Text-to-Image
Diffusers
PyTorch
flow-matching
rectified-flow
dit
dual-stream-dit
sprint
anime
aipracticecafe commited on
Commit
ad580cf
·
verified ·
1 Parent(s): a14a45e

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +215 -0
README.md ADDED
@@ -0,0 +1,215 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: openrail++
3
+ library_name: diffusers
4
+ tags:
5
+ - text-to-image
6
+ - flow-matching
7
+ - rectified-flow
8
+ - dit
9
+ - dual-stream-dit
10
+ - sprint
11
+ - anime
12
+ - pytorch
13
+ widget:
14
+ - output:
15
+ url: images/sample_asuka_zoom.png
16
+ text: "1girl, souryuu asuka langley, neon genesis evangelion, rebuild of evangelion, evangelion: 3.0 you can \\(not\\) redo, momoko \\(momopoco\\), toosaka asagi, ebifurya, sensitive, masterpiece, good score, absurdres, highres, best quality, amazing quality, great quality, solo, long hair, blue eyes, orange hair, breasts, parted lips, official alternate costume, pilot suit, red bodysuit, black eye patch, animal hat, cabbie hat, cat hat, black hat, gloves, hat, hat horns, jacket, mecha pilot suit, bodysuit under clothes, red gloves, red jacket, track jacket, turtleneck, turtleneck jacket, interface headset, multicolored clothes, interface headset \\(evangelion\\), upper body, simple background"
17
+ - output:
18
+ url: images/sample_hoshino_ai.png
19
+ text: 1girl, hoshino ai, oshi no ko, ebifurya, sensitive, masterpiece, good score, absurdres, highres, best quality, amazing quality, great quality, dress, gloves, hair ornament, idol, idol clothes, long hair, looking at viewer, microphone, one eye closed, open mouth, pink dress, pink gloves, pointing, purple eyes, purple hair, solo, star-shaped pupils, star, star in eye, symbol-shaped pupils, symbol in eye, anime coloring, black shirt, brooch, frilled gloves, frilled shirt, frilled sleeves, frills, hair between eyes, heart, heart brooch, jewelry, messy hair, one side up, pink shirt, rabbit hair ornament, shiny clothes, shiny skin, shirt, short sleeves, solo, star \(symbol\), turtleneck, wing brooch, wings, yellow shirt, upper body, simple background
20
+ - output:
21
+ url: images/sample_asuka_bodysuit.png
22
+ text: "1girl, souryuu asuka langley, neon genesis evangelion, rebuild of evangelion, evangelion: 3.0 you can \\(not\\) redo, momoko \\(momopoco\\), toosaka asagi, ebifurya, sensitive, masterpiece, good score, absurdres, highres, best quality, amazing quality, great quality, solo, long hair, blue eyes, orange hair, breasts, official alternate costume, bodysuit, breasts, brown hair, closed mouth, clothing cutout, dark background, expressionless, eyepatch, falling petals, floating hair, hair ornament, holding, lance of longinus \\(evangelion\\), long hair, looking at viewer, looking down, mecha pilot suit, medium breasts, petals, plugsuit \\(evangelion\\), polearm, red bodysuit, solo, spear, standing, taut clothes, weapon, holding weapon, holding polearm, plugsuit, interface headset, light particle,upper body"
23
+ - output:
24
+ url: images/sample_yuri_claire.png
25
+ text: 2girls, claire francois, rae taylor, watashi no oshi wa akuyaku reijou, sensitive, masterpiece, good score, absurdres, highres, best quality, amazing quality, great quality, :d, ^ ^, blonde hair, blue eyes, blush, white ascot, bow, brown hair, closed eyes, collared shirt, cropped jacket, drill hair, drill sidelocks, floating hair, frills, gem, hair bow, half updo, holding another's arm, jacket, long hair, long sleeves, multiple girls, red bow, red jacket, school uniform, shirt, short hair, smile, yuri, brown eyes, multiple girls, white shirt, ascot, brooch, frilled jacket, jewelry, medium breasts, very long hair, upper body
26
+ datasets:
27
+ - aipracticecafe-mirror/curated-danbooru-2026-256px-flux2-vae
28
+ - aipracticecafe/curated-danbooru-2026
29
+ ---
30
+
31
+ # Asuka-FM: 200M Dual-Stream SPRINT DiT (Flow Matching)
32
+
33
+ An ultra-efficient, 200M-parameter **Dual-Stream Diffusion Transformer
34
+ (DiT)** trained from scratch on consumer hardware (1x NVIDIA RTX 4090 in ~50
35
+ hours) using **Rectified Flow / Flow Matching**, **SPRINT** token-dropping,
36
+ and **Continuous Aspect-Normalized RoPE**.
37
+
38
+ Despite its compact size, the model generates high-fidelity anime art with
39
+ accurate character reproduction, prompt alignment, and zero-shot camera
40
+ control (panning and zoom) at inference time.
