Spaces:
Paused
Paused
File size: 4,830 Bytes
cd458ae | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 | # Copyright 2026 Krea AI and The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import torch
from ...configuration_utils import FrozenDict
from ...image_processor import VaeImageProcessor
from ...models import AutoencoderKLQwenImage
from ...utils import logging
from ..modular_pipeline import ModularPipelineBlocks, PipelineState
from ..modular_pipeline_utils import ComponentSpec, InputParam, OutputParam
from .modular_pipeline import Krea2ModularPipeline
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
# auto_docstring
class Krea2DecodeStep(ModularPipelineBlocks):
"""
Step that unpacks the denoised packed latents back to the spatial grid, de-normalizes them with the VAE's
per-channel statistics, and decodes them through the Qwen-Image VAE into images.
Components:
vae (`AutoencoderKLQwenImage`) image_processor (`VaeImageProcessor`)
Inputs:
output_type (`str`, *optional*, defaults to pil):
Output format: 'pil', 'np', 'pt'.
height (`int`, *optional*, defaults to 1024):
The height in pixels of the generated image.
width (`int`, *optional*, defaults to 1024):
The width in pixels of the generated image.
latents (`Tensor`):
The denoised packed latents (B, image_seq_len, in_channels) from the denoising loop.
Outputs:
images (`list`):
Generated images.
"""
model_name = "krea2"
@property
def description(self) -> str:
return (
"Step that unpacks the denoised packed latents back to the spatial grid, de-normalizes them with the "
"VAE's per-channel statistics, and decodes them through the Qwen-Image VAE into images."
)
@property
def expected_components(self) -> list[ComponentSpec]:
return [
ComponentSpec("vae", AutoencoderKLQwenImage),
ComponentSpec(
"image_processor",
VaeImageProcessor,
# Effective pixel-to-token downsampling factor: vae_scale_factor (8) * patch_size (2).
config=FrozenDict({"vae_scale_factor": 16}),
default_creation_method="from_config",
),
]
@property
def inputs(self) -> list[InputParam]:
return [
InputParam.template("output_type", default="pil"),
InputParam.template("height", default=1024),
InputParam.template("width", default=1024),
InputParam(
name="latents",
required=True,
type_hint=torch.Tensor,
description="The denoised packed latents (B, image_seq_len, in_channels) from the denoising loop.",
),
]
@property
def intermediate_outputs(self) -> list[OutputParam]:
return [OutputParam.template("images")]
@torch.no_grad()
def __call__(self, components: Krea2ModularPipeline, state: PipelineState) -> PipelineState:
block_state = self.get_block_state(state)
vae = components.vae
p = components.patch_size
latents = block_state.latents
batch_size, _, channels = latents.shape
height = p * (int(block_state.height) // (components.vae_scale_factor * p))
width = p * (int(block_state.width) // (components.vae_scale_factor * p))
latents = latents.view(batch_size, height // p, width // p, channels // (p * p), p, p)
latents = latents.permute(0, 3, 1, 4, 2, 5)
latents = latents.reshape(batch_size, channels // (p * p), 1, height, width)
latents = latents.to(vae.dtype)
latents_mean = (
torch.tensor(vae.config.latents_mean).view(1, vae.config.z_dim, 1, 1, 1).to(latents.device, latents.dtype)
)
latents_std = 1.0 / torch.tensor(vae.config.latents_std).view(1, vae.config.z_dim, 1, 1, 1).to(
latents.device, latents.dtype
)
latents = latents / latents_std + latents_mean
image = vae.decode(latents, return_dict=False)[0][:, :, 0]
block_state.images = components.image_processor.postprocess(image, output_type=block_state.output_type)
self.set_block_state(state, block_state)
return components, state
|