Download trout.py from FrameXlabs/Trout: direct link, hf CLI and curl.
- Browser
- Download file 46 kB
-
https://huggingface.co/FrameXlabs/Trout/resolve/main/trout.py
- Command line
-
hf download hf://FrameXlabs/Trout/trout.py
-
curl -L -o trout.py https://huggingface.co/FrameXlabs/Trout/resolve/main/trout.py
46 kB
"""
trout.py -- a from-scratch diffusion language model for 16x16 pixel art.
Trout treats a 16x16 sprite as a 256-token "sentence" over a learned 512-color
palette (token 0 = transparent). A bidirectional transformer denoiser starts
from a fully masked canvas and unmasks all pixels in parallel over ~16 steps,
conditioned on a frozen CLIP text tower. No VAE, no latent space, no
autoregression -- absorbing-state masked (discrete) diffusion.
Architecture (hourglass dual-resolution denoiser, 47.6M params for the small
preset), reimplemented from scratch following the design published with
Tessera-small by Jagrat Patel (TheJagStudio), whose model card and open demo
space document every choice below:
* joint [registers | text tokens | pixel tokens] bidirectional attention
* hourglass: depth_in pixel-level blocks -> 2x2 patch-merged core blocks
-> depth_out pixel-level blocks, with a residual skip around the core
* QK-RMSNorm, 2D axial RoPE (theta=100) with sub-pixel-aligned patch centers
* Conv-SwiGLU: depthwise 3x3 conv on the gate branch over image tokens
* adaLN-single conditioning: shared MLP on (pooled text + mask ratio
Fourier features) with per-block learned modulation offsets and
zero-initialized gates (blocks start as identity)
* color-grounded token embeddings: learned vector + MLP(Fourier(OKLab))
* 4 register tokens in the prefix
* self-conditioning: the previous step's predicted color distribution is
projected and added at masked positions
Theory sources (the "knowledge" this model is built on):
* Absorbing-state / masked discrete diffusion:
Sahoo et al., "Simple and Effective Masked Diffusion Language Models",
NeurIPS 2024 (arXiv:2406.07524) -- MDLM
Shi et al., "Simplified and Generalized Masked Diffusion for Discrete
Data", NeurIPS 2024 (arXiv:2406.04329)
Labadie & Ye, "Absorbing State Discrete Diffusion" survey, 2025.
* Parallel confidence-ordered decoding: Chang et al., "MaskGIT: Masked
Generative Image Transformer", CVPR 2022.
* Hourglass transformer: Blecher et al., "Nougat" (arXiv:2308.13418).
* adaLN-single: Peebles & Xie, "DiT" (arXiv:2212.09748); PixArt-alpha.
* Registers: Darcet et al., "Vision Transformers Need Registers" (ICLR 2024).
* QK-norm: Gemma/Henry et al. "Training Transformers with ... QK-Norm".
* Self-conditioning: Chen, "Analog Bits" (arXiv:2208.04202).
* Muon optimizer: Jordan et al. (arXiv:2502.16982) and Moonshot
"Muon is Scalable" (arXiv:2508.02535) for the 0.2*sqrt(max(m,n)) scale.
* OKLab: Bjorn Ottosson, "A perceptual color space for image processing"
(2020) -- public-domain reference formulas.
Written by Z.ai for the user's "Trout" project; trained code lives in
train_trout.py. Apache-2.0.
"""
from __future__ import annotations
import json
import math
import os
from dataclasses import asdict, dataclass, field
from typing import Any, Dict, Iterator, List, Optional, Sequence, Tuple
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
DEFAULT_TEXT_ENCODER = "openai/clip-vit-base-patch32"
# ---------------------------------------------------------------------------
# configuration
# ---------------------------------------------------------------------------
@dataclass
class TroutConfig:
"""Architecture hyper-parameters for the Trout denoiser."""
vocab_size: int = 512 # palette entries incl. transparent (index 0); MASK id == vocab_size
grid_size: int = 16 # image is grid_size x grid_size tokens
dim: int = 512 # residual width
n_heads: int = 8
depth_in: int = 2 # pixel-level blocks before the 2x2 merge
depth_core: int = 8 # patch-level blocks on merged (grid/patch)^2 tokens
depth_out: int = 2 # pixel-level blocks after the unmerge
patch: int = 2 # hourglass merge factor (1 = flat model)
mlp_ratio: float = 3.0 # SwiGLU hidden = mlp_ratio * dim
n_registers: int = 4 # learnable register tokens in the prefix
text_dim: int = 512 # hidden size of the frozen CLIP text tower
text_len: int = 24 # text tokens fed to the denoiser
rope_theta: float = 100.0 # base of the 2D axial RoPE frequencies
conv_ffn: bool = True # depthwise 3x3 conv on the gate branch
color_freqs: int = 6 # Fourier frequencies for OKLab color embeddings
self_cond: bool = True # feed previous-step predictions back as input
@property
def mask_id(self) -> int:
return self.vocab_size
@property
def num_tokens(self) -> int:
return self.grid_size * self.grid_size
@property
def hourglass(self) -> bool:
return self.patch > 1 and self.depth_core > 0
@property
def head_dim(self) -> int:
return self.dim // self.n_heads
@property
def mlp_hidden(self) -> int:
return int(self.dim * self.mlp_ratio)
def validate(self) -> None:
assert self.dim % self.n_heads == 0
assert self.head_dim % 4 == 0, "head_dim must be divisible by 4 (2D axial RoPE)"
if self.hourglass:
assert self.grid_size % self.patch == 0
def to_dict(self) -> Dict[str, Any]:
return asdict(self)
@classmethod
def from_dict(cls, d: Dict[str, Any]) -> "TroutConfig":
import dataclasses as _dc
names = {f.name for f in _dc.fields(cls)}
keep = {k: v for k, v in d.items() if k in names}
cfg = cls(**keep)
cfg.validate()
return cfg
SMALL_PRESET: Dict[str, Any] = dict(
dim=512, n_heads=8, depth_in=2, depth_core=8, depth_out=2,
) # matches Tessera-small; ~47.6M parameters
# ---------------------------------------------------------------------------
# OKLab color math (Bjorn Ottosson's public-domain formulas)
# ---------------------------------------------------------------------------
def _srgb_to_linear_np(c):
return np.where(c <= 0.04045, c / 12.92, ((c + 0.055) / 1.055) ** 2.4)
def _linear_to_srgb_np(c):
c = np.clip(c, 0.0, 1.0)
return np.where(c <= 0.0031308, 12.92 * c, 1.055 * np.maximum(c, 1e-12) ** (1 / 2.4) - 0.055)
def srgb_to_oklab(rgb: np.ndarray) -> np.ndarray:
"""rgb: (..., 3) in [0,255]. Returns OKLab (..., 3) float32."""
