Image-to-Image
PEFT
Safetensors
File size: 16,807 Bytes
43bff6c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
#!/usr/bin/env python3
"""
Layer decomposition inference — Qwen-Image-Layered + GRPO LoRA.

Decomposes a source image into ordered RGBA layers (background + objects).

============================================================================
RECOMMENDED INFERENCE SETTINGS  ***  USE THESE  ***
============================================================================
    Sampler : Heun (2nd order)   <- the only sampler here; always used
    Steps   : 50                 <- --steps 50   (default)
    CFG     : 1.0 (disabled)     <- --guidance-scale 1.0 (default)
    Size    : 640 px             <- --size 640   (default)
    Layers  : 4                  <- --num-layers 4 (default)

All published results were produced with exactly these settings, and they are
the defaults below. Lowering the step count or raising CFG noticeably degrades
the decomposition. Note 50 Heun steps == ~100 model evaluations (Heun is
2nd order: two forward passes per step) — that is expected, not a bug.
============================================================================

Usage:
    python decompose.py --input photo.png --output out/
    python decompose.py --input images/ --output out/ --transparent

Output per image:
    out/<name>/source.png      resized input
    out/<name>/composite.png   layers recomposited (sanity check vs source)
    out/<name>/layer_0.png     background (inpainted behind the objects)
    out/<name>/layer_1..N.png  object layers, back-to-front
"""

import argparse
import os
from pathlib import Path

import numpy as np
import torch
from PIL import Image

IMAGE_EXTS = {".png", ".jpg", ".jpeg", ".webp", ".bmp", ".tiff"}

# Qwen-Image prompt template (must match training/inference used for the LoRA)
PROMPT_TEMPLATE = (
    "<|im_start|>system\nDescribe the image by detailing the color, shape, size, "
    "texture, quantity, text, spatial relationships of the objects and background:"
    "<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
)
PROMPT_TEMPLATE_DROP_IDX = 34
DEFAULT_PROMPT = "a clean, well composed image"
VAE_SCALE_FACTOR = 8


# ---------------------------------------------------------------------------
# Latent helpers
# ---------------------------------------------------------------------------

def rgb_to_rgba(image):
    b, _, h, w = image.shape
    alpha = torch.ones(b, 1, h, w, device=image.device, dtype=image.dtype)
    return torch.cat([image, alpha], dim=1)


def normalize_latents(latents, vae):
    mean = torch.tensor(vae.config.latents_mean).view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
    std = torch.tensor(vae.config.latents_std).view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
    return (latents - mean) / std


def denormalize_latents(latents, vae):
    mean = torch.tensor(vae.config.latents_mean).view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
    inv_std = (1.0 / torch.tensor(vae.config.latents_std)).view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
    return latents / inv_std + mean


def pack_latents(latents, batch_size, num_channels, height, width, num_frames):
    latents = latents.view(batch_size, num_frames, num_channels, height // 2, 2, width // 2, 2)
    latents = latents.permute(0, 1, 3, 5, 2, 4, 6)
    return latents.reshape(batch_size, num_frames * (height // 2) * (width // 2), num_channels * 4)


def unpack_latents(latents, height, width, num_layers, vae_scale_factor=VAE_SCALE_FACTOR):
    batch_size, _, channels = latents.shape
    frames = num_layers + 1
    h = 2 * (int(height) // (vae_scale_factor * 2))
    w = 2 * (int(width) // (vae_scale_factor * 2))
    latents = latents.view(batch_size, frames, h // 2, w // 2, channels // 4, 2, 2)
    latents = latents.permute(0, 1, 4, 2, 5, 3, 6)
    latents = latents.reshape(batch_size, frames, channels // 4, h, w)
    return latents.permute(0, 2, 1, 3, 4)


@torch.no_grad()
def encode_condition_image(vae, image):
    """RGB image tensor in [-1,1], (B,3,H,W) -> packed condition latents."""
    b = image.shape[0]
    vae_input_ch = getattr(vae.config, "input_channels", 3)
    if vae_input_ch == 4 and image.shape[1] == 3:
        image = rgb_to_rgba(image)
    elif vae_input_ch == 3 and image.shape[1] == 4:
        image = image[:, :3]

    image_5d = image.unsqueeze(2).to(dtype=vae.dtype)
    dist = vae.encode(image_5d)
    if hasattr(dist, "latent_dist"):
        latents = dist.latent_dist.mode()
    elif hasattr(dist, "mode"):
        latents = dist.mode()
    else:
        latents = dist

    latents = normalize_latents(latents, vae)
    z_dim, lh, lw = latents.shape[1], latents.shape[3], latents.shape[4]
    latents = latents.permute(0, 2, 1, 3, 4)
    return pack_latents(latents, b, z_dim, lh, lw, 1).to(dtype=torch.bfloat16)


