import json import os import torch from pathlib import Path from PIL import Image import math import numpy as np import torch import torch.nn.functional as F import torch.utils.checkpoint from torch import nn from PIL import Image SIZES = (16, 32, 64) def timestep_embedding(t, dim, max_period=10000): half = dim // 2 freqs = torch.exp(-math.log(max_period) * torch.arange(half, device=t.device, dtype=torch.float32) / half) args = t.float()[:, None] * freqs[None, :] return torch.cat([torch.cos(args), torch.sin(args)], dim=-1) class RMSNorm(nn.Module): def __init__(self, dim, eps=1e-6): super().__init__() self.weight = nn.Parameter(torch.ones(dim)) self.eps = eps def forward(self, x): dtype = x.dtype x = x.float() x = x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) return x.to(dtype) * self.weight def _rope_angles(pos, n_freq, theta=10000.0): inv = 1.0 / (theta ** (torch.arange(n_freq, device=pos.device, dtype=torch.float32) / n_freq)) return pos.float()[:, None] * inv[None, :] def rotate_half(x): a, b = x.chunk(2, dim=-1) return torch.cat([-b, a], dim=-1) _ROPE_CACHE = {} def build_rope(n, text_len, head_dim, device, p=1): # image token positions are in PIXEL coordinates (patch centres), so patch 1 and patch 2 share one coordinate system key = (n, text_len, head_dim, str(device), p) if key not in _ROPE_CACHE: n_freq = head_dim // 4 coords = torch.arange(n, device=device, dtype=torch.float32) * p + (p - 1) / 2.0 ys, xs = torch.meshgrid(coords, coords, indexing="ij") txt_y = -(torch.arange(text_len, device=device, dtype=torch.float32) + 1.0) pos_y = torch.cat([txt_y, ys.reshape(-1)]) pos_x = torch.cat([torch.zeros(text_len, device=device), xs.reshape(-1)]) ang = torch.cat([_rope_angles(pos_y, n_freq), _rope_angles(pos_x, n_freq)], dim=-1) _ROPE_CACHE[key] = (torch.cat([ang.cos(), ang.cos()], dim=-1), torch.cat([ang.sin(), ang.sin()], dim=-1)) return _ROPE_CACHE[key] class DiTBlock(nn.Module): def __init__(self, dim, n_heads, mlp_hidden): super().__init__() self.n_heads = n_heads self.head_dim = dim // n_heads self.norm1 = RMSNorm(dim) self.qkv = nn.Linear(dim, dim * 3, bias=False) self.proj_out = nn.Linear(dim, dim, bias=False) self.norm2 = RMSNorm(dim) self.fc1 = nn.Linear(dim, mlp_hidden, bias=False) self.fc3 = nn.Linear(dim, mlp_hidden, bias=False) self.fc2 = nn.Linear(mlp_hidden, dim, bias=False) self.table = nn.Parameter(torch.zeros(6, dim)) def forward(self, x, cos, sin, mod): s1, sc1, g1, s2, sc2, g2 = [m + self.table[i].to(m.dtype) for i, m in enumerate(mod)] h = self.norm1(x) * (1 + sc1) + s1 B, S, _ = h.shape qkv = self.qkv(h).view(B, S, 3, self.n_heads, self.head_dim).permute(2, 0, 3, 1, 4) q, k, v = qkv.unbind(0) qf, kf = q.float(), k.float() q = (qf * cos + rotate_half(qf) * sin).to(v.dtype) k = (kf * cos + rotate_half(kf) * sin).to(v.dtype) o = F.scaled_dot_product_attention(q, k, v) o = o.transpose(1, 2).reshape(B, S, -1) x = x + self.proj_out(o) * g1 h = self.norm2(x) * (1 + sc2) + s2 x = x + self.fc2(F.silu(self.fc1(h)) * self.fc3(h)) * g2 return x class PixelArtDiT(nn.Module): def __init__(self, dim=768, n_layers=18, n_heads=12, mlp_ratio=3.0, patch_small=1, patch_large=2, large_from=64, in_channels=4, text_len=64, enc_dim=2048): super().