41
+
42
+ - **Developed by**: [aipracticecafe](https://huggingface.co/aipracticecafe)
43
+ - **Repository**: [deeplearningcafe/asuka-fm](https://github.com/deeplearningcafe/asuka-fm)
44
+ - **Dataset**: [curated-danbooru-2026](https://huggingface.co/datasets/aipracticecafe/curated-danbooru-2026) (~337K samples)
45
+ - **Architecture**: Dual-Stream DiT (SPRINT-aligned, SwiGLU, Sandwich RMSNorm)
46
+ - **Parameters**: ~200M (Backbone) + 230M (Text Encoder)
47
+ - **Objective**: Conditional Flow Matching (Logit-Normal timestep sampling)
48
+ - **License**: OpenRAIL++
49
+
50
+ ---
51
+
52
+ ## Key Architecture & Features
53
+
54
+ ### 1. SPRINT Dual-Stream Transformer Backbone
55
+ The base architecture is based on the results from (i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models)[https://arxiv.org/abs/2606.11289]
56
+ - **Minimalist Conditioning**: Uses prepended time tokens rather than heavy
57
+ per-block AdaLN modulation, reducing parameter overhead.
58
+ - **Sandwich Normalization & SwiGLU**: RMSNorm placed before and after
59
+ attention and FFN operations ensures training stability in bfloat16.
60
+ - **SPRINT Acceleration**: Employs structured group-wise $2 \times 2$ patch
61
+ dropping ($75\%$ drop ratio during early pretraining), cutting compute
62
+ demands by half while retaining global semantic context.
63
+
64
+ ### 2. HDM Continuous Aspect-Normalized RoPE (Camera Control)
65
+ Based on the [Patch Diffusion: Faster and More Data-Efficient](https://arxiv.org/abs/2304.12526) implementation proposed in the [HDM](https://arxiv.org/abs/2509.06068) paper. Early experiments with discrete integer position IDs caused severe multi-head
66
+ replication when deviating from trained aspect ratios. We resolve this by
67
+ adopting continuous aspect-normalized coordinate mapping which was the original idea proposed in HDM:
68
+
69
+ $$r_H = \sqrt{\frac{H}{W}}, \quad r_W = \sqrt{\frac{W}{H}}$$
70
+
71
+ $$y_{\text{pos}} = \left(\frac{y_{\text{center}}}{H}\right) \cdot (2 r_H) - r_H, \quad
72
+ x_{\text{pos}} = \left(\frac{x_{\text{center}}}{W}\right) \cdot (2 r_W) - r_W$$
73
+
74
+ Manipulating $(x_{\text{pos}}, y_{\text{pos}})$ directly controls the visual
75
+ framing at inference time:
76
+ - **X-Shift / Y-Shift**: Translates the virtual camera viewport horizontally
77
+ or vertically.
78
+ - **Zoom**: Scales spatial coordinates around the focal center.
79
+
80
+ ### 3. Components
81
+ - **Text Encoder**: `LiquidAI/LFM2.5-230M` (unidirectional causal LM hidden
82
+ states passed via a 2-layer Transformer Text Adapter).
83
+ - **Autoencoder**: `kaiyuyue/FLUX.2-dev-vae` ($C_{\text{lat}}=32$, $f=8$,
84
+ patch size $p=2$, effective image token size $16 \times 16$).
85
+
86
+ ---
87
+
88
+ ## Training Recipe (1x RTX 4090, ~50h)
89
+
90
+ The model was pretrained on **337,000 curated Danbooru samples** at 256px base
91
+ resolution with shifted square cropping and cached in-RAM latents.
92
+
93
+ | Phase | Epochs | Token Drop | RoPE Format | Description |
94
+ |---|---|---|---|---|
95
+ | **Phase 1** | 100 ep | 0.75 (SPRINT) | Discrete | Fast semantic convergence |
96
+ | **Phase 2** | 20 ep | 0.75 (SPRINT) | Continuous | RoPE coordinates alignment |
97
+ | **Phase 3** | 25 ep | 0.00 (Dense) | Continuous | High-frequency detail recovery |
98
+ | **Phase 4** | 25 ep | 0.00 (Dense) | Continuous | Cosine LR decay to minimum LR |
99
+
100
+ - **Total Epochs**: 175
101
+ - **Optimizer**: 8-bit AdamW (`bitsandbytes`), $\beta=(0.9, 0.95)$, WD=0.01
102
+ - **Learning Rate**: $5 \times 10^{-4}$ with warmup and cosine annealing only for last stage
103
+ - **Effective Batch Size**: 256 (Batch size 64 $\times$ 4 gradient accumulation)
104
+ - **Precision**: Mixed Precision (`torch.bfloat16` with `torch.compile`)
105
+
106
+ ---
107
+
108
+ ## Inference and WebUI
109
+ The training repository has a simple Gradio UI to perform sampling with camera controls.