@torch.no_grad()
def decode_layers(vae, latents, height, width, num_layers):
    b = latents.shape[0]
    unpacked = unpack_latents(latents, height, width, num_layers)
    unpacked = denormalize_latents(unpacked, vae)
    layer_latents = unpacked[:, :, 1:].permute(0, 2, 1, 3, 4)  # drop base frame
    _, _, c, h, w = layer_latents.shape
    decoded = vae.decode(layer_latents.reshape(b * num_layers, c, 1, h, w).to(dtype=vae.dtype),
                         return_dict=False)[0]
    decoded = decoded.squeeze(2).float().clamp(-1.0, 1.0)
    _, c_out, h_out, w_out = decoded.shape
    return decoded.reshape(b, num_layers, c_out, h_out, w_out)


def composite_layers(layers):
    """Porter-Duff 'over', back-to-front."""
    if layers.shape[2] == 3:
        return layers.mean(dim=1)
    result = torch.zeros_like(layers[:, 0, :3])
    for i in range(layers.shape[1]):
        rgb = layers[:, i, :3]
        alpha = (layers[:, i, 3:4] + 1.0) / 2.0
        result = rgb * alpha + result * (1.0 - alpha)
    return result


def tensor_to_pil(t, mode="RGB"):
    arr = ((t.clamp(-1, 1) + 1) / 2 * 255).byte().permute(1, 2, 0).cpu().numpy()
    return Image.fromarray(arr, mode=mode)


def layer_to_pil_white_bg(layer):
    if layer.shape[0] == 4:
        rgb, alpha = layer[:3], (layer[3:4] + 1.0) / 2.0
        return tensor_to_pil(rgb * alpha + torch.ones_like(rgb) * (1.0 - alpha))
    return tensor_to_pil(layer)


def layer_to_pil_rgba(layer):
    layer = layer.clamp(-1, 1).float()
    rgba = torch.cat([(layer[:3] + 1) / 2, (layer[3:4] + 1) / 2], dim=0)
    arr = (rgba * 255).byte().permute(1, 2, 0).cpu().numpy()
    return Image.fromarray(arr, mode="RGBA")


# ---------------------------------------------------------------------------
# Denoising — Heun (2nd order)
# ---------------------------------------------------------------------------

def transformer_forward(transformer, hidden_states, timestep, encoder_hidden_states,
                        img_shapes, encoder_hidden_states_mask=None, additional_t_cond=None):
    return transformer(
        hidden_states=hidden_states,
        timestep=timestep / 1000,
        encoder_hidden_states=encoder_hidden_states,
        encoder_hidden_states_mask=encoder_hidden_states_mask,
        img_shapes=img_shapes,
        guidance=None,
        additional_t_cond=additional_t_cond,
        return_dict=False,
    )[0]


def build_img_shapes(num_layers, packed_h, packed_w):
    # (num_layers + 1) generated frames + 1 condition-image entry
    return [[*[(1, packed_h, packed_w) for _ in range(num_layers + 1)], (1, packed_h, packed_w)]]


@torch.no_grad()
def denoise(transformer, scheduler, latents, timesteps, prompt_embeds, prompt_mask,
            img_shapes, condition_latents, guidance_scale,
            neg_embeds=None, neg_mask=None, additional_t_cond=None):
    """Heun's method (2nd order) over the scheduler's sigmas."""
    gen_seq_len = latents.shape[1]
    do_cfg = guidance_scale > 1.0 and neg_embeds is not None

    def velocity(lat, t_val):
        ts = t_val.expand(1).to(lat.dtype)
        model_input = lat if condition_latents is None else torch.cat([lat, condition_latents], dim=1)
        v = transformer_forward(transformer, model_input, ts, prompt_embeds,
                                img_shapes, prompt_mask, additional_t_cond)[:, :gen_seq_len]
        if do_cfg:
            v_neg = transformer_forward(transformer, model_input, ts, neg_embeds,
                                        img_shapes, neg_mask, additional_t_cond)[:, :gen_seq_len]
            v = v_neg + guidance_scale * (v - v_neg)
        return v

