Buckets:
| import"../chunks/DsnmJJEf.js";import{i as L,h as P,C as K,H as v,a as k,D as o,E as J,s as z}from"../chunks/X8cAGukj.js";import{p as j,o as D,s as e,f as U,a as b,b as q,c as a,d as w,n as t,r as n}from"../chunks/CQ2f15a1.js";const E='{"title":"AutoencoderKLPH3Image","local":"autoencoderklph3image","sections":[{"title":"AutoencoderKLPH3Image","local":"diffusers.AutoencoderKLPH3Image","sections":[],"depth":2}],"depth":1}';var N=w('<meta name="hf:doc:metadata"/>'),V=w(`<p></p> <!> <!> <p>The image VAE of <a href="https://huggingface.co/PrunaAI/P-H3-Image" rel="nofollow">P-H3-Image</a> by Pruna AI. It uses the encoder and latent space of MiniMax-H3’s <a href="./autoencoderkl_minimax_h3">video VAE</a>, with a ViT decoder that decodes a single latent frame into an image.</p> <p>It follows the conventions of <a href="/docs/diffusers/pr_15212/en/api/models/autoencoderkl_minimax_h3#diffusers.AutoencoderKLMiniMaxH3">AutoencoderKLMiniMaxH3</a>:</p> <ul><li>Images are passed as single-frame videos, <code>(batch_size, channels, 1, height, width)</code>, with 16x spatial compression.</li> <li>Latents are normalized per channel with <code>latents_mean</code> / <code>latents_std</code> instead of a <code>scaling_factor</code>.</li> <li>Pixels are ImageNet-normalized RGB over <code>[0, 1]</code>: <code>encode</code> expects <code>(pixel - imagenet_mean) / imagenet_std</code> and <code>decode</code> returns that space.</li> <li>Spatial tiling is on by default, with 256 pixel tiles and at least 64 pixels of overlap.</li></ul> <!> <p>The encoder and the decoder’s output projection stay in float32 whatever <code>dtype</code> the model is loaded in.</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>The image VAE of <a href="https://huggingface.co/PrunaAI/P-H3-Image" rel="nofollow">P-H3-Image</a>: MiniMax-H3’s VAE encoder and latent | |
| space, with a <code>PH3ImageViTDecoder</code> that decodes a single latent frame.</p> <p>Images are passed as single-frame videos, <code>(batch_size, channels, 1, height, width)</code>. As in <a href="/docs/diffusers/pr_15212/en/api/models/autoencoderkl_minimax_h3#diffusers.AutoencoderKLMiniMaxH3">AutoencoderKLMiniMaxH3</a>, pixels are ImageNet-normalized RGB over <code>[0, 1]</code> (mean <code>(0.485, 0.456, 0.406)</code>, std <code>(0.229, 0.224, 0.225)</code>) and spatial tiling is on by default.</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 batch of images into latents.</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 a batch of latent images.</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>Enable tiled VAE encoding/decoding. When this option is enabled, the VAE splits the frames into tiles, encodes | |
| or decodes each tile separately and linearly blends the overlaps back together. This lowers the memory | |
| requirement and allows processing larger frames.</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>Encode then decode a batch of images.</p></div></div> <!> <p></p>`,1);function C(I,y){j(y,!1),D(()=>{new URLSearchParams(window.location.search).get("fw")}),L();var l=V();P("c22the",g=>{var f=N();z(f,"content",E),b(g,f)});var c=e(U(l),2);K(c,{containerStyle:"float: inline-end; margin-inline-start: 10px; display: inline-flex; position: relative; z-index: 10;"});var m=e(c,2);v(m,{title:"AutoencoderKLPH3Image",local:"autoencoderklph3image",headingTag:"h1"});var p=e(m,8);k(p,{code:"aW1wb3J0JTIwdG9yY2glMEFmcm9tJTIwZGlmZnVzZXJzJTIwaW1wb3J0JTIwQXV0b2VuY29kZXJLTFBIM0ltYWdlJTBBJTBBdmFlJTIwJTNEJTIwQXV0b2VuY29kZXJLTFBIM0ltYWdlLmZyb21fcHJldHJhaW5lZCglMjJQcnVuYUFJJTJGUC1IMy1JbWFnZSUyMiUyQyUyMHN1YmZvbGRlciUzRCUyMnZhZSUyMiUyQyUyMGR0eXBlJTNEdG9yY2guYmZsb2F0MTYpLnRvKCUwQSUyMCUyMCUyMCUyMCUyMmN1ZGElMjIlMEEpJTIwJTIwJTIzJTIwb3IlMjAlMjJtcHMlMjIlMkMlMjAlMjJ4cHUlMjIlMkMlMjAlMjJjcHUlMjI=",highlighted:`<span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoencoderKLPH3Image | |