110
+ [ui_app.py](https://github.com/deeplearningcafe/asuka-fm/blob/main/scripts/ui_app.py)
111
+
112
+ ### Recommended Settings
113
+ - **Sampling Algorithm**: Forward Euler ODE Solver (Rectified Flow)
114
+ - **Inference Steps**: 25 – 35
115
+ - **CFG Scale**: 4.5 – 6.0
116
+ - **Time Shift**: 1.0 – 3.0 (higher values enhance contrast/saturation)
117
+
118
+ ### Tag Formatting Template
119
+ ```text
120
+ 1girl/1boy, character name, series name, artists, rating, quality tags, actions/clothing
121
+ ```
122
+
123
+ Negative prompt:
124
+ ```text
125
+ sketch, monochrome, english text, realistic, traditional media, very displeasing, displeasing, bad score, worse score, bad quality,worst quality, worst detail, watermark, signature, blurry, bad hands, bad anatomy, bad fingers, extra fingers, extra limbs, deformed limbs, comic, speech bubble, lowres
126
+ ```
127
+
128
+ ### Python Sampling Script
129
+
130
+ ```python
131
+ import torch
132
+ from omegaconf import OmegaConf
133
+ from src.models.factory import load_trainable_model
134
+ from src.diffusion.schedules import LinearSchedule
135
+ from src.diffusion.sampling import generate_samples
136
+
137
+ cfg = OmegaConf.load("config.yaml")
138
+ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
139
+ dtype = torch.bfloat16
140
+
141
+ unet, text_encoder, vae, tokenizer, _ = load_trainable_model(
142
+ models_path=cfg.paths.models,
143
+ device=device,
144
+ dtype=dtype,
145
+ resume_from_checkpoint="path_to_checkpoint",
146
+ model_type="sprint_dual",
147
+ model_cfg=cfg.models,
148
+ autocast_dtype=dtype,
149
+ )
150
+
151
+ schedule = LinearSchedule(device=device)
152
+
153
+ sample_configs = [
154
+ {
155
+ "prompt": "1girl, souryuu asuka langley, neon genesis evangelion, "
156
+ "masterpiece, absurdres",
157
+ "negative_prompt": "very displeasing, displeasing, bad score",
158
+ "height": 256,
159
+ "width": 384,
160
+ "sample_steps": 30,
161
+ "cfg_scale": 6.0,
162
+ "shift": 1.0,
163
+ "zoom": 1.0,
164
+ "x_shift": 0.0,
165
+ "y_shift": 0.0,
166
+ "seed": 42,
167
+ }
168
+ ]
169
+
170
+ images = generate_samples(
171
+ unet=unet,
172
+ text_encoder=text_encoder,
173
+ tokenizer=tokenizer,
174
+ vae=vae,
175
+ schedule=schedule,
176
+ sample_configs=sample_configs,
177
+ global_batch_size=1,
178
+ diffusion_type="flow_matching",
179
+ device=device,
180
+ dtype=dtype,
181
+ autocast_dtype=dtype,
182
+ use_unet_mult=False,
183
+ vae_mean=cfg.models.vae_mean,
184
+ vae_std=cfg.models.vae_std,
185
+ in_channels=cfg.models.in_channels,
186
+ coord_system="aspect_norm",
187
+ )
188
+
189
+ images[0].save("output.png")
190
+ ```
191
+
192
+ ---
193
+
194
+ ## Content Advisory & Limitations
195
+
196
+ - **Resolution**: Trained on 256px shifted crops. While continuous RoPE allows
197
+ sampling at 384px and 512px, high resolutions may exhibit slight softness
198
+ until a dedicated 512px/1024px fine-tuning pass is completed.
199
+ - **Complex Hands**: Small 200M parameter count prioritizes overall composition
200
+ and character concepts over fine micro-structures like fingernails.
201
+ - **NSFW & Sensitive Content:** The model was trained on a dataset that contains uncensored anime-style art, including sensitive, questionable, and explicit material (`g`, `s`, `q`, `e` ratings). Use appropriate tag filtering during training if a SFW-only model is required.
202
+ - **Legal Disclaimer:** Images are sourced from Danbooru for research, educational, and generative modeling purposes under fair use. Copyright of all underlying illustrations belongs to their respective creators.
203
+ ---
204
+
205
+ ## References
206
+ * **i1**: [i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models](https://arxiv.org/abs/2606.11289)
207
+ * **Patch Diffusion**: [Patch Diffusion: Faster and More Data-Efficient](https://arxiv.org/abs/2304.12526)
208
+ * **Home-made Diffusion Model from Scratch to Hatch**: [HDM](https://arxiv.org/abs/2509.06068)
209
+ * **Danbooru Dataset Toolkit**: Dataset processed and formatted using the [Danbooru Dataset Toolkit](https://github.com/deeplearningcafe/danbooru-dataset-toolkit).
210
+
211
+ ## License
212
+
213
+ This model is licensed under the **CreativeML Open RAIL++-M License**. You are
214
+ free to use, modify, and distribute the model weights provided you include the
215
+ license text and respect the behavioral use restrictions.