x = np.asarray(rgb, dtype=np.float64) / 255.0
lin = _srgb_to_linear_np(x)
r, g, b = lin[..., 0], lin[..., 1], lin[..., 2]
l = 0.4122214708 * r + 0.5363325363 * g + 0.0514459929 * b
m = 0.2119034982 * r + 0.6806995451 * g + 0.1073969566 * b
s = 0.0883024619 * r + 0.2817188376 * g + 0.6299787005 * b
l_, m_, s_ = np.cbrt(np.maximum(l, 0)), np.cbrt(np.maximum(m, 0)), np.cbrt(np.maximum(s, 0))
L = 0.2104542553 * l_ + 0.7936177850 * m_ - 0.0040720468 * s_
A = 1.9779984951 * l_ - 2.4285922050 * m_ + 0.4505937099 * s_
B = 0.0259040371 * l_ + 0.7827717662 * m_ - 0.8086757660 * s_
return np.stack([L, A, B], -1).astype(np.float32)
def _torch_srgb_to_linear(c):
return torch.where(c <= 0.04045, c / 12.92, ((c + 0.055) / 1.055) ** 2.4)
def srgb_to_oklab_t(rgb: torch.Tensor) -> torch.Tensor:
x = rgb.float() / 255.0
lin = _torch_srgb_to_linear(x)
r, g, b = lin[..., 0], lin[..., 1], lin[..., 2]
l = 0.4122214708 * r + 0.5363325363 * g + 0.0514459929 * b
m = 0.2119034982 * r + 0.6806995451 * g + 0.1073969566 * b
s = 0.0883024619 * r + 0.2817188376 * g + 0.6299787005 * b
l_ = torch.pow(l.clamp(min=0), 1 / 3)
m_ = torch.pow(m.clamp(min=0), 1 / 3)
s_ = torch.pow(s.clamp(min=0), 1 / 3)
L = 0.2104542553 * l_ + 0.7936177850 * m_ - 0.0040720468 * s_
A = 1.9779984951 * l_ - 2.4285922050 * m_ + 0.4505937099 * s_
B = 0.0259040371 * l_ + 0.7827717662 * m_ - 0.8086757660 * s_
return torch.stack([L, A, B], -1).float()
def rgb_to_key(rgb: np.ndarray) -> np.ndarray:
rgb = rgb.astype(np.int64)
return (rgb[..., 0] << 16) | (rgb[..., 1] << 8) | rgb[..., 2]
def key_to_rgb(key: np.ndarray) -> np.ndarray:
key = np.asarray(key, dtype=np.int64)
return np.stack([(key >> 16) & 255, (key >> 8) & 255, key & 255], -1).astype(np.uint8)
# ---------------------------------------------------------------------------
# palette: 512 colors, frequency-weighted k-means in OKLab, snapped to data
# ---------------------------------------------------------------------------
def _weighted_kmeans(x: torch.Tensor, w: torch.Tensor, k: int, iters: int = 60,
seed: int = 0) -> torch.Tensor:
"""Weighted k-means with k-means++ init. x: (U,3), w: (U,). Returns (k,3)."""
g = torch.Generator().manual_seed(seed)
U = x.shape[0]
w = w.float()
first = torch.multinomial(w, 1, generator=g).item()
centers = [x[first]]
d2 = ((x - x[first]) ** 2).sum(1)
for _ in range(1, k):
p = (w * d2)
if p.sum() <= 0:
j = int(torch.randint(U, (1,), generator=g))
else:
j = torch.multinomial(p.clamp(min=0), 1, generator=g).item()
centers.append(x[j])
d2 = torch.minimum(d2, ((x - x[j]) ** 2).sum(1))
C = torch.stack(centers).clone()
for _ in range(iters):
# nearest assignment, chunked to bound memory
idx = torch.empty(U, dtype=torch.long)
mind = torch.empty(U)
c2 = (C * C).sum(1)
for s in range(0, U, 1 << 18):
p = x[s:s + (1 << 18)]
d = (p * p).sum(1, keepdim=True) - 2 * p @ C.T + c2[None]
dv, di = d.min(1)
idx[s:s + p.shape[0]] = di
mind[s:s + p.shape[0]] = dv.clamp(min=0)
sums = torch.zeros_like(C).index_add_(0, idx, x * w[:, None])
ws = torch.zeros(k).index_add_(0, idx, w)
newC = torch.where((ws <= 0)[:, None], C, sums / ws.clamp(min=1e-12)[:, None])
nempty = int((ws <= 0).sum())
if nempty > 0:
worst = torch.topk(w * mind, nempty).indices
newC[ws <= 0] = x[worst]
shift = (newC - C).norm(dim=1).max().item()
C = newC
if shift < 1e-7:
break
return C
class Palette:
"""A fixed list of colors. Index 0 = transparent; 1..K-1 opaque sRGB."""
def __init__(self, colors: np.ndarray, name: str = "custom"):
colors = np.asarray(colors, dtype=np.uint8).reshape(-1, 3).copy()
colors[0] = 0
self.colors = colors
self.name = name
lab = srgb_to_oklab(colors)
lab[0] = 0.0
self.lab = lab
@property
def size(self) -> int:
return int(self.colors.shape[0])
def __len__(self) -> int:
return self.size
def rgba_table(self) -> np.ndarray:
t = np.concatenate([self.colors, np.full((self.size, 1), 255, np.uint8)], 1)
t[0] = 0
return t
def tokens_to_rgba(self, tokens) -> np.ndarray:
if isinstance(tokens, torch.Tensor):
tokens = tokens.detach().cpu().numpy()
tokens = np.asarray(tokens).astype(np.int64)
table = np.concatenate([self.rgba_table(), np.array([[255, 0, 255, 255]], np.uint8)], 0)
return table[np.clip(tokens, 0, self.size)]
def nearest_index(self, rgb: np.ndarray) -> np.ndarray:
"""Nearest opaque entry (1..K-1) in OKLab for (M,3) uint8."""
if rgb.shape[0] == 0:
return np.zeros(0, np.int64)
pts = torch.from_numpy(srgb_to_oklab(rgb))
cen = torch.from_numpy(self.lab[1:])
idx, _ = self._nearest(pts, cen)
return idx.numpy() + 1
@staticmethod
def _nearest(points: torch.Tensor, centers: torch.Tensor, chunk: int = 1 << 18):
idx = torch.empty(points.shape[0], dtype=torch.long)
dist = torch.empty(points.shape[0])
c2 = (centers * centers).sum(1)
for s in range(0, points.shape[0], chunk):
p = points[s:s + chunk]
d = (p * p).sum(1, keepdim=True) - 2 * p @ centers.T + c2[None]
dv, di = d.min(1)
idx[s:s + p.shape[0]] = di
dist[s:s + p.shape[0]] = dv.clamp(min=0)
return idx, dist
def build_lut(self, keys: np.ndarray) -> Dict[str, np.ndarray]:
keys = np.unique(np.asarray(keys, dtype=np.int64))
idx = self.nearest_index(key_to_rgb(keys))
return {"keys": keys, "idx": idx.astype(np.uint16)}
def quantize_rgba(self, rgba: np.ndarray, alpha_threshold: int = 128,
lut: Optional[dict] = None) -> np.ndarray:
"""uint8 RGBA (...,4) -> uint16 tokens (...); alpha<thresh -> transparent."""
