Hemera-Nano

The grid above is deliberately cherry-picked after full-resolution visual review. Exact prompts, negative prompts, seeds, CFG, NFE, hashes, and selection notes are available in assets/showcase/showcase.jsonl.

Hemera-Nano is a 29,756,564-parameter, 256×256 text-to-image denoiser trained from scratch on Photonyx under a micro-budget. Frozen pretrained representation models provide image latents and text embeddings; only the denoiser was trained.

This release is a completed research artifact, but it does not pass every pre-registered success criterion. The automated confirmation composite was positive at 95% confidence, but the candidate did not win in every seed. Blinded human preference ratings and official GenEval detector scoring also remain pending.

Model details

• Backbone: standard flat DiT with global self-attention and adaLN-Zero
• Conditioning: pooled frozen CLIP text representation
• Feed-forward network: dense SwiGLU
• Objective: rectified-flow velocity prediction with logit-normal timesteps
• Trainable parameters: 29,756,564
• Forward compute: 2.489 GFLOPs per sample
• Frozen VAE: dc-ai/dc-ae-lite-f32c32-diffusers
• Frozen text encoder: openai/clip-vit-base-patch32
• Native resolution: 256×256
• Training dataset: Photonyx revision 35978bff3d6f38f4e432c25506168aaf8bd7e26f
• Audited examples: 427,260
• Training examples seen: 736,104,192
• Mean training throughput: 5,234 images/s on an RTX 5090
• Peak training VRAM: 11.47 GiB
• Released checkpoint: best-validation EMA at step 320,000
• Released checkpoint SHA-256: 7746f789384ba6349b2798d8959583e9e66aa06cb2e5231c9602d401ada8111d
• Terminal checkpoint: step 958,469
• Terminal checkpoint SHA-256: 58d1996b6183844230a23102a0808e761cdb017c05490ded7de00b594ffdb4ea
• Exported model.safetensors SHA-256: 3a033dd0ade8178003adaa43270711b382200ec6dffa6f4e565a6c9d4763c6ee
• Locked inference: CFG 5.0, NFE 15

The best-validation checkpoint is the release default. The later terminal checkpoint is retained for analysis and reproducibility, not presented as the best model.

Usage

Install the small Hemera inference package directly from its source repository. The denoiser, VAE, and text encoder are downloaded and cached automatically on first use.

pip install "git+https://github.com/pauhidalgoo/diffusion-architectures.git@v1.0.0"

from hemera import HemeraPipeline

pipeline = HemeraPipeline.from_pretrained(
    "pauhidalgoo/hemera-nano",
    device="auto",
    dtype="bfloat16",
)

images = pipeline(
    [
        "a crystal alpine lake reflecting snow capped mountains at sunrise",
        "blue hydrangea blossoms in a white vase, soft window light",
    ],
    negative_prompt="blurry, malformed, distorted, text, watermark, logo",
    num_inference_steps=15,
    guidance_scale=5.0,
    seed=1234,
)
images[0].save("alpine-lake.png")
images[1].save("hydrangeas.png")

The pipeline also supports prompt batches, per-prompt negative prompts, torch.Generator objects, deterministic seeds, latent output, device selection, and dtype selection.

Training data

The pinned Photonyx audit accepted 427,260 rows:

Split	Rows
Train	418,792
Validation	4,278
Test	4,190

Source	Rows
DiffusionDB	231,217
Safe Commons	196,043

Accepted license labels were CC0, CC0-1.0, and public domain. The immutable manifest hash is 7d4fa4f570640747d957876a52b307bfd3aab460477bd74099c65a55d32692cf.

Locked test evaluation

Inference settings were selected from 15 CFG/NFE candidates on validation before the test split was generated once. The realized test set contains every available test row (4,190), despite the 5,000 requested-sample ceiling.

Metric	Value
Photonyx test CMMD	0.55408
Canonical FID	21.83977
Precision	0.72840
Recall	0.20621
HPSv2	0.19745
CLIP alignment	0.25158
SigLIP alignment	0.12078

These values characterize this release under the frozen Hemera protocol. They should not be compared directly with numbers computed using different datasets, feature encoders, image sizes, sample counts, or estimators.