__init__() self.dim = dim self.n_heads = n_heads self.patch_small = patch_small self.patch_large = patch_large self.large_from = large_from self.in_channels = in_channels self.text_len = text_len self.enc_dim = enc_dim patch_set = sorted({patch_small, patch_large}) # one patch-in / patch-out layer per patch size; the transformer body is shared self.x_projs = nn.ModuleDict({str(p): nn.Linear(in_channels * p * p, dim, bias=False) for p in patch_set}) self.outs = nn.ModuleDict({str(p): nn.Linear(dim, in_channels * p * p) for p in patch_set}) self.res_emb = nn.Embedding(len(SIZES), dim) # which resolution regime (16 / 32 / 64) # frozen-encoder features -> model width self.txt_norm = RMSNorm(enc_dim) self.txt_proj = nn.Linear(enc_dim, dim, bias=False) self.pad_emb = nn.Parameter(torch.randn(dim) * 0.02) # replaces padded positions self.null_txt = nn.Parameter(torch.randn(text_len, dim) * 0.02) # learned "no caption" (CFG) self.time_mlp = nn.Sequential(nn.Linear(dim, dim), nn.SiLU(), nn.Linear(dim, dim)) self.mod = nn.Linear(dim, 6 * dim) self.blocks = nn.ModuleList( [DiTBlock(dim, n_heads, int(dim * mlp_ratio)) for _ in range(n_layers)] ) self.norm_out = RMSNorm(dim) self.grad_ckpt = False nn.init.normal_(self.res_emb.weight, std=0.02) nn.init.zeros_(self.mod.weight) nn.init.zeros_(self.mod.bias) for o in self.outs.values(): nn.init.zeros_(o.weight) nn.init.zeros_(o.bias) def patch_for(self, size): return self.patch_large if size >= self.large_from else self.patch_small @staticmethod def res_index(size): if size not in SIZES: raise ValueError(f"unsupported image size {size}; supported sizes: {SIZES}") return SIZES.index(size) def patchify(self, x, p): B, C, H, W = x.shape nh, nw = H // p, W // p x = x.reshape(B, C, nh, p, nw, p).permute(0, 2, 4, 1, 3, 5) return x.reshape(B, nh * nw, C * p * p), nh, nw def unpatchify(self, x, nh, nw, p): B = x.shape[0] x = x.reshape(B, nh, nw, self.in_channels, p, p).permute(0, 3, 1, 4, 2, 5) return x.reshape(B, self.in_channels, nh * p, nw * p) def forward(self, x, t, txt, txt_mask, drop=None): # txt: (B, text_len, enc_dim) encoder features | txt_mask: (B, text_len) bool # drop: optional (B,) bool -> replace that sample's caption with the learned null caption B, _, H, W = x.shape p = self.patch_for(H) proj = self.x_projs[str(p)] img, nh, nw = self.patchify(x, p) img = proj(img.to(proj.weight.dtype)) txt = self.txt_proj(self.txt_norm(txt.to(self.txt_proj.weight.dtype))) mask = txt_mask.bool() txt = torch.where(mask.unsqueeze(-1), txt, self.pad_emb) if drop is not None: d = drop.bool() txt = torch.where(d.view(B, 1, 1), self.null_txt.unsqueeze(0), txt) mask = torch.where(d.view(B, 1), torch.ones_like(mask), mask) seq = torch.cat([txt, img], dim=1) t_emb = timestep_embedding(t * 1000.0, self.dim).to(self.time_mlp[0].weight.dtype) keep = mask.unsqueeze(-1).to(txt.dtype) pooled = (txt * keep).sum(dim=1) / keep.sum(dim=1).clamp(min=1.0) cond = self.time_mlp(t_emb) + pooled + self.res_emb.weight[self.res_index(H)] mod = self.mod(F.silu(cond)).view(B, 6, 1, self.dim).unbind(1) cos, sin = build_rope(nh, self.text_len, self.dim // self.n_heads, x.device, p) h = seq for blk in self.blocks: if self.grad_ckpt and self.training: h = torch.utils.checkpoint.checkpoint(blk, h, cos, sin, mod, use_reentrant=False) else: h = blk(h, cos, sin, mod) h = self.norm_out(h)[:, self.text_len:] return self.unpatchify(self.outs[str(p)](h), nh, nw, p) def load_text_encoder(model_id, device="cuda", dtype=torch.bfloat16): from transformers import AutoModel, AutoTokenizer tokenizer = AutoTokenizer.from_pretrained(model_id) if tokenizer.pad_token is None: tokenizer.pad_token = tokenizer.eos_token last_err = None for kw in ({"dtype": dtype, "attn_implementation": "sdpa"}, {"torch_dtype": dtype, "attn_implementation": "sdpa"}, {"dtype": dtype}, {"torch_dtype": dtype}): try: enc = AutoModel.from_pretrained(model_id, trust_remote_code=True, **kw) break except (TypeError, ValueError, ImportError) as e: last_err = e else: raise last_err enc = enc.to(device).eval().requires_grad_(False) return tokenizer, enc @torch.no_grad() def encode_prompts(tokenizer, text_encoder, prompts, text_len, device): prompts = [p if isinstance(p, str) and p.strip() else "." for p in prompts] tok = tokenizer(prompts, padding="max_length", truncation=True, max_length=text_len, return_tensors="pt") ids = tok["input_ids"].to(device) mask = tok["attention_mask"].to(device) out = text_encoder(input_ids=ids, attention_mask=mask) feats = out.last_hidden_state if hasattr(out, "last_hidden_state") else out[0] return feats, mask.bool() def decode_latents(x): x = x.float().clamp(-1, 1).add(1).mul(127.5).round().to(torch.uint8) return [Image.fromarray(np.ascontiguousarray(arr.transpose(1, 2, 0))) for arr in x.cpu().numpy()] def sprite_grid(images, scale=8, cols=None, bg=(40, 40, 40, 255)): cols = cols or int(math.ceil(math.sqrt(len(images)))) rows = math.ceil(len(images) / cols) w, h = images[0].size grid = Image.new("RGBA", (cols * w, rows * h), bg) for i, im in enumerate(images): grid.alpha_composite(im.convert("RGBA"), dest=((i % cols) * w, (i // cols) * h)) return grid.resize((cols * w * scale, rows * h * scale), Image.NEAREST) @torch.no_grad() def generate(model, prompts, encode_fn, size=32, steps=30, guidance=3.0, seed=0, device="cuda", num_per_prompt=1): # encode_fn(list[str]) -> (features, bool mask) was_training = model.training model.eval() dev = torch.device(device) gen = torch.Generator(device=dev).manual_seed(int(seed)) if isinstance(prompts, str): prompts = [prompts] expanded = [p for p in prompts for _ in range(num_per_prompt)] B = len(expanded) feats, mask = encode_fn(expanded) use_cfg = guidance != 1.0 if use_cfg: feats2 = torch.cat([feats, feats]) mask2 = torch.cat([mask, mask]) drop2 = torch.cat([torch.zeros(B, dtype=torch.bool, device=dev), torch.ones(B, dtype=torch.bool, device=dev)]) x = torch.randn(B, model.in_channels, size, size, device=dev, generator=gen) ts = torch.linspace(1.0, 0.0, steps + 1, device=dev) for i in range(steps): if use_cfg: xi, f, m, d, t = torch.cat([x, x]), feats2, mask2, drop2, ts[i].expand(2 * B) else: xi, f, m, d, t = x, feats, mask, None, ts[i].expand(B) with torch.autocast(device_type=dev.type, dtype=torch.bfloat16, enabled=dev.type == "cuda"): v = model(xi, t, f, m, d) v = v.float() if use_cfg: vc, vu = v.chunk(2) v = vu + guidance * (vc - vu) x = x + (ts[i + 1] - ts[i]) * v model.train(was_training) return decode_latents(x) import json import os from pathlib import Path try: from diffusers import DiffusionPipeline from diffusers.configuration_utils import ConfigMixin, register_to_config from diffusers.models.modeling_utils import ModelMixin from diffusers.pipelines.pipeline_utils import ImagePipelineOutput from diffusers.utils.torch_utils import randn_tensor _HAS_DIFFUSERS = True except ImportError: _HAS_DIFFUSERS = False if _HAS_DIFFUSERS: class PixelArtDiTModel(ModelMixin, ConfigMixin): config_name = "config.json" @register_to_config def __init__(self, dim=768, n_layers=18, n_heads=12, mlp_ratio=3.0, patch_small=1, patch_large=2, large_from=64, in_channels=4, text_len=64, enc_dim=2048, text_encoder_id="MongoDB/voyage-4-nano"): super().