rgba = np.asarray(rgba)
shape = rgba.shape[:-1]
flat = rgba.reshape(-1, 4)
opaque = flat[:, 3] >= alpha_threshold
out = np.zeros(flat.shape[0], np.uint16)
if opaque.any():
keys = rgb_to_key(flat[opaque, :3])
if lut is None:
lut = self.build_lut(keys)
pos = np.searchsorted(lut["keys"], keys)
pos = np.clip(pos, 0, len(lut["keys"]) - 1)
hit = lut["keys"][pos] == keys
vals = lut["idx"][pos]
if not hit.all():
miss = ~hit
vals = vals.copy()
vals[miss] = self.nearest_index(key_to_rgb(keys[miss]))
out[opaque] = vals
return out.reshape(shape)
def soft_targets(self, tau: float = 0.05) -> torch.Tensor:
"""(K,K) row-stochastic matrix for perceptual label smoothing:
row c = softmax over opaque colors of -dE(c,.)^2 / tau^2."""
lab = torch.from_numpy(self.lab).double()
d2 = torch.cdist(lab, lab) ** 2
logits = -d2 / (tau * tau)
logits[:, 0] = -math.inf
S = torch.softmax(logits, dim=1)
S[0] = 0.0
S[0, 0] = 1.0
return S.float()
def to_dict(self) -> Dict[str, Any]:
return {"name": self.name, "colors": self.colors.tolist()}
@classmethod
def from_dict(cls, d: Dict[str, Any]) -> "Palette":
return cls(np.array(d["colors"], dtype=np.uint8), name=d.get("name", "custom"))
def save(self, path: str) -> None:
with open(path, "w") as f:
json.dump(self.to_dict(), f)
@classmethod
def load(cls, path: str) -> "Palette":
with open(path) as f:
return cls.from_dict(json.load(f))
def preview(self, cell: int = 12, cols: int = 32) -> "Image":
from PIL import Image
rows = math.ceil(self.size / cols)
img = np.full((rows * cell, cols * cell, 3), 255, np.uint8)
for i in range(self.size):
r, c = divmod(i, cols)
if i == 0:
yy, xx = np.mgrid[0:cell, 0:cell]
img[r * cell:(r + 1) * cell, c * cell:(c + 1) * cell] = np.where(
(((yy // 3) + (xx // 3)) % 2 == 0)[..., None], 200, 140)
else:
img[r * cell:(r + 1) * cell, c * cell:(c + 1) * cell] = self.colors[i]
return Image.fromarray(img)
@classmethod
def build_kmeans(cls, keys: np.ndarray, counts: np.ndarray, size: int = 512,
weight_power: float = 0.5, iters: int = 60, seed: int = 0) -> "Palette":
"""Data-driven palette: 1 transparent + size-1 opaque colors chosen by
frequency-weighted (tempered by weight_power) k-means in OKLab, with
centroids snapped to real dataset colors so exact hues survive."""
keys = np.asarray(keys, dtype=np.int64)
counts = np.asarray(counts, dtype=np.float64)
k = size - 1
rgb = key_to_rgb(keys)
if keys.shape[0] <= k:
colors = np.concatenate([np.zeros((1, 3), np.uint8), rgb], 0)
return cls._sorted(colors, name=f"exact{len(colors)}")
x = torch.from_numpy(srgb_to_oklab(rgb))
w = torch.from_numpy(counts ** weight_power)
C = _weighted_kmeans(x, w, k, iters=iters, seed=seed)
snap_idx, _ = cls._nearest(C, x)
snap_idx = snap_idx.numpy()
chosen = list(dict.fromkeys(snap_idx.tolist()))
if len(chosen) < k: # fill duplicate slots with worst-represented colors
cen = x[torch.as_tensor(chosen)]
_, dist = cls._nearest(x, cen)
order = torch.argsort(w * dist, descending=True).numpy()
for j in order:
if len(chosen) >= k:
break
if int(j) not in chosen:
chosen.append(int(j))
colors = np.concatenate([np.zeros((1, 3), np.uint8), rgb[np.array(chosen[:k])]], 0)
return cls._sorted(colors, name=f"kmeans{size}")
@classmethod
def _sorted(cls, colors: np.ndarray, name: str) -> "Palette":
"""Order opaque colors by hue bucket then lightness (readable palette)."""
opaque = colors[1:]
lab = srgb_to_oklab(opaque)
chroma = np.sqrt(lab[:, 1] ** 2 + lab[:, 2] ** 2)
hue = np.arctan2(lab[:, 2], lab[:, 1])
bucket = np.where(chroma < 0.03, -1, np.floor((hue + np.pi) / (2 * np.pi) * 12)).astype(np.int64)
order = np.lexsort((lab[:, 0], bucket))
return cls(np.concatenate([colors[:1], opaque[order]], 0), name=name)
def quantization_report(palette: Palette, rgba: np.ndarray, tokens: np.ndarray,
max_pixels: int = 2_000_000) -> Dict[str, float]:
"""Mean/p95 OKLab error introduced by quantization on opaque pixels."""