Per-source slices

Source	Samples	CMMD	Precision	Recall	HPSv2	CLIP
DiffusionDB	2,335	0.50640	0.78630	0.29807	0.20252	0.25005
Safe Commons	1,855	1.00148	0.66361	0.09704	0.19108	0.25351

The large source gap is a material limitation and likely reflects differences in captions, content, and image distributions.

Confirmation and success status

The automated confirmation compared two matched dense recipes across seeds 101, 202, and 303 with 256 paired prompts and 2,000 cluster-bootstrap resamples.

• Relative composite improvement: 0.00745
• 95% confidence interval: [0.00210, 0.01222]
• Candidate wins by seed: positive, positive, negative
• Pre-registered confirmation gate: failed

The checkpoint reload and automated final suite passed. The overall pre-registered success claim remains false because:

1. the confirmation candidate did not beat its baseline in every seed;
2. blinded human preference ratings are not complete;
3. official GenEval detector scoring is not complete.

This release therefore reports a strong micro-budget artifact and useful negative/positive results, not a fully confirmed state-of-the-art claim.

Memorization diagnostics

The locked test generations were compared with a deterministic, source-stratified subset of 50,000 training images. Similarity is diagnostic only; a high embedding similarity is not proof of copying.

Encoder	Mean nearest cosine	99th percentile	Maximum
CLIP ViT-L/14@336	0.82467	0.93448	0.95997
DINOv2-small	0.63442	0.86550	0.92881

The highest matches require blinded visual inspection before drawing any memorization conclusion.

Additional protocols

• Human-study candidate images: 200 across 10 balanced categories. Matched baseline generation, three-rater judgments, confidence intervals, and inter-rater agreement remain pending.
• Safety/bias audit images: 400 from 100 prompts. This suite documents behavior and does not establish safety.
• Official GenEval-format images: 2,212 (553 prompts × 4). Official detector scoring was not run because its pinned PyTorch 1.12/CUDA 11.3 environment is incompatible with RTX 5090 execution. The generations and evaluator revision are preserved for later isolated scoring.

Cost

• Exploratory wall-clock estimate: €9.94
• Decision-sprint wall-clock estimate: €0.74
• Final training ledger: €20.25
• Complete final-instance wall-clock estimate, including evaluation and packaging: €23.05
• Total recorded project wall-clock estimate: €33.73

The original €10 exploratory target was met. The complete final instance exceeded the original €20 target by about €3.05 but remained within the later explicitly authorized contingency. These are estimates derived from recorded hourly prices; stopped periods and provider billing details can make them differ from the actual invoice.

Intended use

Hemera-Nano is intended for research and education around micro-budget text-to-image training. It is not intended for safety-critical, identity-sensitive, medical, legal, or factual visual tasks.

Limitations

Photonyx mixes real, artwork, and synthetic sources and may contain caption, selection, aesthetic, demographic, and cultural biases. The dataset audit retained only the two large sources present in the pinned release; smaller sources described in the dataset README were not present in the audited revision.

A small 256×256 model struggles with people, faces, hands, text, counting, precise spatial relations, uncommon concepts, and multi-entity composition. The deterministic uncurated sample grid shows strong performance on simple animals, food, landscapes, architecture, and artistic media, alongside clear failures on people and relational prompts.

The frozen CLIP encoder has a 77-token limit, and pooled conditioning discards token-level spatial detail. Photonyx captions have a median length of 314 characters and a 95th percentile of 776 characters, so substantial prompt truncation is expected. This likely limits compositional fidelity.

Outputs may reproduce stereotypes or unsafe associations present in the data and frozen representation models. Automated safety prompts do not prove that the model is safe.

Licensing

Hemera's original code and released denoiser weights are Apache-2.0. Photonyx is CC-BY as declared by its owner and is attributed by name and repository link throughout this card. The released Safetensors file contains only Hemera's 29.76M-parameter denoiser; it does not redistribute CLIP or DC-AE weights. Those frozen dependencies are downloaded separately from their upstream Hub repositories and remain subject to their respective terms. The component-level record and release rationale are in LICENSE_AUDIT.md.

Reproducibility

The verified release bundle contains the exact YAML configuration, immutable dataset and prompt manifests, dependency lock, run manifest, source-tree hash, checkpoint hashes, telemetry, validation sweep, locked inference record, generated test pairs, metric features, milestone canaries, failure records, cost ledgers, source archive, all checkpoints, and exported Safetensors.