape <code>(batch_size, num_samples)</code>) — | |
| Mono waveform in <code>[-1, 1]</code> at <code>sample_rate</code>.`,name:"audio"},{anchor:"diffusers.MMAudioVAE.encode.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, defaults to <code>True</code>) — | |
| Whether to return an <a href="/docs/diffusers/pr_14521/en/api/models/autoencoderkl_qwenimage21#diffusers.models.modeling_outputs.AutoencoderKLOutput">AutoencoderKLOutput</a> instead of a plain tuple.`,name:"return_dict"}]}),s(2),n(l);var u=e(l,2),j=r(u);o(j,{name:"decode",anchor:"diffusers.MMAudioVAE.decode",source:"https://github.com/huggingface/diffusers/blob/vr_14521/src/diffusers/models/autoencoders/autoencoder_mmaudio.py#L420",parameters:[{name:"z",val:": torch.Tensor"},{name:"return_dict",val:": bool = True"}],parametersDescription:[{anchor:"diffusers.MMAudioVAE.decode.z",description:`<strong>z</strong> (<code>torch.Tensor</code> of shape <code>(batch_size, latent_channels, num_latent_frames)</code>) — | |
| Latents, already divided by <code>scaling_factor</code>.`,name:"z"},{anchor:"diffusers.MMAudioVAE.decode.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, defaults to <code>True</code>) — | |
| Whether to return a <code>~models.autoencoder_kl.DecoderOutput</code> instead of a plain tuple.`,name:"return_dict"}],returnDescription:`<script context="module">export const metadata = 'undefined';<\/script> | |
| <p>The mel spectrogram of shape <code>(batch_size, mel_bins, num_mel_frames)</code>, ready for <a | |
| href="/docs/diffusers/pr_14521/en/api/models/kandinsky6_vae#diffusers.MMAudioVocoder" | |
| >MMAudioVocoder</a>.</p> | |
| `}),s(2),n(u);var y=e(u,2),G=r(y);o(G,{name:"forward",anchor:"diffusers.MMAudioVAE.forward",source:"https://github.com/huggingface/diffusers/blob/vr_14521/src/diffusers/models/autoencoders/autoencoder_mmaudio.py#L439",parameters:[{name:"sample",val:": torch.Tensor"},{name:"sample_posterior",val:": bool = False"},{name:"return_dict",val:": bool = True"},{name:"generator",val:": torch.Generator | None = None"}],parametersDescription:[{anchor:"diffusers.MMAudioVAE.forward.sample",description:`<strong>sample</strong> (<code>torch.Tensor</code> of shape <code>(batch_size, num_samples)</code>) — | |
| Mono waveform in <code>[-1, 1]</code> at <code>sample_rate</code> to encode and reconstruct as a mel spectrogram.`,name:"sample"},{anchor:"diffusers.MMAudioVAE.forward.sample_posterior",description:`<strong>sample_posterior</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) — | |
| Whether to sample from the latent posterior instead of using its mode.`,name:"sample_posterior"},{anchor:"diffusers.MMAudioVAE.forward.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) — | |
| Whether to return a <code>~models.autoencoder_kl.DecoderOutput</code> instead of a plain tuple.`,name:"return_dict"},{anchor:"diffusers.MMAudioVAE.forward.generator",description:`<strong>generator</strong> (<code>torch.Generator</code>, <em>optional</em>) — | |
| A <a href="https://pytorch.org/docs/stable/generated/torch.Generator.html" rel="nofollow"><code>torch.Generator</code></a> to make sampling | |
| deterministic.`,name:"generator"}],returnDescription:`<script context="module">export const metadata = 'undefined';<\/script> | |
| <p>If <code>return_dict</code> is True, a <code>~models.autoencoder_kl.DecoderOutput</code> is returned, otherwise a plain | |
| <code>tuple</code> is returned. Its <code>sample</code> is the reconstructed mel spectrogram of shape <code>(batch_size, mel_bins, num_mel_frames)</code>.</p> | |
| `,returnType:`<script context="module">export const metadata = 'undefined';<\/script> | |
| <p><code>~models.autoencoder_kl.DecoderOutput</code> or <code>tuple</code></p> | |
