Glint-Research/Anthos-1

Anthos

⚠️ IMPORTANT NOTICE

1. It's tiny. 984K params. 120K steps on Oxford Flowers 102. 256×256 out.
2. Petals are negotiable. It can draw a rose. Sometimes the rose has too many petals, or stamens in odd places, or a stem that becomes a vine. That's the territory at 0.98M params.
3. Not Stable Diffusion. Class-conditional DiT-Nano/2. No text encoder, no safety filter, no upscaler. The 102 Oxford Flowers classes are the entire vocabulary.

Quick Stats

Stat	Value
Parameters	983,808
Architecture	DiT-Nano/2 (6 blocks, hidden 96, 4 heads, patch 2, SwiGLU)
Training Steps	120,000
Training Time	~18 min on an RTX Pro 6000
Precision	bfloat16
Output	256 × 256
Latents	32 × 32 × 4
Classes	102
Loss	1.843 → 0.880 (flow-matching MSE)
Sampling	Heun, 50 steps, CFG 4.0

What Is This?

Anthos means flower in Greek. That's the whole naming story.

Trained it as a rectified flow rather than diffusion, so the network predicts the velocity field between noise and data. Architecture is just a regular DiT — adaLN-Zero, SwiGLU MLPs, sin-cos pos embed. SD-VAE for the latent side, Heun or Euler at sample time.

Dataset: every image in Oxford Flowers 102 (8,189 across train+val+test), plus a horizontal-flip copy. Encoded once through the VAE and dumped into VRAM as BF16 channels_last. 130 MB on the card. GPULatentLoader shuffles on-GPU and yields batches straight out of VRAM, so a step is just a forward pass.

120K steps in 18 minutes on the RTX Pro 6000. Loss went 1.84 → 0.88. We saved a sample grid every 2K steps to watch the convergence.

Meant as a sanity check on the training loop. It turned out to be a flower generator.

Samples

4×4, class-conditional, step 120K, CFG 4.0, Heun 50. Each tile is a different Oxford Flowers class. Some are clearly the right flower. Some are a guess.

Model Specifications

Parameter	Value
Architecture	DiT
Variant	DiT-Nano/2
Depth	6
Hidden Size	96
Heads	4 (head dim 24)
Patch Size	2
Grid	16 × 16 = 256 tokens
MLP	SwiGLU, ratio 2.0
Norm	LayerNorm, no affine on block norms (adaLN)
Attention	QK-LayerNorm, SDPA
Conditioning	AdaLN-Zero, t + y
Class Dropout	0.1
Class Embed	102 + 1 (null)
Pos Embed	2D sin-cos, frozen
VAE	stabilityai/sd-vae-ft-ema, 8× downsample, 4 ch
VAE Scale	0.18215
Output Ch	4 (no learned sigma)

Training Details

Parameter	Value
Dataset	Oxford Flowers 102 (train+val+test, 8,189 imgs)
Aug	identity + hflip = 16,378 latents
Storage	all in VRAM, channels_last BF16
Batch	256
Grad Accum	1
Optim	AdamW, β=(0.9, 0.95), wd=0, fused
LR	1e-4, 1K-step linear warmup, then constant
Grad Clip	1.0
EMA	0.9999
t Sampler	logit-normal (μ=0, σ=1)
Loss	flow-matching MSE on velocity
CFG Dropout	0.1 (10% labels → null token)
Precision	BF16 autocast, FP32 reductions
Compile	torch.compile(mode="max-autotune")
GPU	RTX Pro 6000, 96 GB, sm_120
Step Rate	~111 it/s
Wall	1078s for 120K steps

Benchmarks

Loss curve, sampled from the log: 1.843 → 1.71 (1K) → 1.31 (10K) → 1.04 (50K) → 0.91 (100K) → 0.880 (120K). Monotone down. Didn't bother with FID or IS.

Usage

from pipeline import AnthosPipeline

pipe = AnthosPipeline(repo_dir=".")

# every class
imgs = pipe(classes="all", seed=0)
imgs[0].save("out.png")  # class 0 is pink primrose, not a rose, by the way

# specific names or ids, comma-separated
imgs = pipe("rose,sunflower,daffodil", n_per_class=2, seed=42)
for i, im in enumerate(imgs):
    im.save(f"flower_{i:02d}.png")

# fiddling
imgs = pipe(73, steps=100, cfg_scale=2.5, sampler="euler", seed=7)

CLI:

python pipeline.py "rose,sunflower,daffodil" --n-per-class 2 --seed 42 --out out.png

For the Gradio demo, see this.

Files

File	What it is
model.safetensors	EMA weights, 3.95 MB
config.json	arch + sampling config
modeling.py	DiT + samplers
pipeline.py	AnthosPipeline
classes.txt	102 names, id⟂tab⟂name
convert_checkpoint.py	final.pt → safetensors
sample_grid.png	4×4 grid, step 120K
requirements.txt	deps

Limitations

• 102 classes, hard-coded. There's no prompt, so no "sunset over a meadow."
• Output is 256×256. Bigger needs an upscaler.
• 0.98M params, which is enough for a rose but absolutely not enough for Stable Diffusion.
• A few classes stayed rough through training. "Barberton daisy" and "mexican petunia" in particular. Oxford Flowers 102 is class-imbalanced and we didn't rebalance it.
• No FID/IS numbers. We looked at the samples.
• Don't use it in a textbook or in production.

Citation

@misc{anthos2026,
  author = {Glint Research},
  title  = {Anthos: a 984K-parameter class-conditional DiT on Oxford Flowers 102},
  year   = {2026},
  publisher = {Hugging Face},
  url    = {https://huggingface.co/Glint-Research/Anthos}
}

---

Built by Glint Research.