flat = rgba.reshape(-1, 4)
tok = tokens.reshape(-1)
if flat.shape[0] > max_pixels:
sel = np.random.default_rng(0).integers(0, flat.shape[0], max_pixels)
flat, tok = flat[sel], tok[sel]
op = flat[:, 3] >= 128
if not op.any():
return {"mean_dE": 0.0, "p95_dE": 0.0, "exact_frac": 1.0}
src = srgb_to_oklab(flat[op, :3])
dst = palette.lab[tok[op].astype(np.int64)]
d = np.linalg.norm(src - dst, axis=1)
exact = (palette.colors[tok[op].astype(np.int64)] == flat[op, :3]).all(1)
return {"mean_dE": float(d.mean()), "p95_dE": float(np.percentile(d, 95)),
"max_dE": float(d.max()), "exact_frac": float(exact.mean())}
# ---------------------------------------------------------------------------
# model building blocks
# ---------------------------------------------------------------------------
class RMSNorm(nn.Module):
def __init__(self, dim: int, eps: float = 1e-6, affine: bool = True):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(dim)) if affine else None
def forward(self, x: torch.Tensor) -> torch.Tensor:
dtype = x.dtype
x32 = x.float()
y = x32 * torch.rsqrt(x32.pow(2).mean(-1, keepdim=True) + self.eps)
if self.weight is not None:
y = y * self.weight.float()
return y.to(dtype)
def _rope_freqs(head_dim: int, theta: float) -> torch.Tensor:
n = head_dim // 4 # rotation pairs per axis
return 1.0 / (theta ** (torch.arange(n, dtype=torch.float64) / n))
def build_rope(n_prefix: int, grid: int, head_dim: int, theta: float,
stride: int = 1) -> Tuple[torch.Tensor, torch.Tensor]:
"""cos/sin tables for 2D axial RoPE. Image token (i,j) at a level with
`stride` pixels per token sits at pixel-space centre
(i*stride + (stride-1)/2, j*stride + (stride-1)/2); prefix angle 0."""
f = _rope_freqs(head_dim, theta)
ii, jj = torch.meshgrid(torch.arange(grid, dtype=torch.float64),
torch.arange(grid, dtype=torch.float64), indexing="ij")
off = (stride - 1) / 2.0
rows = (ii.reshape(-1) * stride + off)[:, None] * f[None]
cols = (jj.reshape(-1) * stride + off)[:, None] * f[None]
ang = torch.cat([rows, cols], 1) # (grid^2, head_dim/2)
ang = torch.cat([torch.zeros(n_prefix, ang.shape[1], dtype=ang.dtype), ang], 0)
return ang.cos().float(), ang.sin().float()
def apply_rope(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor:
# x: (B,H,N,hd); cos/sin: (N, hd/2)
x1, x2 = x.float().chunk(2, dim=-1)
out = torch.cat([x1 * cos - x2 * sin, x1 * sin + x2 * cos], dim=-1)
return out.to(x.dtype)
def fourier_features(x: torch.Tensor, n_freqs: int) -> torch.Tensor:
"""x: (...,C) roughly in [-1,1] -> (..., C*2*n_freqs) sin/cos features."""
freqs = (2.0 ** torch.arange(n_freqs, dtype=torch.float32)) * math.pi
ang = x.float().unsqueeze(-1) * freqs
return torch.cat([ang.sin(), ang.cos()], dim=-1).flatten(-2)
class Attention(nn.Module):
def __init__(self, dim: int, n_heads: int):
super().__init__()
self.n_heads = n_heads
self.head_dim = dim // n_heads
self.qkv = nn.Linear(dim, 3 * dim, bias=False)
self.proj = nn.Linear(dim, dim, bias=False)
self.q_norm = RMSNorm(self.head_dim)
self.k_norm = RMSNorm(self.head_dim)
def forward(self, x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor:
B, N, D = x.shape
qkv = self.qkv(x).view(B, N, 3, self.n_heads, self.head_dim).permute(2, 0, 3, 1, 4)
q, k, v = qkv.unbind(0)
q = apply_rope(self.q_norm(q), cos, sin)
k = apply_rope(self.k_norm(k), cos, sin)
o = F.scaled_dot_product_attention(q, k, v)
return self.proj(o.transpose(1, 2).reshape(B, N, D))
class ConvSwiGLU(nn.Module):
"""SwiGLU whose gate branch gets a residual depthwise 3x3 conv over the
image grid (local pixel structure)."""
def __init__(self, dim: int, hidden: int, conv: bool = True):
super().__init__()
self.w12 = nn.Linear(dim, 2 * hidden, bias=False)
self.w3 = nn.Linear(hidden, dim, bias=False)
self.dw = nn.Conv2d(hidden, hidden, 3, padding=1, groups=hidden, bias=True) if conv else None
def forward(self, x: torch.Tensor, n_prefix: int, hw: Tuple[int, int]) -> torch.Tensor:
a, g = self.w12(x).chunk(2, dim=-1)
if self.dw is not None:
B, _, C = g.shape
H, W = hw
g_pre, g_img = g[:, :n_prefix], g[:, n_prefix:]
conv = self.dw(g_img.transpose(1, 2).reshape(B, C, H, W)).flatten(2).transpose(1, 2)
g = torch.cat([g_pre, g_img + conv.to(g_img.dtype)], dim=1)
return self.w3(a * F.silu(g))
class Block(nn.Module):
def __init__(self, cfg: TroutConfig):
super().__init__()
self.norm1 = RMSNorm(cfg.dim, affine=False)
self.attn = Attention(cfg.dim, cfg.n_heads)
self.norm2 = RMSNorm(cfg.dim, affine=False)
self.mlp = ConvSwiGLU(cfg.dim, cfg.mlp_hidden, conv=cfg.conv_ffn)
# per-block offsets added to the shared adaLN-single modulation
# (shift/scale/gate) x (attn/mlp)
self.mod_bias = nn.Parameter(torch.zeros(6, cfg.dim))
def forward(self, x, mod, cos, sin, n_prefix: int, hw: Tuple[int, int]):
m = (mod + self.mod_bias).unsqueeze(2) # (B,6,1,D)
sh1, sc1, g1, sh2, sc2, g2 = m.unbind(1)
h = self.norm1(x) * (1 + sc1) + sh1
x = x + g1 * self.attn(h, cos, sin)
h = self.norm2(x) * (1 + sc2) + sh2
x = x + g2 * self.mlp(h, n_prefix, hw)
return x
# ---------------------------------------------------------------------------
# the model
# ---------------------------------------------------------------------------
class TroutModel(nn.Module):
"""Dual-resolution (hourglass) masked-diffusion transformer denoiser."""