__init__() self.inner = PixelArtDiT(dim=dim, n_layers=n_layers, n_heads=n_heads, mlp_ratio=mlp_ratio, patch_small=patch_small, patch_large=patch_large, large_from=large_from, in_channels=in_channels, text_len=text_len, enc_dim=enc_dim) def forward(self, x, t, txt, txt_mask, drop=None): return self.inner(x, t, txt, txt_mask, drop) @property def text_len(self): return self.inner.text_len @property def in_channels(self): return self.inner.in_channels class PixelArtDiTPipeline(DiffusionPipeline): def __init__(self, transformer, scheduler): super().__init__() self.register_modules(transformer=transformer, scheduler=scheduler) def _get_text_encoder(self, device): # frozen text encoder is downloaded from the Hub id stored in transformer/config.json if getattr(self, "_text_encoder", None) is None: self._tokenizer, self._text_encoder = load_text_encoder( self.transformer.config.text_encoder_id, device=device) else: self._text_encoder.to(device) return self._tokenizer, self._text_encoder @torch.no_grad() def __call__(self, prompt, height=32, width=32, num_inference_steps=30, guidance_scale=3.0, num_images_per_prompt=1, generator=None, output_type="pil", return_dict=True): if height != width or height not in (16, 32, 64): raise ValueError("PixelArtDiT supports square images of size 16, 32 or 64.") if isinstance(prompt, str): prompt = [prompt] prompts = [p for p in prompt for _ in range(num_images_per_prompt)] B = len(prompts) device = self._execution_device model = self.transformer dtype = model.dtype tokenizer, text_encoder = self._get_text_encoder(device) feats, mask = encode_prompts(tokenizer, text_encoder, prompts, model.text_len, device) use_cfg = guidance_scale != 1.0 if use_cfg: feats_in = torch.cat([feats, feats]) mask_in = torch.cat([mask, mask]) drop = torch.cat([torch.zeros(B, dtype=torch.bool, device=device), torch.ones(B, dtype=torch.bool, device=device)]) else: feats_in, mask_in, drop = feats, mask, None self.scheduler.set_timesteps(num_inference_steps, device=device) latents = randn_tensor((B, model.in_channels, height, width), generator=generator, device=device, dtype=torch.float32) for i, t in enumerate(self.scheduler.timesteps): sigma = self.scheduler.sigmas[i].to(device=device, dtype=torch.float32) x_in = torch.cat([latents, latents]) if use_cfg else latents t_in = sigma.expand(x_in.shape[0]) v = model(x_in.to(dtype), t_in, feats_in, mask_in, drop).float() if use_cfg: vc, vu = v.chunk(2) v = vu + guidance_scale * (vc - vu) latents = self.scheduler.step(v, t, latents).prev_sample if output_type == "latent": out = latents elif output_type == "np": out = ((latents.clamp(-1, 1) + 1) / 2).permute(0, 2, 3, 1).cpu().numpy() else: out = decode_latents(latents) if not return_dict: return (out,) return ImagePipelineOutput(images=out) else: PixelArtDiTModel = PixelArtDiT def load_pixel_art_dit(repo_or_path, device="cuda", dtype=torch.bfloat16): from safetensors.torch import load_file if os.path.isdir(repo_or_path): root = Path(repo_or_path) else: from huggingface_hub import snapshot_download root = Path(snapshot_download(repo_or_path)) with open(root / "transformer" / "config.json") as f: cfg = json.load(f) cfg = {k: v for k, v in cfg.items() if not k.startswith("_") and k != "text_encoder_id"} model = PixelArtDiT(**cfg) sd = load_file(root / "transformer" / "diffusion_pytorch_model.safetensors") sd = {k[len("inner."):] if k.startswith("inner.") else k: v for k, v in sd.items()} model.load_state_dict(sd) return model.to(device=device, dtype=dtype).eval()