Revera — Neural Image Upscaling

01 · Live demo

Upload an image.
See what's left behind the pixels.

Supports JPG, PNG, WebP up to 20 MB. Processing runs locally in your browser via a WebAssembly build of the model. Nothing leaves your device — except quality metrics, which we aggregate to improve the model.

revera-lab / session / anonymous

wasm-runtime · v4.2.1

↑

Drop image here

or select manually

Browse

JPG PNG WEBP up to 20 MB max 4096px on longest side

Initializing... 0% · 0.0s

Done. Compare for yourself.

ORIGINAL

REVERA × 2

200% · pixel-level

Input

— × —

Output

— × —

LPIPS gain

+0.000

Time

— s

02 · Pipeline

Three stages. Nothing more.

The Revera-LD architecture is built on the classic latent diffusion paradigm with modifications: a Variational Autoencoder compresses the image into an 8-dimensional latent space, a 50-step DPM++ 2M sampler refines the latent, and a VAE decoder restores the pixels. All components are jointly trained on 14M paired and 220M unpaired images.

STAGE 01

Latent Encoding

VAE encoder compresses the input image into an 8-channel latent representation with a downsampling factor of 8. This allows working in a space 64× smaller than pixel space, reducing inference time by 11× compared to pixel-space diffusion.

ENCODER · VAE-HQ / 86M
LATENT DIM · 8 × H/8 × W/8
COMPRESSION · 1 : 48

STAGE 02

Diffusion Sampling

A 1.2B-parameter U-Net performs 50 steps of DPM++ 2M sampling with classifier-free guidance. At each step, the model predicts noise and refines the latent. Input resolution is preserved via text-conditioned skip-connections.

U-NET · 1.21B params
SAMPLER · DPM++ 2M · 50 steps
CFG · 7.5
LATENT NOISE · σ_max = 14.6

STAGE 03

Decoder + Refine

VAE decoder restores pixels from the refined latent. An additional perceptual-refinement block corrects high-frequency details via a pre-trained VGG-19+LPIPS discriminator. Final pass is adaptive noise reduction.

DECODER · VAE-HQ / 86M
REFINE · VGG-19 + LPIPS
POST · adaptive NR · unsharp

03 · Model card

Technical specifications

Transparency is part of our methodology. All hyperparameters, datasets, and metrics are published openly. The model card is updated with every release.

revera-ld-v4.2.bin STABLE

architecturelatent-diffusion

params total1,412,882,304

encoderVAE-HQ · 86M

u-net1.21B

decoderVAE-HQ · 86M

samplerDPM++ 2M

steps (default)50

cfg scale7.5

latent channels8

max output16,384 × 16,384

training data14M pairs

unpaired aug220M images

fp16 inferenceyes

licenseRAIL-M · commercial

Trained on 14 million paired images.

The dataset is compiled from publicly licensed stock photos (CC0, CC-BY), museum archives (Rijksmuseum, Met Open Access), and synthetic pairs generated via a realistic downscale pipeline. Each pair underwent manual curatorial review.

An additional 220M unpaired images were used for the pretrain stage in a self-supervised (MAE) regime. Fine-tuning was performed on 8× H100 for 14 days with a batch size of 768.

F.01

Face-aware refinement

A separate attention block for facial regions. Prevents hallucinations around eyes, mouth, and ears.

F.02

Texture consistency loss

Multi-scale Gram-matrix loss ensures consistent preservation of fabric, skin, and wood textures across the entire image.

F.03

Color fidelity lock

The original color tone is preserved with ΔE < 1.4 in Lab space. No 'enhanced' colors like our competitors.

F.04

Anti-hallucination guard

A CLIP-based artifact detector rejects generations with >0.7 hallucination confidence and retries the step with reduced CFG.

05 · Benchmarks

Comparison with SOTA models.

Metrics calculated on the DIV8K-test dataset (480 images) and DIV2K-validation (100 images). Lower is better for LPIPS and FID, higher is better for CLIPIQA and user study.

Model	Params	LPIPS ↓	FID ↓	CLIPIQA ↑	User Study ↑	Inference (A100)
Revera-LD v4.2 this paper · 2024	1.41B	0.082	12.4	0.687	87%	3.2s
Real-ESRGAN x4 Wang et al. · 2021	16.7M	0.142	24.7	0.612	64%	0.4s
SwinIR Liang et al. · 2021	11.8M	0.156	28.3	0.598	58%	0.6s
Stable Diffusion x4 Rombach et al. · 2022	1.50B	0.094	14.8	0.642	71%	4.1s
Topaz Gigapixel v7 proprietary · 2024	—	0.098	16.2	0.651	73%	—
Upscayl (Real-ESRGAN) open-source · 2024	16.7M	0.142	24.7	0.612	64%	0.4s
HAT-L Chen et al. · 2022	40.1M	0.118	19.6	0.628	69%	1.2s
Bicubic (baseline) —	—	0.241	42.1	0.514	22%	0.01s

06 · API

Integrates in five lines.