def __init__(self, cfg: TroutConfig, palette_lab: Optional[torch.Tensor] = None,
palette_rgb: Optional[torch.Tensor] = None):
super().__init__()
cfg.validate()
self.cfg = cfg
K, D, G = cfg.vocab_size, cfg.dim, cfg.grid_size
self.mask_id = cfg.mask_id
if palette_lab is None:
palette_lab = torch.zeros(K, 3)
if palette_rgb is None:
palette_rgb = torch.zeros(K, 3, dtype=torch.uint8)
assert palette_lab.shape == (K, 3), f"palette has {palette_lab.shape[0]} colors, config says {K}"
# the palette travels with the checkpoint
self.register_buffer("palette_lab", palette_lab.float().clone())
self.register_buffer("palette_rgb", palette_rgb.to(torch.uint8).clone())
# ---- token embeddings (color-grounded) ----
self.tok_emb = nn.Embedding(K + 1, D) # +1 for MASK
n_feat = 3 + 3 * 2 * cfg.color_freqs + 1
self.color_mlp = nn.Sequential(nn.Linear(n_feat, D), nn.SiLU(), nn.Linear(D, D))
self.pos_emb = nn.Parameter(torch.zeros(1, G * G, D))
self.sc_proj = nn.Linear(K, D, bias=False) if cfg.self_cond else None
# ---- text conditioning ----
self.text_norm = RMSNorm(cfg.text_dim)
self.text_proj = nn.Linear(cfg.text_dim, D)
self.null_text = nn.Parameter(torch.zeros(1, cfg.text_len, D))
self.pool_mlp = nn.Sequential(RMSNorm(cfg.text_dim), nn.Linear(cfg.text_dim, D), nn.SiLU(), nn.Linear(D, D))
self.null_cond = nn.Parameter(torch.zeros(1, D))
self.ratio_mlp = nn.Sequential(nn.Linear(2 * 8, D), nn.SiLU(), nn.Linear(D, D))
self.registers = nn.Parameter(torch.zeros(1, cfg.n_registers, D))
self.t_block = nn.Sequential(nn.SiLU(), nn.Linear(D, 6 * D))
# ---- hourglass backbone ----
p = cfg.patch
self.blocks_in = nn.ModuleList([Block(cfg) for _ in range(cfg.depth_in)])
self.down = nn.Sequential(RMSNorm(p * p * D), nn.Linear(p * p * D, D))
self.blocks_core = nn.ModuleList([Block(cfg) for _ in range(cfg.depth_core)])
self.up = nn.Linear(D, p * p * D)
self.blocks_out = nn.ModuleList([Block(cfg) for _ in range(cfg.depth_out)])
self.norm_out = RMSNorm(D, affine=False)
self.final_mod = nn.Sequential(nn.SiLU(), nn.Linear(D, 2 * D))
self.head = nn.Linear(D, K)
# ---- fixed tables ----
n_pre = cfg.n_registers + cfg.text_len
self.n_prefix = n_pre
cf, sf = build_rope(n_pre, G, cfg.head_dim, cfg.rope_theta, stride=1)
self.register_buffer("rope_cos", cf, persistent=False)
self.register_buffer("rope_sin", sf, persistent=False)
cc, sc = build_rope(n_pre, G // cfg.patch, cfg.head_dim, cfg.rope_theta, stride=cfg.patch)
self.register_buffer("rope_cos_core", cc, persistent=False)
self.register_buffer("rope_sin_core", sc, persistent=False)
self.register_buffer("color_feats", self._color_features(self.palette_lab), persistent=False)
self.grad_checkpointing = False
self._init_weights()
# ------------------------------------------------------------------
def _color_features(self, lab: torch.Tensor) -> torch.Tensor:
x = torch.stack([lab[:, 0] * 2 - 1, lab[:, 1] / 0.4, lab[:, 2] / 0.4], -1)
alpha = torch.ones(lab.shape[0], 1)
alpha[0] = 0.0 # token 0 = transparent
x = x * alpha
return torch.cat([x, fourier_features(x, self.cfg.color_freqs), alpha], -1)
def set_palette(self, palette_lab: torch.Tensor, palette_rgb: torch.Tensor) -> None:
self.palette_lab.copy_(palette_lab.float())
self.palette_rgb.copy_(palette_rgb.to(torch.uint8))
self.color_feats = self._color_features(self.palette_lab.detach().cpu()).to(self.palette_lab.device)
def _load_from_state_dict(self, state_dict, prefix, *args, **kwargs):
super()._load_from_state_dict(state_dict, prefix, *args, **kwargs)
self.color_feats = self._color_features(self.palette_lab.detach().cpu()).to(self.palette_lab.device)
def _init_weights(self) -> None:
def init(m):
if isinstance(m, nn.Linear):
nn.init.trunc_normal_(m.weight, std=0.02)
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.Embedding):
nn.init.normal_(m.weight, std=0.02)
self.apply(init)
nn.init.trunc_normal_(self.pos_emb, std=0.02)
nn.init.normal_(self.null_text, std=0.02)
nn.init.normal_(self.null_cond, std=0.02)
nn.init.normal_(self.registers, std=0.02)
# zero-init: conditioning projections, gates, head, self-cond, up path
for m in [self.t_block[-1], self.final_mod[-1], self.head]:
nn.init.zeros_(m.weight)
nn.init.zeros_(m.bias)
if self.sc_proj is not None:
nn.init.zeros_(self.sc_proj.weight)
nn.init.zeros_(self.up.weight)
nn.init.zeros_(self.up.bias)
for blk in list(self.blocks_in) + list(self.blocks_core) + list(self.blocks_out):
if blk.mlp.dw is not None:
nn.init.zeros_(blk.mlp.dw.weight)
nn.init.zeros_(blk.mlp.dw.bias)
# ------------------------------------------------------------------
def _merge(self, h: torch.Tensor) -> torch.Tensor:
B, N, D = h.shape
G, p = self.cfg.grid_size, self.cfg.patch
g = G // p
h = h.view(B, g, p, g, p, D).permute(0, 1, 3, 2, 4, 5).reshape(B, g * g, p * p * D)
return self.down(h)
def _split(self, h: torch.Tensor) -> torch.Tensor:
B = h.shape[0]
G, p, D = self.cfg.grid_size, self.cfg.patch, self.cfg.dim
g = G // p
h = self.up(h).view(B, g, g, p, p, D).permute(0, 1, 3, 2, 4, 5).reshape(B, G * G, D)
return h
def _run(self, blk: Block, xs, mod, cos, sin, n_pre: int, hw):
if self.grad_checkpointing and self.training and torch.is_grad_enabled():
return torch.utils.checkpoint.checkpoint(blk, xs, mod, cos, sin, n_pre, hw,
use_reentrant=False)