| vae = AutoencoderKLPH3Image.from_pretrained(<span class="hljs-string">"PrunaAI/P-H3-Image"</span>, subfolder=<span class="hljs-string">"vae"</span>, dtype=torch.bfloat16).to( | |
| <span class="hljs-string">"cuda"</span> | |
| ) <span class="hljs-comment"># or "mps", "xpu", "cpu"</span>`,lang:"python",wrap:!1});var u=e(p,4);v(u,{title:"AutoencoderKLPH3Image",local:"diffusers.AutoencoderKLPH3Image",headingTag:"h2"});var r=e(u,2),h=a(r);o(h,{name:"class diffusers.AutoencoderKLPH3Image",anchor:"diffusers.AutoencoderKLPH3Image",source:"https://github.com/huggingface/diffusers/blob/vr_15212/src/diffusers/models/autoencoders/autoencoder_kl_p_h3_image.py#L493",parameters:[{name:"in_channels",val:": int = 3"},{name:"out_channels",val:": int = 3"},{name:"latent_channels",val:": int = 24"},{name:"block_out_channels",val:": tuple = (128, 256, 256, 512, 512, 1024)"},{name:"layers_per_block",val:": int = 2"},{name:"spatial_downsample_factors",val:": tuple = (2, 2, 2, 2, 1, 1)"},{name:"temporal_downsample_factors",val:": tuple = (1, 2, 2, 1, 1, 1)"},{name:"norm_num_groups",val:": int = 32"},{name:"norm_eps",val:": float = 1e-06"},{name:"spatial_padding_mode",val:": str = 'reflect'"},{name:"decoder_num_layers",val:": int = 36"},{name:"decoder_num_attention_heads",val:": int = 32"},{name:"decoder_attention_head_dim",val:": int = 64"},{name:"decoder_num_register_tokens",val:": int = 4"},{name:"decoder_ffn_mult",val:": int = 4"},{name:"decoder_rope_theta",val:": float = 100.0"},{name:"decoder_rope_dim_ratio",val:": float = 0.75"},{name:"decoder_norm_eps",val:": float = 1e-05"},{name:"latents_mean",val:": tuple = (0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0)"},{name:"latents_std",val:": tuple = (1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0)"}]});var s=e(h,6),T=a(s);o(T,{name:"encode",anchor:"diffusers.AutoencoderKLPH3Image.encode",source:"https://github.com/huggingface/diffusers/blob/vr_15212/src/diffusers/models/autoencoders/autoencoder_kl_p_h3_image.py#L728",parameters:[{name:"x",val:": Tensor"},{name:"return_dict",val:": bool = True"}],parametersDescription:[{anchor:"diffusers.AutoencoderKLPH3Image.encode.x",description:`<strong>x</strong> (<code>torch.Tensor</code>) — | |
| Input batch of images as single-frame videos, shape <code>(batch_size, in_channels, 1, height, width)</code>.`,name:"x"},{anchor:"diffusers.AutoencoderKLPH3Image.encode.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) — | |
| Whether to return a <a href="/docs/diffusers/pr_15212/en/api/models/autoencoderkl_qwenimage#diffusers.models.modeling_outputs.AutoencoderKLOutput">AutoencoderKLOutput</a> instead of a plain | |
| tuple.`,name:"return_dict"}],returnDescription:` | |
| <p>The latent distribution of the encoded images. Note that the latents are normalized with <code>latents_mean</code> / | |
| <code>latents_std</code> afterwards.</p> | |
| `}),t(2),n(s);var i=e(s,2),x=a(i);o(x,{name:"decode",anchor:"diffusers.AutoencoderKLPH3Image.decode",source:"https://github.com/huggingface/diffusers/blob/vr_15212/src/diffusers/models/autoencoders/autoencoder_kl_p_h3_image.py#L757",parameters:[{name:"z",val:": Tensor"},{name:"return_dict",val:": bool = True"}],parametersDescription:[{anchor:"diffusers.AutoencoderKLPH3Image.decode.z",description:`<strong>z</strong> (<code>torch.Tensor</code>) — | |
| Input batch of latents, shape <code>(batch_size, latent_channels, 1, height, width)</code>.`,name:"z"},{anchor:"diffusers.AutoencoderKLPH3Image.decode.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) — | |
| Whether to return a <a href="/docs/diffusers/pr_15212/en/api/models/autoencoderkl_qwenimage#diffusers.models.autoencoders.vae.DecoderOutput">DecoderOutput</a> instead of a plain tuple.`,name:"return_dict"}],returnDescription:` | |
| <p>The decoded images, shape <code>(batch_size, out_channels, 1, height, width)</code>.</p> | |
| `,returnType:` | |
| <p><a | |
| href="/docs/diffusers/pr_15212/en/api/models/autoencoderkl_qwenimage#diffusers.models.autoencoders.vae.DecoderOutput" | |
| >DecoderOutput</a> or <code>tuple</code></p> | |