    sigmas = scheduler.sigmas
    for i, t in enumerate(timesteps):
        sigma = sigmas[i]
        sigma_next = sigmas[i + 1] if i + 1 < len(sigmas) else torch.tensor(0.0, device=t.device)
        dt = sigma_next - sigma

        v1 = velocity(latents, t)
        latents_mid = latents + dt * v1
        if sigma_next > 0:                       # Heun corrector
            v2 = velocity(latents_mid, sigma_next * 1000)
            latents = latents + dt * 0.5 * (v1 + v2)
        else:
            latents = latents_mid
    return latents


# ---------------------------------------------------------------------------
# Pipeline
# ---------------------------------------------------------------------------

def compute_aspect_resize(orig_w, orig_h, max_size):
    """Keep aspect ratio, max dim = max_size, both dims divisible by 16."""
    scale = max_size / max(orig_w, orig_h)
    return (max(int(round(orig_w * scale / 16)) * 16, 16),
            max(int(round(orig_h * scale / 16)) * 16, 16))


def collect_images(input_path):
    p = Path(input_path)
    if p.is_file():
        return [p]
    if p.is_dir():
        return sorted(f for f in p.iterdir() if f.suffix.lower() in IMAGE_EXTS)
    raise FileNotFoundError(f"input not found: {input_path}")


def main():
    ap = argparse.ArgumentParser(
        description="Layer decomposition (Heun / 50 steps / cfg 1.0 recommended).",
        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
    )
    ap.add_argument("--input", required=True, help="image file or directory of images")
    ap.add_argument("--output", required=True, help="output directory")
    ap.add_argument("--lora", default=str(Path(__file__).parent / "checkpoint-600"),
                    help="path to the LoRA adapter directory")
    ap.add_argument("--base-model", default="Qwen/Qwen-Image-Layered",
                    help="base model (HuggingFace id or local path)")
    # --- recommended settings (defaults) ---
    ap.add_argument("--steps", type=int, default=50, help="Heun steps (RECOMMENDED: 50)")
    ap.add_argument("--guidance-scale", type=float, default=1.0, help="CFG (RECOMMENDED: 1.0 = off)")
    ap.add_argument("--num-layers", type=int, default=4)
    ap.add_argument("--size", type=int, default=640, help="max image dimension")
    # --- misc ---
    ap.add_argument("--prompt", default=DEFAULT_PROMPT)
    ap.add_argument("--seed", type=int, default=42)
    ap.add_argument("--transparent", action="store_true",
                    help="save layers as RGBA with real alpha (for compositing) "
                         "instead of composited onto white")
    ap.add_argument("--device", default="cuda")
    ap.add_argument("--cache-dir", default=None, help="HuggingFace cache directory")
    args = ap.parse_args()

    if args.steps != 50 or args.guidance_scale != 1.0:
        print(f"WARNING: using steps={args.steps}, cfg={args.guidance_scale}. "
              f"Recommended is steps=50, cfg=1.0 — results may be worse.\n")

    device = torch.device(args.device)
    images = collect_images(args.input)
    if not images:
        raise SystemExit(f"no images found in {args.input}")
    os.makedirs(args.output, exist_ok=True)
    print(f"{len(images)} image(s) | Heun, {args.steps} steps, cfg {args.guidance_scale}, "
          f"{args.size}px, {args.num_layers} layers")

    # --- load base model -----------------------------------------------
    print(f"Loading base model: {args.base_model}")
    from diffusers import DiffusionPipeline

    pipe = DiffusionPipeline.from_pretrained(
        args.base_model, torch_dtype=torch.bfloat16,
        trust_remote_code=True, cache_dir=args.cache_dir,
    )
    transformer = pipe.transformer.to(device).eval()
    vae = pipe.vae.to(device).eval()
    scheduler = pipe.scheduler
    tokenizer = getattr(pipe, "tokenizer", None)
    text_encoder = getattr(pipe, "text_encoder", None)
    del pipe
    transformer.requires_grad_(False)
    vae.requires_grad_(False)