return blk(xs, mod, cos, sin, n_pre, hw)
def embed_table(self) -> torch.Tensor:
K = self.cfg.vocab_size
w = self.tok_emb.weight
colors = w[:K] + self.color_mlp(self.color_feats).to(w.dtype)
return torch.cat([colors, w[K:]], 0)
def forward(self, x: torch.Tensor, text_states: torch.Tensor, text_pooled: torch.Tensor,
mask_ratio: Optional[torch.Tensor] = None,
sc_probs: Optional[torch.Tensor] = None,
drop_text: Optional[torch.Tensor] = None) -> torch.Tensor:
"""
x: (B,N) long palette ids or mask_id
text_states: (B,L,text_dim) frozen text-encoder token states
text_pooled: (B,text_dim) pooled text embedding
mask_ratio: (B,) fraction masked (computed from x if None)
sc_probs: (B,N,K) previous-step predicted distribution or None
drop_text: (B,) bool -> use learned null condition (CFG)
returns logits (B,N,K) float32
"""
cfg = self.cfg
B, N = x.shape
D, G = cfg.dim, cfg.grid_size
is_mask = x == self.mask_id
if mask_ratio is None:
mask_ratio = is_mask.float().mean(1)
# ---- image tokens ----
h = F.embedding(x, self.embed_table()).float() + self.pos_emb
if self.sc_proj is not None and sc_probs is not None:
h = h + self.sc_proj(sc_probs.to(self.sc_proj.weight.dtype)).float() * is_mask.unsqueeze(-1)
# ---- text tokens + global conditioning ----
t = self.text_proj(self.text_norm(text_states.float())).float()
c = self.pool_mlp(text_pooled.float()).float()
if drop_text is not None:
t = torch.where(drop_text.view(B, 1, 1), self.null_text, t)
c = torch.where(drop_text.view(B, 1), self.null_cond, c)
r = mask_ratio.float().view(B, 1) * 2 - 1
g = c + self.ratio_mlp(fourier_features(r, 8).view(B, -1)).float()
mod = self.t_block(g).float().view(B, 6, D)
prefix = torch.cat([self.registers.expand(B, -1, -1), t], 1)
n_pre = prefix.shape[1]
xs = torch.cat([prefix, h], 1)
# ---- pixel level (in) ----
for blk in self.blocks_in:
xs = self._run(blk, xs, mod, self.rope_cos, self.rope_sin, n_pre, (G, G))
# ---- patch level (core) ----
pre, img = xs[:, :n_pre], xs[:, n_pre:]
skip = img
xs = torch.cat([pre, self._merge(img).float()], 1)
gc = G // cfg.patch
for blk in self.blocks_core:
xs = self._run(blk, xs, mod, self.rope_cos_core, self.rope_sin_core, n_pre, (gc, gc))
pre, core = xs[:, :n_pre], xs[:, n_pre:]
xs = torch.cat([pre, skip + self._split(core).float()], 1)
# ---- pixel level (out) ----
for blk in self.blocks_out:
xs = self._run(blk, xs, mod, self.rope_cos, self.rope_sin, n_pre, (G, G))
img = xs[:, n_pre:]
shift, scale = self.final_mod(g).float().unsqueeze(1).chunk(2, dim=-1)
img = self.norm_out(img) * (1 + scale) + shift
return self.head(img).float()
def num_params(self, trainable_only: bool = True) -> int:
return sum(p.numel() for p in self.parameters() if p.requires_grad or not trainable_only)
# ---------------------------------------------------------------------------
# sampler: parallel confidence-ordered unmasking (MaskGIT-style) with CFG,
# Gumbel-annealed choice, spatial NMS and self-conditioning feedback
# ---------------------------------------------------------------------------
def filter_logits(logits: torch.Tensor, top_k: int = 0, top_p: float = 1.0,
min_p: float = 0.0) -> torch.Tensor:
if top_k and top_k > 0 and top_k < logits.shape[-1]:
kth = torch.topk(logits, top_k, dim=-1).values[..., -1:]
logits = logits.masked_fill(logits < kth, -math.inf)
if min_p and min_p > 0:
p = logits.softmax(-1)
logits = logits.masked_fill(p < min_p * p.amax(-1, keepdim=True), -math.inf)
if top_p is not None and top_p < 1.0:
sorted_l, idx = torch.sort(logits, dim=-1, descending=True)
sp = sorted_l.softmax(-1)
remove = (sp.cumsum(-1) - sp) > top_p
sorted_l = sorted_l.masked_fill(remove, -math.inf)
logits = torch.full_like(logits, -math.inf).scatter(-1, idx, sorted_l)
return logits
def cfg_weight(scale: float, schedule: str, frac: float) -> float:
if schedule == "constant":
return scale
if schedule == "linear":
return 1.0 + (scale - 1.0) * frac
if schedule == "cosine":
return 1.0 + (scale - 1.0) * (1 - math.cos(math.pi * frac)) / 2
raise ValueError(f"unknown cfg schedule {schedule!r}")
def _gumbel(shape, device, generator):
u = torch.rand(shape, device=device, generator=generator).clamp_(1e-10, 1 - 1e-10)
return -torch.log(-torch.log(u))
def _nms_eligible(score: torch.Tensor, masked: torch.Tensor, certain: torch.Tensor,
grid: int) -> torch.Tensor:
B, N = score.shape
cand = masked & ~certain
s = torch.where(cand, score, torch.full_like(score, -math.inf)).view(B, 1, grid, grid)
local_max = F.max_pool2d(s, 3, stride=1, padding=1).view(B, N)
return certain | (cand & (score >= local_max))
@torch.no_grad()
def sample(model: TroutModel, text_states: torch.Tensor, text_pooled: torch.Tensor,
steps: int = 16, cfg_scale: float = 3.0, cfg_schedule: str = "constant",
temperature: float = 1.0, temperature_end: Optional[float] = None,
top_k: int = 0, top_p: float = 1.0, min_p: float = 0.0,
choice_temperature: float = 2.0, spatial_nms: bool = True,
nms_threshold: float = 0.9, polish_steps: int = 0, polish_frac: float = 0.08,
init_tokens: Optional[torch.Tensor] = None, keep_mask: Optional[torch.Tensor] = None,
self_cond: bool = True, generator: Optional[torch.Generator] = None,
return_history: bool = False):
"""Generate palette-token canvases (B, N); inpainting via init_tokens+keep_mask."""