| `}),n(y),n(c);var V=e(c,2);t(V,{title:"MMAudioVocoder",local:"diffusers.MMAudioVocoder",headingTag:"h2"});var p=e(V,4),x=r(p);o(x,{name:"class diffusers.MMAudioVocoder",anchor:"diffusers.MMAudioVocoder",source:"https://github.com/huggingface/diffusers/blob/vr_14521/src/diffusers/models/autoencoders/mmaudio_vocoder.py#L154",parameters:[{name:"num_mels",val:": int = 128"},{name:"upsample_initial_channel",val:": int = 1536"},{name:"upsample_rates",val:": tuple[int, ...] = (8, 4, 2, 2, 2, 2)"},{name:"upsample_kernel_sizes",val:": tuple[int, ...] = (16, 8, 4, 4, 4, 4)"},{name:"resblock_kernel_sizes",val:": tuple[int, ...] = (3, 7, 11)"},{name:"resblock_dilation_sizes",val:": tuple[tuple[int, ...], ...] = ((1, 3, 5), (1, 3, 5), (1, 3, 5))"}],parametersDescription:[{anchor:"diffusers.MMAudioVocoder.num_mels",description:`<strong>num_mels</strong> (<code>int</code>, defaults to <code>128</code>) — | |
| Number of mel bins of the input spectrogram. Must match the paired <a href="/docs/diffusers/pr_14521/en/api/models/kandinsky6_vae#diffusers.MMAudioVAE">MMAudioVAE</a>’s <code>mel_bins</code>.`,name:"num_mels"},{anchor:"diffusers.MMAudioVocoder.upsample_initial_channel",description:`<strong>upsample_initial_channel</strong> (<code>int</code>, defaults to <code>1536</code>) — | |
| Width of the first layer.`,name:"upsample_initial_channel"},{anchor:"diffusers.MMAudioVocoder.upsample_rates",description:`<strong>upsample_rates</strong> (<code>tuple[int, ...]</code>, defaults to <code>(8, 4, 2, 2, 2, 2)</code>) — | |
| Upsampling factors of the vocoder stages. Their product is the total upsampling factor and must match the | |
| paired <a href="/docs/diffusers/pr_14521/en/api/models/kandinsky6_vae#diffusers.MMAudioVAE">MMAudioVAE</a>’s <code>hop_length</code>.`,name:"upsample_rates"},{anchor:"diffusers.MMAudioVocoder.upsample_kernel_sizes",description:`<strong>upsample_kernel_sizes</strong> (<code>tuple[int, ...]</code>, defaults to <code>(16, 8, 4, 4, 4, 4)</code>) — | |
| Transposed-convolution kernel sizes of the vocoder stages.`,name:"upsample_kernel_sizes"},{anchor:"diffusers.MMAudioVocoder.resblock_kernel_sizes",description:`<strong>resblock_kernel_sizes</strong> (<code>tuple[int, ...]</code>, defaults to <code>(3, 7, 11)</code>) — | |
| Kernel sizes of the residual blocks.`,name:"resblock_kernel_sizes"},{anchor:"diffusers.MMAudioVocoder.resblock_dilation_sizes",description:`<strong>resblock_dilation_sizes</strong> (<code>tuple[tuple[int, ...], ...]</code>, defaults to <code>((1, 3, 5), (1, 3, 5), (1, 3, 5))</code>) — | |
| Dilations of the residual blocks.`,name:"resblock_dilation_sizes"}]});var w=e(x,4),L=r(w);o(L,{name:"forward",anchor:"diffusers.MMAudioVocoder.forward",source:"https://github.com/huggingface/diffusers/blob/vr_14521/src/diffusers/models/autoencoders/mmaudio_vocoder.py#L213",parameters:[{name:"mel",val:": torch.Tensor"},{name:"return_dict",val:": bool = True"}],parametersDescription:[{anchor:"diffusers.MMAudioVocoder.forward.mel",description:`<strong>mel</strong> (<code>torch.Tensor</code> of shape <code>(batch_size, num_mels, num_mel_frames)</code>) — | |
| Mel spectrogram, as decoded by <a href="/docs/diffusers/pr_14521/en/api/models/kandinsky6_vae#diffusers.MMAudioVAE">MMAudioVAE</a>.`,name:"mel"},{anchor:"diffusers.MMAudioVocoder.forward.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, defaults to <code>True</code>) — | |
| Whether to return a <code>~models.autoencoder_kl.DecoderOutput</code> instead of a plain tuple.`,name:"return_dict"}],returnDescription:`<script context="module">export const metadata = 'undefined';<\/script> | |
| <p>The waveform of shape <code>(batch_size, 1, num_samples)</code> in <code>[-1, 1]</code>.</p> | |
| `}),n(w),n(p);var Z=e(p,2);O(Z,{source:"https://github.com/huggingface/diffusers/blob/main/docs/source/en/api/models/kandinsky6_vae.md"}),s(2),S(z,f),Y()}export{ne as component}; | |
Xet Storage Details
- Size:
- 25.3 kB
- Xet hash:
- 22f9b02f857b1ebc1daf8ae98defc860a4ea6b3732aa0ed6c4c3dd8724a11928
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.