    # --- apply LoRA ------------------------------------------------------
    print(f"Loading LoRA: {args.lora}")
    from peft import PeftModel
    transformer = PeftModel.from_pretrained(transformer, args.lora)
    transformer.eval()

    in_channels = getattr(transformer.config, "in_channels", 64)
    additional_t_cond = torch.zeros(1, dtype=torch.long, device=device)

    # --- prompt encoding -------------------------------------------------
    if text_encoder is None or tokenizer is None:
        raise SystemExit("base model has no text encoder/tokenizer")
    text_encoder = text_encoder.to(device).eval()
    text_encoder.requires_grad_(False)

    @torch.no_grad()
    def encode_prompt(text):
        tokens = tokenizer([PROMPT_TEMPLATE.format(text)], padding=True,
                           return_tensors="pt").to(text_encoder.device)
        out = text_encoder(input_ids=tokens.input_ids,
                           attention_mask=tokens.attention_mask,
                           output_hidden_states=True)
        hidden = out.hidden_states[-1]
        mask_bool = tokens.attention_mask.bool()
        lengths = mask_bool.sum(dim=1)
        chunks = torch.split(hidden[mask_bool], lengths.tolist(), dim=0)
        chunks = [c[PROMPT_TEMPLATE_DROP_IDX:] for c in chunks]   # drop system prefix
        max_len = max(c.size(0) for c in chunks)
        embeds = torch.stack([torch.cat([c, c.new_zeros(max_len - c.size(0), c.size(1))]) for c in chunks])
        mask = torch.stack([
            torch.cat([torch.ones(c.size(0), dtype=torch.long, device=c.device),
                       torch.zeros(max_len - c.size(0), dtype=torch.long, device=c.device)])
            for c in chunks
        ])
        return embeds.to(dtype=torch.bfloat16, device=device), mask.to(device)

    prompt_embeds, prompt_mask = encode_prompt(args.prompt)
    neg_embeds, neg_mask = encode_prompt("")

    # scheduler may or may not accept sigmas/mu
    import inspect
    supports_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters)

    # --- per-image loop --------------------------------------------------
    for idx, img_path in enumerate(images):
        name = img_path.stem
        out_dir = Path(args.output) / name
        out_dir.mkdir(parents=True, exist_ok=True)
        print(f"[{idx + 1}/{len(images)}] {img_path.name}", flush=True)

        image = Image.open(img_path).convert("RGB")
        tw, th = compute_aspect_resize(*image.size, args.size)
        source = image.resize((tw, th), Image.LANCZOS)

        packed_h = (th // VAE_SCALE_FACTOR) // 2
        packed_w = (tw // VAE_SCALE_FACTOR) // 2

        img_t = torch.from_numpy(np.array(source)).permute(2, 0, 1).float() / 127.5 - 1.0
        img_t = img_t.unsqueeze(0).to(device)
        condition_latents = encode_condition_image(vae, img_t)

        # dynamic-shift mu from the condition sequence length
        mu = (condition_latents.shape[1] / (256 * 256 / 16 / 16)) ** 0.5
        if supports_sigmas:
            scheduler.set_timesteps(sigmas=np.linspace(1.0, 0, args.steps + 1)[:-1],
                                    device=device, mu=mu)
        else:
            scheduler.set_timesteps(num_inference_steps=args.steps, device=device)

        img_shapes = build_img_shapes(args.num_layers, packed_h, packed_w)
        seq_len = (args.num_layers + 1) * packed_h * packed_w

        gen = torch.Generator(device=device).manual_seed(args.seed + idx)
        latents = torch.randn(1, seq_len, in_channels, device=device,
                              dtype=torch.bfloat16, generator=gen)

        latents = denoise(transformer, scheduler, latents, scheduler.timesteps,
                          prompt_embeds, prompt_mask, img_shapes, condition_latents,
                          args.guidance_scale, neg_embeds, neg_mask, additional_t_cond)

        decoded = decode_layers(vae, latents, th, tw, args.num_layers)
        composite = composite_layers(decoded)

        source.save(out_dir / "source.png")
        tensor_to_pil(composite[0]).save(out_dir / "composite.png")
        for li in range(args.num_layers):
            layer = decoded[0, li]
            pil = layer_to_pil_rgba(layer) if args.transparent else layer_to_pil_white_bg(layer)
            pil.save(out_dir / f"layer_{li}.png")

    print(f"\nDone -> {args.output}/<name>/{{source,composite,layer_0..{args.num_layers - 1}}}.png")


if __name__ == "__main__":
    main()