cfg = model.cfg
device = text_pooled.device
B, N, K, M, G = text_pooled.shape[0], cfg.num_tokens, cfg.vocab_size, cfg.mask_id, cfg.grid_size
if init_tokens is not None and keep_mask is not None:
init_tokens = init_tokens.to(device).long()
x = torch.where(keep_mask.to(device).bool(), init_tokens, torch.full_like(init_tokens, M))
else:
x = torch.full((B, N), M, dtype=torch.long, device=device)
masked = x == M
conf_map = torch.where(masked, torch.zeros(B, N, device=device),
torch.full((B, N), math.inf, device=device))
n_start = masked.sum(1)
use_cfg = cfg_scale is not None and abs(cfg_scale - 1.0) > 1e-6
if use_cfg:
ts2 = torch.cat([text_states, text_states])
tp2 = torch.cat([text_pooled, text_pooled])
drop2 = torch.cat([torch.zeros(B, dtype=torch.bool, device=device),
torch.ones(B, dtype=torch.bool, device=device)])
sc = torch.zeros(B, N, K, device=device) if (cfg.self_cond and self_cond) else None
history = []
def predict(x_in, sc_in, frac):
ratio = (x_in == M).float().mean(1)
if use_cfg:
sc2 = torch.cat([sc_in, sc_in]) if sc_in is not None else None
out = model(torch.cat([x_in, x_in]), ts2, tp2, torch.cat([ratio, ratio]),
sc2, drop2).float()
lc, lu = out.chunk(2)
w = cfg_weight(cfg_scale, cfg_schedule, frac)
return lu + w * (lc - lu)
return model(x_in, text_states, text_pooled, ratio, sc_in, None).float()
for k in range(steps):
if not masked.any():
break
frac = (k + 1) / steps
logits = predict(x, sc, frac)
probs_full = logits.softmax(-1)
T = temperature if temperature_end is None else temperature + (temperature_end - temperature) * (k / max(1, steps - 1))
if T <= 0:
sampled = logits.argmax(-1)
p_sel = probs_full.gather(-1, sampled.unsqueeze(-1)).squeeze(-1)
else:
probs = filter_logits(logits / T, top_k, top_p, min_p).softmax(-1)
sampled = torch.multinomial(probs.view(-1, K), 1, generator=generator).view(B, N)
p_sel = probs.gather(-1, sampled.unsqueeze(-1)).squeeze(-1)
n_masked = masked.sum(1)
remaining = torch.floor(n_start.float() * math.cos(math.pi / 2 * frac)).long()
if k == steps - 1:
remaining = torch.zeros_like(remaining)
n_commit = torch.minimum((n_masked - remaining).clamp(min=1), n_masked)
score = torch.log(p_sel.clamp(min=1e-12))
if choice_temperature > 0: # Gumbel-annealed choice (MaskGIT)
score = score + choice_temperature * (1.0 - frac) * _gumbel(score.shape, device, generator)
priority = score
if spatial_nms: # don't commit two uncertain neighbours in one step
elig = _nms_eligible(score, masked, p_sel >= nms_threshold, G)
priority = score + elig.float() * 1e4
priority = torch.where(masked, priority, torch.full_like(priority, -math.inf))
ranks = priority.argsort(1, descending=True).argsort(1)
commit = (ranks < n_commit.unsqueeze(1)) & masked
x = torch.where(commit, sampled, x)
conf_map = torch.where(commit, p_sel, conf_map)
masked = masked & ~commit
if sc is not None:
sc = probs_full
if return_history:
history.append(x.clone())
# ---- optional polish: re-mask the least confident committed pixels ----
n_re = max(1, int(round(polish_frac * N)))
for _ in range(polish_steps):
low = torch.topk(-conf_map, n_re, dim=1).indices
remask = torch.zeros(B, N, dtype=torch.bool, device=device).scatter_(1, low, True)
remask &= torch.isfinite(conf_map)
x_in = torch.where(remask, torch.full_like(x, M), x)
logits = predict(x_in, sc, 1.0)
probs_full = logits.softmax(-1)
new = logits.argmax(-1)
x = torch.where(remask, new, x)
conf_map = torch.where(remask, probs_full.gather(-1, new.unsqueeze(-1)).squeeze(-1), conf_map)
if sc is not None:
sc = probs_full
if return_history:
history.append(x.clone())
if return_history:
return x, history
return x
# ---------------------------------------------------------------------------
# checkpoint helpers (same schema semantics as the Tessera distribution:
# ema weights + model_config + palette + text-encoder id + step)
# ---------------------------------------------------------------------------
def save_checkpoint(path: str, model: TroutModel, palette: Palette, step: int,
ema_state: Optional[Dict[str, torch.Tensor]] = None,
raw_state: Optional[Dict[str, torch.Tensor]] = None,
text_encoder: str = DEFAULT_TEXT_ENCODER, half: bool = False) -> None:
def prep(sd):
out = {}
for k, v in sd.items():
out[k] = (v.half() if half and v.is_floating_point() else v).cpu()
return out
ckpt = {"model_config": model.cfg.to_dict(), "palette": palette.to_dict(),
"text_encoder": text_encoder, "step": step}
if ema_state is not None:
ckpt["ema"] = prep(ema_state)
if raw_state is not None:
ckpt["model"] = prep(raw_state)
torch.save(ckpt, path)
def load_model_from_checkpoint(path: str, device="cpu", use_ema: bool = True):
ckpt = torch.load(path, map_location="cpu", weights_only=False)
mcfg = TroutConfig.from_dict(ckpt["model_config"])
palette = Palette.from_dict(ckpt["palette"])
model = TroutModel(mcfg, torch.from_numpy(palette.lab), torch.from_numpy(palette.colors))
key = "ema" if (use_ema and "ema" in ckpt) else "model"
sd = {k: v.float() if v.is_floating_point() else v for k, v in ckpt[key].items()}
model.load_state_dict(sd)
model.to(device).eval().requires_grad_(False)
return model, palette, ckpt
# ---------------------------------------------------------------------------
# visualization helpers
# ---------------------------------------------------------------------------
def tokens_to_images(tokens, palette: Palette, grid: int = 16) -> List:
from PIL import Image
if isinstance(tokens, torch.Tensor):
tokens = tokens.detach().cpu().numpy()
toks = np.asarray(tokens).reshape(-1, grid * grid)
imgs = []
for row in toks:
rgba = palette.tokens_to_rgba(row.reshape(grid, grid))
imgs.append(Image.fromarray(rgba, mode="RGBA"))
return imgs
def upscale(image, scale: int = 16, checker: bool = True):
"""Nearest-neighbour upscale with an optional checkerboard background."""
from PIL import Image
w, h = image.size
big = image.resize((w * scale, h * scale), Image.NEAREST)
if not checker:
return big
bg = Image.new("RGBA", big.size, (255, 255, 255, 255))
s = 4 * 2 # checker cell in big pixels
yy, xx = np.mgrid[0:big.size[1], 0:big.size[0]]
mask = (((yy // s) + (xx // s)) % 2 == 0)
light = Image.fromarray(np.where(mask[..., None], np.array([238, 238, 238, 255], np.uint8),
np.array([204, 204, 204, 255], np.uint8)), "RGBA")
light.alpha_composite(big)
return light
def make_grid(images, ncols: int = 8, scale: int = 8, checker: bool = True):
from PIL import Image
cells = [upscale(im, scale, checker) for im in images]
cw, ch = cells[0].size
rows = math.ceil(len(cells) / ncols)
out = Image.new("RGBA", (ncols * cw, rows * ch), (255, 255, 255, 255))
for i, c in enumerate(cells):
out.alpha_composite(c, ((i % ncols) * cw, (i // ncols) * ch))
return out
# ---------------------------------------------------------------------------
# frozen CLIP text tower wrapper
# ---------------------------------------------------------------------------
def load_tokenizer(name: str):
from transformers import AutoTokenizer
return AutoTokenizer.from_pretrained(name)
def tokenize_texts(tokenizer, texts: Sequence[str], text_len: int) -> Tuple[np.ndarray, np.ndarray]:
"""Fixed-length tokenization; pad with pad id; over-long prompts truncated
while keeping the final EOS token."""