| `}),t(2),n(i);var d=e(i,2),M=a(d);o(M,{name:"enable_tiling",anchor:"diffusers.AutoencoderKLPH3Image.enable_tiling",source:"https://github.com/huggingface/diffusers/blob/vr_15212/src/diffusers/models/autoencoders/autoencoder_kl_p_h3_image.py#L575",parameters:[{name:"tile_sample_min_height",val:": int | None = None"},{name:"tile_sample_min_width",val:": int | None = None"},{name:"tile_sample_min_overlap_height",val:": int | None = None"},{name:"tile_sample_min_overlap_width",val:": int | None = None"}],parametersDescription:[{anchor:"diffusers.AutoencoderKLPH3Image.enable_tiling.tile_sample_min_height",description:`<strong>tile_sample_min_height</strong> (<code>int</code>, <em>optional</em>) — | |
| The tile height in pixel space. Frames taller than this are split along the height dimension.`,name:"tile_sample_min_height"},{anchor:"diffusers.AutoencoderKLPH3Image.enable_tiling.tile_sample_min_width",description:`<strong>tile_sample_min_width</strong> (<code>int</code>, <em>optional</em>) — | |
| The tile width in pixel space. Frames wider than this are split along the width dimension.`,name:"tile_sample_min_width"},{anchor:"diffusers.AutoencoderKLPH3Image.enable_tiling.tile_sample_min_overlap_height",description:`<strong>tile_sample_min_overlap_height</strong> (<code>int</code>, <em>optional</em>) — | |
| The minimum overlap, in pixels, between two consecutive vertical tiles.`,name:"tile_sample_min_overlap_height"},{anchor:"diffusers.AutoencoderKLPH3Image.enable_tiling.tile_sample_min_overlap_width",description:`<strong>tile_sample_min_overlap_width</strong> (<code>int</code>, <em>optional</em>) — | |
| The minimum overlap, in pixels, between two consecutive horizontal tiles.`,name:"tile_sample_min_overlap_width"}]}),t(2),n(d);var _=e(d,2),A=a(_);o(A,{name:"forward",anchor:"diffusers.AutoencoderKLPH3Image.forward",source:"https://github.com/huggingface/diffusers/blob/vr_15212/src/diffusers/models/autoencoders/autoencoder_kl_p_h3_image.py#L784",parameters:[{name:"sample",val:": Tensor"},{name:"sample_posterior",val:": bool = False"},{name:"generator",val:": typing.Optional[torch.Generator] = None"},{name:"return_dict",val:": bool = True"}],parametersDescription:[{anchor:"diffusers.AutoencoderKLPH3Image.forward.sample",description:`<strong>sample</strong> (<code>torch.Tensor</code>) — | |
| Input batch of images as single-frame videos, shape <code>(batch_size, in_channels, 1, height, width)</code>.`,name:"sample"},{anchor:"diffusers.AutoencoderKLPH3Image.forward.sample_posterior",description:`<strong>sample_posterior</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) — | |
| Whether to sample the posterior instead of taking its mode.`,name:"sample_posterior"},{anchor:"diffusers.AutoencoderKLPH3Image.forward.generator",description:`<strong>generator</strong> (<code>torch.Generator</code>, <em>optional</em>) — | |
| Generator used when <code>sample_posterior=True</code>.`,name:"generator"},{anchor:"diffusers.AutoencoderKLPH3Image.forward.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) — | |
| Whether to return a <a href="/docs/diffusers/pr_15212/en/api/models/autoencoderkl_qwenimage#diffusers.models.autoencoders.vae.DecoderOutput">DecoderOutput</a> instead of a plain tuple.`,name:"return_dict"}],returnDescription:` | |
| <p>The round-tripped images, shape <code>(batch_size, out_channels, 1, height, width)</code>.</p> | |
| `,returnType:` | |
| <p><a | |
| href="/docs/diffusers/pr_15212/en/api/models/autoencoderkl_qwenimage#diffusers.models.autoencoders.vae.DecoderOutput" | |
| >DecoderOutput</a> or <code>tuple</code></p> | |
| `}),t(2),n(_),n(r);var H=e(r,2);J(H,{source:"https://github.com/huggingface/diffusers/blob/main/docs/source/en/api/models/autoencoderkl_p_h3_image.md"}),t(2),b(I,l),q()}export{C as component}; | |
Xet Storage Details
- Size:
- 12.5 kB
- Xet hash:
- 35e87221112a5e045ae41acbe29e621288ddf7d32b9b21dac982199a99b9061e
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.