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
|
mod clip;
mod sampling;
mod vae;
use candle::{DType, IndexOp, Tensor};
use candle_transformers::models::mmdit::model::{Config as MMDiTConfig, MMDiT};
use crate::clip::StableDiffusion3TripleClipWithTokenizer;
use crate::vae::{build_sd3_vae_autoencoder, sd3_vae_vb_rename};
use anyhow::{Ok, Result};
use clap::Parser;
#[derive(Clone, Debug, Copy, PartialEq, Eq, clap::ValueEnum)]
enum Which {
#[value(name = "3-medium")]
V3Medium,
#[value(name = "3.5-large")]
V3_5Large,
#[value(name = "3.5-large-turbo")]
V3_5LargeTurbo,
}
impl Which {
fn is_3_5(&self) -> bool {
match self {
Self::V3Medium => false,
Self::V3_5Large | Self::V3_5LargeTurbo => true,
}
}
}
#[derive(Parser)]
#[command(author, version, about, long_about = None)]
struct Args {
/// The prompt to be used for image generation.
#[arg(
long,
default_value = "A cute rusty robot holding a candle torch in its hand, \
with glowing neon text \"LETS GO RUSTY\" displayed on its chest, \
bright background, high quality, 4k"
)]
prompt: String,
#[arg(long, default_value = "")]
uncond_prompt: String,
/// Run on CPU rather than on GPU.
#[arg(long)]
cpu: bool,
/// Enable tracing (generates a trace-timestamp.json file).
#[arg(long)]
tracing: bool,
/// Use flash_attn to accelerate attention operation in the MMDiT.
#[arg(long)]
use_flash_attn: bool,
/// The height in pixels of the generated image.
#[arg(long, default_value_t = 1024)]
height: usize,
/// The width in pixels of the generated image.
#[arg(long, default_value_t = 1024)]
width: usize,
/// The model to use.
#[arg(long, default_value = "3-medium")]
which: Which,
/// The seed to use when generating random samples.
#[arg(long)]
num_inference_steps: Option<usize>,
// CFG scale.
#[arg(long)]
cfg_scale: Option<f64>,
// Time shift factor (alpha).
#[arg(long, default_value_t = 3.0)]
time_shift: f64,
/// The seed to use when generating random samples.
#[arg(long)]
seed: Option<u64>,
}
fn main() -> Result<()> {
use tracing_chrome::ChromeLayerBuilder;
use tracing_subscriber::prelude::*;
let Args {
prompt,
uncond_prompt,
cpu,
tracing,
use_flash_attn,
height,
width,
num_inference_steps,
cfg_scale,
time_shift,
seed,
which,
} = Args::parse();
let _guard = if tracing {
let (chrome_layer, guard) = ChromeLayerBuilder::new().build();
tracing_subscriber::registry().with(chrome_layer).init();
Some(guard)
} else {
None
};
let device = candle_examples::device(cpu)?;
let default_inference_steps = match which {
Which::V3_5Large => 28,
Which::V3_5LargeTurbo => 4,
Which::V3Medium => 28,
};
let num_inference_steps = num_inference_steps.unwrap_or(default_inference_steps);
let default_cfg_scale = match which {
Which::V3_5Large => 4.0,
Which::V3_5LargeTurbo => 1.0,
Which::V3Medium => 4.0,
};
let cfg_scale = cfg_scale.unwrap_or(default_cfg_scale);
let api = hf_hub::api::sync::Api::new()?;
let (mmdit_config, mut triple, vb) = if which.is_3_5() {
let sai_repo = {
let name = match which {
Which::V3_5Large => "stabilityai/stable-diffusion-3.5-large",
Which::V3_5LargeTurbo => "stabilityai/stable-diffusion-3.5-large-turbo",
Which::V3Medium => unreachable!(),
};
api.repo(hf_hub::Repo::model(name.to_string()))
};
let clip_g_file = sai_repo.get("text_encoders/clip_g.safetensors")?;
let clip_l_file = sai_repo.get("text_encoders/clip_l.safetensors")?;
let t5xxl_file = sai_repo.get("text_encoders/t5xxl_fp16.safetensors")?;
let model_file = {
let model_file = match which {
Which::V3_5Large => "sd3.5_large.safetensors",
Which::V3_5LargeTurbo => "sd3.5_large_turbo.safetensors",
Which::V3Medium => unreachable!(),
};
sai_repo.get(model_file)?
};
let triple = StableDiffusion3TripleClipWithTokenizer::new_split(
&clip_g_file,
&clip_l_file,
&t5xxl_file,
&device,
)?;
let vb = unsafe {
candle_nn::VarBuilder::from_mmaped_safetensors(&[model_file], DType::F16, &device)?
};
(MMDiTConfig::sd3_5_large(), triple, vb)
} else {
let sai_repo = {
let name = "stabilityai/stable-diffusion-3-medium";
api.repo(hf_hub::Repo::model(name.to_string()))
};
let model_file = sai_repo.get("sd3_medium_incl_clips_t5xxlfp16.safetensors")?;
let vb_fp16 = unsafe {
candle_nn::VarBuilder::from_mmaped_safetensors(&[&model_file], DType::F16, &device)?
};
let vb_fp32 = unsafe {
candle_nn::VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)?
};
let triple = StableDiffusion3TripleClipWithTokenizer::new(
vb_fp16.pp("text_encoders"),
vb_fp32.pp("text_encoders"),
)?;
(MMDiTConfig::sd3_medium(), triple, vb_fp16)
};
let (context, y) = triple.encode_text_to_embedding(prompt.as_str(), &device)?;
let (context_uncond, y_uncond) =
triple.encode_text_to_embedding(uncond_prompt.as_str(), &device)?;
let context = Tensor::cat(&[context, context_uncond], 0)?;
let y = Tensor::cat(&[y, y_uncond], 0)?;
let mmdit = MMDiT::new(
&mmdit_config,
use_flash_attn,
vb.pp("model.diffusion_model"),
)?;
if let Some(seed) = seed {
device.set_seed(seed)?;
}
let start_time = std::time::Instant::now();
let x = sampling::euler_sample(
&mmdit,
&y,
&context,
num_inference_steps,
cfg_scale,
time_shift,
height,
width,
)?;
let dt = start_time.elapsed().as_secs_f32();
println!(
"Sampling done. {num_inference_steps} steps. {:.2}s. Average rate: {:.2} iter/s",
dt,
num_inference_steps as f32 / dt
);
let img = {
let vb_vae = vb.rename_f(sd3_vae_vb_rename).pp("first_stage_model");
let autoencoder = build_sd3_vae_autoencoder(vb_vae)?;
// Apply TAESD3 scale factor. Seems to be significantly improving the quality of the image.
// https://github.com/comfyanonymous/ComfyUI/blob/3c60ecd7a83da43d694e26a77ca6b93106891251/nodes.py#L721-L723
autoencoder.decode(&((x / 1.5305)? + 0.0609)?)?
};
let img = ((img.clamp(-1f32, 1f32)? + 1.0)? * 127.5)?.to_dtype(candle::DType::U8)?;
candle_examples::save_image(&img.i(0)?, "out.jpg")?;
Ok(())
}
|
