NanoDreamer

NanoDreamer is a compact, unconditional image-generation model trained to synthesize 256 × 256 images of liminal interior spaces. It is designed for fully local inference in a web browser with ONNX Runtime Web.

The recommended FP16 model is 47.9 MiB. WebGPU is used when available, with WebAssembly CPU inference as a compatibility fallback. After the model has been downloaded, image generation runs on the user's device without a cloud inference server.

Open the interactive browser demo

Technical overview

NanoDreamer is not a diffusion model. It generates an image in a single model forward pass and does not perform iterative denoising. There is no prompt, caption, text encoder, reference image, mask, or class input.

The exported graph contains the StyleGAN mapping network, truncation, synthesis network, fixed synthesis noise, RGB conversion, and final output scaling. A fixed latent vector and truncation value produce a deterministic result.

The NanoDreamer training, export, and browser deployment pipeline is custom. The underlying StyleGAN2 generator and discriminator components are based on OpenMMLab MMagic.

Model interface

The model has two FP32 inputs:

• latent: shape [1, 512], normally sampled from a standard Gaussian distribution.
• truncation: shape [1], controlling movement toward the mean latent.

It returns one FP32 tensor named image with shape [1, 256, 256, 3]. The output uses NHWC layout, RGB channel order, and values in the range [0, 1].

Available files

The FP16 graph uses reduced-precision internal weights and tensors while preserving FP32 model inputs and output. Actual runtime memory consumption is higher than the download size because the runtime also allocates intermediate tensors and provider-specific buffers.

Training

NanoDreamer was trained at 256 × 256 resolution using StyleGAN2 with adaptive discriminator augmentation. Training also used R1 discriminator regularization, generator path-length regularization, mixed precision, and an exponential moving average of the generator weights.

The training dataset contains 20,000 synthetic JPEG images generated with the Z-Image diffusion model. They depict rooms, corridors, openings, surfaces, lighting, and architecturally ambiguous liminal spaces. Only the images were used for training; captions and source prompts were not provided to NanoDreamer.

The deployed model was exported from the EMA checkpoint at iteration 54,698. With a training batch size of 64, this represents approximately 3.50 million real-image presentations, including repeated and augmented views.

Intended use

NanoDreamer is intended for local generative-art experiments, compact GAN deployment research, latent-space exploration, and ONNX Runtime WebGPU testing. It is not a general-purpose image generator, text-to-image model, or architectural design tool.

Limitations

• The model only represents its narrow synthetic liminal-interior training distribution.
• It cannot follow text instructions or control specific image content.
• Generated geometry may contain impossible perspective, fused surfaces, repeated structures, or inconsistent openings.
• The latent space is not guaranteed to be semantically disentangled.
• Outputs may reproduce biases and artifacts found in the synthetic dataset.
• Browser compatibility and generation speed depend on the device, browser, GPU driver, and ONNX Runtime version.

License

NanoDreamer code and model artifacts are released under the MIT License. Third-party training components and runtimes retain their respective licenses.