Python SDK, REST API, ready-made wrappers for Node.js, Go, Rust. SLA 99.95%, edge inference in 12 regions, automatic fallback.

Python · simple synchronous upload

Install the package via pip, pass a file path or PIL object, and get the result. The enhance_faces and restore_textures parameters enable the respective model blocks.

Asynchronous mode, batch processing, webhooks, and streaming output are available on the Studio tier and above.

upscale.py · python 3.11

# pip install revera-sdk
from revera import Client
from PIL import Image

client = Client(api_key="rv_live_8f2a...c41d")

# sync mode · x4 · face-aware
result = client.upscale(
    image="portrait.jpg",
    scale=4,
    model="revera-ld-v4.2",
    enhance_faces=True,
    restore_textures=True,
    output_format="png",
)

print(f"Output: {result.width}×{result.height}")
print(f"LPIPS gain: +{result.metrics.lpips_gain:.3f}")
print(f"Inference: {result.metrics.inference_ms}ms")

result.save("portrait_4k.png")

07 · Pricing

Pay for the result, not the subscription.

All tiers include access to Revera-LD v4.2, parallel processing, REST API, and the web interface. Prices are in euros, VAT included.

Free

0€/mo

for evaluation

5 images per month
Max 2048px output
Scale ×2, ×4
Web interface
API access
Batch processing
No watermark

popular

Pro

14€/mo

for professionals

500 images per month
Max 8192px output
Scale ×2, ×4, ×8
Web + REST API
Batch processing (up to 20)
No watermark
Priority queue

Studio

49€/mo

for teams

Unlimited images
Max 16,384px output
Scale ×2, ×4, ×8, ×16
API + Webhooks + Streaming
Unlimited batch
Commercial license
SLA 99.95% · 24/7 support

08 · FAQ

Frequently asked questions.

Didn't find an answer? Email us at support@revera.ai — we reply within 6 hours.

Does it really work in the browser without sending data to a server?

Yes. The WASM build of the model is 280 MB and downloads once, cached in IndexedDB. All computations run on your device via WebGPU (where supported) or fallback to WebGL2. Quality metrics (LPIPS, inference time) are sent to the server only if you consent in the cookie banner.

Why is the result sometimes almost identical to the original?

If the original image already contains the maximum amount of real information (e.g., shot on a full-frame camera with no noise), the model has nothing to reconstruct — and Revera-LD correctly returns nearly the same image, only slightly refining high-frequency contours. This is a feature of the task itself: you cannot restore what wasn't there. For heavily compressed or low-resolution images, the difference will be immediately noticeable.

Can I use it commercially?

Yes, on the Pro and Studio tiers — full RAIL-M commercial license. On the Free tier, the result can only be used for personal purposes and with a watermark. All rights to the original image remain yours; we do not store uploaded files for more than 24 hours, even for debugging.

What does 'no hallucinations' mean? Isn't that an exaggeration?

It is impossible to completely eliminate hallucinations in any generative model. We use a CLIP-based artifact detector at each sampler step and retry the step with a reduced CFG if the hallucination confidence is >0.7. This reduces the frequency of noticeable artifacts to 2.4% in our internal tests (vs 11.7% for Real-ESRGAN, 8.2% for Stable Diffusion x4).

Does the model support anime, drawings, or 3D renders?

Yes, but with caveats. Revera-LD v4.2 is primarily trained on photographic content. For anime/illustrations, we recommend enabling the domain=illustration parameter — it switches to a fine-tuned branch. A separate model, revera-3d-v1.0 (beta), is available for 3D renders.

Can I train the model on my own data?

Yes, the Enterprise tier (from 399 €/mo) includes fine-tuning on your data (minimum 800 paired images). Training takes 18–36 hours on our A100s. The finished model is deployed in an isolated instance.

Why is it so expensive compared to Upscayl (which is free)?

Upscayl uses Real-ESRGAN — a 2021 model with 16M parameters. Revera-LD v4.2 has 1.4B parameters, latent diffusion, a 50-step sampler. The quality difference is visible to the naked eye on complex textures (skin, hair, fabric). Compare LPIPS and user study scores in the Benchmarks section — this isn't marketing, it's measurable metrics.

We restore the detail
your camera failed
to capture.

Upload an image.
See what's left behind the pixels.