File size: 9,960 Bytes
c85ad6e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
# Copyright 2024 NVIDIA 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

import numpy as np
from typing import Callable, Dict, List, Optional, Tuple, Union

import torch
from PIL import Image

from diffusers.utils.torch_utils import randn_tensor

try:
    from .pipeline_pixeldit import PixelDiTPipeline
    from .pipeline_output import PixelDiTPipelineOutput
except ImportError:
    from pipeline_pixeldit import PixelDiTPipeline
    from pipeline_output import PixelDiTPipelineOutput


def _to_pixel_tensor(image, width, height, device, dtype):
    """Convert a PIL Image or float tensor to [B, 3, H, W] in [-1, 1]."""
    if isinstance(image, Image.Image):
        image = image.convert("RGB").resize((width, height))
        image = np.array(image, dtype=np.float32)
        image = torch.from_numpy(image).permute(2, 0, 1).div(127.5).sub(1.0)
        image = image.unsqueeze(0)
    elif isinstance(image, np.ndarray):
        if image.dtype == np.uint8:
            image = torch.from_numpy(image.astype(np.float32)).div(127.5).sub(1.0)
        else:
            image = torch.from_numpy(image.astype(np.float32)).mul(2.0).sub(1.0)
        if image.dim() == 3:
            image = image.permute(2, 0, 1).unsqueeze(0)
    elif isinstance(image, torch.Tensor):
        if image.dim() == 3:
            image = image.unsqueeze(0)
        if image.is_floating_point() and image.max() <= 1.0 + 1e-4:
            image = image.mul(2.0).sub(1.0)
    return image.to(device=device, dtype=dtype)


class PixelDiTImg2ImgPipeline(PixelDiTPipeline):
    """
    Img2img pipeline for PixelDiT.

    Inherits everything from :class:`PixelDiTPipeline` — same model, same text encoder,
    same LoRA API, same schedulers.

    Pass an input image and a ``strength`` value to control how much the image is modified:
    ``strength=1.0`` equals pure text-to-image generation; ``strength=0.1`` barely changes
    the input. Because PixelDiT is a pixel-space model (no VAE), noise is injected directly
    on the pixel tensor using the flow-matching formula:
    ``x_t = (1 − σ) · image + σ · noise``

    Note: PixelDiT needs ≥ 45 total denoising steps for clean output. With low ``strength``
    the effective step count drops — keep ``num_inference_steps`` at 50+ to compensate.

    Example::

        from diffusers.pipelines.pixeldit import PixelDiTImg2ImgPipeline
        from PIL import Image
        import torch

        pipe = PixelDiTImg2ImgPipeline.from_pretrained(
            "madtune/pixeldit-diffusers", torch_dtype=torch.bfloat16
        )
        pipe.to("cuda")

        init = Image.open("photo.jpg").convert("RGB")
        out  = pipe(
            prompt="a cinematic landscape, golden hour",
            image=init,
            strength=0.75,
            num_inference_steps=50,
        ).images[0]
        out.save("img2img_out.png")
    """

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
        """
        Load from the same HF repo as :class:`PixelDiTPipeline`.
        Internally loads a T2I pipeline, then transfers its components into this class.
        """
        import diffusers
        from .modeling_pixeldit_hf import PixelDiTModel
        if not hasattr(diffusers, "PixelDiTModel"):
            diffusers.PixelDiTModel = PixelDiTModel
        t2i = PixelDiTPipeline.from_pretrained(pretrained_model_name_or_path, **kwargs)
        return cls(
            transformer=t2i.transformer,
            scheduler=t2i.scheduler,
            text_encoder=t2i.text_encoder,
            tokenizer=t2i.tokenizer,
        )

    @torch.no_grad()
    def __call__(
        self,
        prompt: Union[str, List[str]],
        image: Union[Image.Image, torch.Tensor, np.ndarray],
        strength: float = 0.8,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        height: int = 512,
        width: int = 512,
        num_inference_steps: int = 20,
        guidance_scale: float = 3.5,
        flow_shift: Optional[float] = None,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        output_type: str = "pil",
        return_dict: bool = True,
        cross_attention_kwargs: Optional[Dict[str, any]] = None,
        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        **kwargs,
    ) -> Union[PixelDiTPipelineOutput, Tuple]:
        """
        Args:
            prompt: Text prompt(s) guiding image generation.
            image: Input image. Accepts PIL ``Image``, ``numpy.ndarray`` (H×W×3 uint8 or float),
                or ``torch.Tensor`` (3×H×W or B×3×H×W).
            strength: How much to transform the input (0 < strength ≤ 1). ``1.0`` = full noise
                (equivalent to t2i). Recommended range: 0.5–0.85.
            negative_prompt: Optional negative prompt(s).
            height: Output height in pixels (must be divisible by 16).
            width: Output width in pixels (must be divisible by 16).
            num_inference_steps: Total scheduler steps. Use ≥ 50 for best quality.
            guidance_scale: CFG scale. ~3.5–7.5 works well.
            flow_shift: Override the scheduler's flow shift at runtime (e.g. 3.0 for 512px,
                4.0 for 1024px). Leaves the scheduler config unchanged if ``None``.
            generator: Torch RNG for reproducibility.
            output_type: ``"pil"`` (default) or ``"np"`` (uint8 numpy array).
            return_dict: If ``True`` returns :class:`PixelDiTPipelineOutput`, else a tuple.
            cross_attention_kwargs: Passed to the attention processor (e.g. ``{"scale": 0.8}``
                to adjust LoRA strength at inference).
            callback_on_step_end: Optional callable invoked at the end of each denoising step.
            callback_on_step_end_tensor_inputs: Names of tensors forwarded to the callback.

        Returns:
            :class:`PixelDiTPipelineOutput` or ``tuple``.
        """
        device = self._execution_device
        dtype  = self.transformer.dtype
        self._guidance_scale = guidance_scale
        lora_scale = (cross_attention_kwargs or {}).get("scale", None)

        if isinstance(prompt, str):
            prompt = [prompt]
        batch_size = len(prompt)

        self.check_inputs(prompt, height, width, negative_prompt)

        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
            prompt,
            device=device,
            dtype=dtype,
            do_classifier_free_guidance=self.do_classifier_free_guidance,
            negative_prompt=negative_prompt,
            lora_scale=lora_scale,
        )

        # Override flow shift if requested (reverts after this call via set_timesteps)
        if flow_shift is not None:
            self.scheduler.config.shift = flow_shift
        self.scheduler.set_timesteps(num_inference_steps, device=device)
        timesteps = self.scheduler.timesteps
        self._num_timesteps = len(timesteps)

        # Skip to the start timestep determined by strength
        t_start   = max(0, int(num_inference_steps * (1.0 - strength)))
        timesteps = timesteps[t_start:]

        if len(timesteps) == 0:
            raise ValueError(
                f"strength={strength} with num_inference_steps={num_inference_steps} "
                "produces 0 denoising steps. Increase strength or num_inference_steps."
            )

        # Preprocess image and add flow-matching noise at sigma_start
        img_tensor = _to_pixel_tensor(image, width, height, device, dtype)
        if img_tensor.shape[0] == 1 and batch_size > 1:
            img_tensor = img_tensor.expand(batch_size, -1, -1, -1).contiguous()

        sigma_start = timesteps[0].float() / 1000.0
        noise   = randn_tensor(img_tensor.shape, generator=generator, device=device, dtype=dtype)
        latents = (1.0 - sigma_start) * img_tensor + sigma_start * noise

        # Denoising loop
        for i, t in enumerate(self.progress_bar(timesteps)):
            if self.do_classifier_free_guidance:
                latent_model_input = torch.cat([latents] * 2)
                embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
            else:
                latent_model_input = latents
                embeds = prompt_embeds

            t_input    = t.expand(latent_model_input.shape[0])
            noise_pred = self.transformer(latent_model_input, t_input, embeds)

            if self.do_classifier_free_guidance:
                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)

            if hasattr(self.scheduler, "scale_model_input"):
                latents = self.scheduler.step(
                    noise_pred, t,
                    self.scheduler.scale_model_input(latents, t),
                    return_dict=False,
                )[0]
            else:
                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]

            if callback_on_step_end is not None:
                cb_kwargs = {k: locals()[k] for k in callback_on_step_end_tensor_inputs}
                callback_on_step_end(i, t, cb_kwargs)

        # Decode (pixel-space — just clamp and normalise)
        image_out = (latents.clamp(-1, 1) + 1) / 2
        image_out = (image_out * 255).byte().permute(0, 2, 3, 1).cpu().numpy()

        if output_type == "pil":
            image_out = [Image.fromarray(img) for img in image_out]

        self.maybe_free_model_hooks()

        if not return_dict:
            return (image_out,)
        return PixelDiTPipelineOutput(images=image_out)