File size: 7,501 Bytes
4450790
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#Original code can be found on: https://github.com/black-forest-labs/flux

from dataclasses import dataclass

import torch
from torch import Tensor, nn

from .layers import (
    DoubleStreamBlock,
    EmbedND,
    LastLayer,
    MLPEmbedder,
    SingleStreamBlock,
    timestep_embedding,
)

from einops import rearrange, repeat
import comfy.ldm.common_dit

@dataclass
class FluxParams:
    in_channels: int
    out_channels: int
    vec_in_dim: int
    context_in_dim: int
    hidden_size: int
    mlp_ratio: float
    num_heads: int
    depth: int
    depth_single_blocks: int
    axes_dim: list
    theta: int
    patch_size: int
    qkv_bias: bool
    guidance_embed: bool


class Flux(nn.Module):
    """
    Transformer model for flow matching on sequences.
    """

    def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs):
        super().__init__()
        self.dtype = dtype
        params = FluxParams(**kwargs)
        self.params = params
        self.patch_size = params.patch_size
        self.in_channels = params.in_channels * params.patch_size * params.patch_size
        self.out_channels = params.out_channels * params.patch_size * params.patch_size
        if params.hidden_size % params.num_heads != 0:
            raise ValueError(
                f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
            )
        pe_dim = params.hidden_size // params.num_heads
        if sum(params.axes_dim) != pe_dim:
            raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}")
        self.hidden_size = params.hidden_size
        self.num_heads = params.num_heads
        self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
        self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations)
        self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations)
        self.guidance_in = (
            MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity()
        )
        self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device)

        self.double_blocks = nn.ModuleList(
            [
                DoubleStreamBlock(
                    self.hidden_size,
                    self.num_heads,
                    mlp_ratio=params.mlp_ratio,
                    qkv_bias=params.qkv_bias,
                    dtype=dtype, device=device, operations=operations
                )
                for _ in range(params.depth)
            ]
        )

        self.single_blocks = nn.ModuleList(
            [
                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations)
                for _ in range(params.depth_single_blocks)
            ]
        )

        if final_layer:
            self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations)

    def forward_orig(
        self,
        img: Tensor,
        img_ids: Tensor,
        txt: Tensor,
        txt_ids: Tensor,
        timesteps: Tensor,
        y: Tensor,
        guidance: Tensor = None,
        control=None,
        transformer_options={},
    ) -> Tensor:
        patches_replace = transformer_options.get("patches_replace", {})
        if img.ndim != 3 or txt.ndim != 3:
            raise ValueError("Input img and txt tensors must have 3 dimensions.")

        # running on sequences img
        img = self.img_in(img)
        vec = self.time_in(timestep_embedding(timesteps, 256).to(img.dtype))
        if self.params.guidance_embed:
            if guidance is None:
                raise ValueError("Didn't get guidance strength for guidance distilled model.")
            vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype))

        vec = vec + self.vector_in(y[:,:self.params.vec_in_dim])
        txt = self.txt_in(txt)

        ids = torch.cat((txt_ids, img_ids), dim=1)
        pe = self.pe_embedder(ids)

        blocks_replace = patches_replace.get("dit", {})
        for i, block in enumerate(self.double_blocks):
            if ("double_block", i) in blocks_replace:
                def block_wrap(args):
                    out = {}
                    out["img"], out["txt"] = block(img=args["img"], txt=args["txt"], vec=args["vec"], pe=args["pe"])
                    return out

                out = blocks_replace[("double_block", i)]({"img": img, "txt": txt, "vec": vec, "pe": pe}, {"original_block": block_wrap})
                txt = out["txt"]
                img = out["img"]
            else:
                img, txt = block(img=img, txt=txt, vec=vec, pe=pe)

            if control is not None: # Controlnet
                control_i = control.get("input")
                if i < len(control_i):
                    add = control_i[i]
                    if add is not None:
                        img += add

        img = torch.cat((txt, img), 1)

        for i, block in enumerate(self.single_blocks):
            if ("single_block", i) in blocks_replace:
                def block_wrap(args):
                    out = {}
                    out["img"] = block(args["img"], vec=args["vec"], pe=args["pe"])
                    return out

                out = blocks_replace[("single_block", i)]({"img": img, "vec": vec, "pe": pe}, {"original_block": block_wrap})
                img = out["img"]
            else:
                img = block(img, vec=vec, pe=pe)

            if control is not None: # Controlnet
                control_o = control.get("output")
                if i < len(control_o):
                    add = control_o[i]
                    if add is not None:
                        img[:, txt.shape[1] :, ...] += add

        img = img[:, txt.shape[1] :, ...]

        img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
        return img

    def forward(self, x, timestep, context, y, guidance, control=None, transformer_options={}, **kwargs):
        bs, c, h, w = x.shape
        patch_size = self.patch_size
        x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))

        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)

        h_len = ((h + (patch_size // 2)) // patch_size)
        w_len = ((w + (patch_size // 2)) // patch_size)
        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
        img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)

        txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
        out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control, transformer_options)
        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h,:w]