eos = tokenizer.eos_token_id
pad = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else eos
n = len(texts)
out = np.full((n, text_len), pad, dtype=np.int32)
lengths = np.zeros(n, dtype=np.int32)
enc = tokenizer(list(texts), add_special_tokens=True, truncation=False)["input_ids"]
for i, ids in enumerate(enc):
lengths[i] = len(ids)
if len(ids) > text_len:
ids = list(ids[:text_len - 1]) + [eos]
out[i, :len(ids)] = ids
return out, lengths
class FrozenTextEncoder:
"""Wraps a Hugging Face CLIP text tower. Not an nn.Module on purpose so its
weights never end up in the trainable model, optimizer or checkpoints."""
def __init__(self, name: str = DEFAULT_TEXT_ENCODER, text_len: int = 24,
device="cpu", dtype: torch.dtype = torch.float32):
from transformers import CLIPTextModel
from transformers.utils import logging as hf_logging
self.name = name
self.text_len = text_len
self.device = torch.device(device)
self.dtype = dtype
self.tokenizer = load_tokenizer(name)
prev = hf_logging.get_verbosity()
hf_logging.set_verbosity_error()
try:
model = CLIPTextModel.from_pretrained(name)
finally:
hf_logging.set_verbosity(prev)
self.model = model.to(device=self.device, dtype=dtype).eval().requires_grad_(False)
self.dim = int(model.config.hidden_size)
def tokenize(self, texts: Sequence[str]) -> torch.Tensor:
ids, _ = tokenize_texts(self.tokenizer, list(texts), self.text_len)
return torch.from_numpy(ids).long().to(self.device)
@torch.no_grad()
def __call__(self, ids: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
out = self.model(input_ids=ids.long().to(self.device))
return out.last_hidden_state, out.pooler_output
@torch.no_grad()
def encode(self, texts: Sequence[str]) -> Tuple[torch.Tensor, torch.Tensor]:
return self(self.tokenize(texts))
"""
trout.py -- a from-scratch diffusion language model for 16x16 pixel art.
Trout treats a 16x16 sprite as a 256-token "sentence" over a learned 512-color
palette (token 0 = transparent). A bidirectional transformer denoiser starts
from a fully masked canvas and unmasks all pixels in parallel over ~16 steps,
conditioned on a frozen CLIP text tower. No VAE, no latent space, no
autoregression -- absorbing-state masked (discrete) diffusion.
Architecture (hourglass dual-resolution denoiser, 47.6M params for the small
preset), reimplemented from scratch following the design published with
Tessera-small by Jagrat Patel (TheJagStudio), whose model card and open demo
space document every choice below:
* joint [registers | text tokens | pixel tokens] bidirectional attention
* hourglass: depth_in pixel-level blocks -> 2x2 patch-merged core blocks
-> depth_out pixel-level blocks, with a residual skip around the core
* QK-RMSNorm, 2D axial RoPE (theta=100) with sub-pixel-aligned patch centers
* Conv-SwiGLU: depthwise 3x3 conv on the gate branch over image tokens
* adaLN-single conditioning: shared MLP on (pooled text + mask ratio
Fourier features) with per-block learned modulation offsets and
zero-initialized gates (blocks start as identity)
* color-grounded token embeddings: learned vector + MLP(Fourier(OKLab))
* 4 register tokens in the prefix
* self-conditioning: the previous step's predicted color distribution is
projected and added at masked positions
Theory sources (the "knowledge" this model is built on):
* Absorbing-state / masked discrete diffusion:
Sahoo et al., "Simple and Effective Masked Diffusion Language Models",
NeurIPS 2024 (arXiv:2406.07524) -- MDLM
Shi et al., "Simplified and Generalized Masked Diffusion for Discrete
Data", NeurIPS 2024 (arXiv:2406.04329)
Labadie & Ye, "Absorbing State Discrete Diffusion" survey, 2025.
* Parallel confidence-ordered decoding: Chang et al., "MaskGIT: Masked
Generative Image Transformer", CVPR 2022.
* Hourglass transformer: Blecher et al., "Nougat" (arXiv:2308.13418).
* adaLN-single: Peebles & Xie, "DiT" (arXiv:2212.09748); PixArt-alpha.
* Registers: Darcet et al., "Vision Transformers Need Registers" (ICLR 2024).
* QK-norm: Gemma/Henry et al. "Training Transformers with ... QK-Norm".
* Self-conditioning: Chen, "Analog Bits" (arXiv:2208.04202).
* Muon optimizer: Jordan et al. (arXiv:2502.16982) and Moonshot
"Muon is Scalable" (arXiv:2508.02535) for the 0.2*sqrt(max(m,n)) scale.
* OKLab: Bjorn Ottosson, "A perceptual color space for image processing"
(2020) -- public-domain reference formulas.
Written by Z.ai for the user's "Trout" project; trained code lives in
train_trout.py. Apache-2.0.
"""
from __future__ import annotations
import json
import math
import os
from dataclasses import asdict, dataclass, field
from typing import Any, Dict, Iterator, List, Optional, Sequence, Tuple
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
DEFAULT_TEXT_ENCODER = "openai/clip-vit-base-patch32"
# ---------------------------------------------------------------------------
# configuration
# ---------------------------------------------------------------------------
class TroutConfig:
"""Architecture hyper-parameters for the Trout denoiser."""
vocab_size: int = 512 # palette entries incl. transparent (index 0); MASK id == vocab_size
grid_size: int = 16 # image is grid_size x grid_size tokens
dim: int = 512 # residual width
n_heads: int = 8
depth_in: int = 2 # pixel-level blocks before the 2x2 merge
depth_core: int = 8 # patch-level blocks on merged (grid/patch)^2 tokens
depth_out: int = 2 # pixel-level blocks after the unmerge
patch: int = 2 # hourglass merge factor (1 = flat model)
mlp_ratio: float = 3.0 # SwiGLU hidden = mlp_ratio * dim
n_registers: int = 4 # learnable register tokens in the prefix
text_dim: int = 512 # hidden size of the frozen CLIP text tower
text_len: int = 24 # text tokens fed to the denoiser
rope_theta: float = 100.0 # base of the 2D axial RoPE frequencies
conv_ffn: bool = True # depthwise 3x3 conv on the gate branch
color_freqs: int = 6 # Fourier frequencies for OKLab color embeddings
self_cond: bool = True # feed previous-step predictions back as input
def mask_id(self) -> int:
return self.vocab_size
def num_tokens(self) -> int:
return self.grid_size * self.grid_size
def hourglass(self) -> bool: