Mindscend Image 0.3
A small class-conditional image generator trained from scratch by Mindscend. It makes 64×64 pictures of ten subjects and runs on your device: in the browser in the Mindscend AI Lab (ONNX Runtime Web, WebGPU) and on iPhone, iPad and Mac (Core ML).
0.3 has the same architecture as 0.2, trained on 2.6× more images for 2.5× more steps. It is better at every guidance strength we measured (table below).
| Architecture | Rectified-flow U-Net (levels ×1, ×2, ×2, ×2; attention at the 2nd level), class-conditional with classifier-free guidance |
| Parameters | 38.3M |
| Resolution | 64×64 RGB |
| Classes | flower, bird, insect, boat, train, mountain, building, water, bridge, sculpture (index 10 = no class) |
| Training data | 156,567 images picked from PD12M by caption, up to 20,000 per class (bridge 7,548, sculpture 6,245, mountain 9,369 had fewer); photos only, no people |
| Training | 200,000 steps, batch 64, AdamW 2e-4, EMA weights, one RTX 3080 Ti |
| Format | ONNX fp32, opset 17 (browser); Core ML FP16 (Apple) |
Quality
FID and Inception score, 2,000 generated samples (200 per class), 30 Euler steps, against the 156,567 training images. 0.2 is measured the same way on the same reference set, so the columns compare directly. One run each (no error bars).
| guidance | 0.2 FID | 0.3 FID | 0.2 IS | 0.3 IS |
|---|---|---|---|---|
| 1.0 | 34.01 | 26.53 | 7.78 | 8.17 |
| 1.5 | 29.21 | 23.90 | 8.19 | 8.42 |
| 2.0 | 27.46 | 24.37 | 8.06 | 8.61 |
| 3.0 | 30.61 | 28.27 | 8.65 | 8.67 |
(Lower FID is better, higher IS is better.) The 0.2 model card's numbers used only 200 samples, so they read much higher; don't compare across the two cards.
It is a research model: subjects are recognisable, details are not. Many PD12M photos are old, so many results look like old photographs.
Use
mindscend-image-0.3.onnx is one velocity call: inputs x (b×3×64×64, float32), t (b, float32, 1 = noise,
0 = image), y (b, int64 class); output v. Sampling with N Euler steps and guidance g, conditional and
unconditional (y = 10) in one batch of 2:
x = randn(3, 64, 64)
for i = N .. 1: vc, vu = model([x, x], [i/N, i/N], [class, 10]); x -= (vu + g * (vc - vu)) / N
image = clamp(x * 0.5 + 0.5, 0, 1)
coreml/MindscendImage.mlpackage is the same model as a Core ML ML program (FP16, fixed batch of 2: the class and
the no-class row), used by the Mindscend AI Lab app on iPhone, iPad and Mac.
config.json holds the class names and the Lab's defaults (20 steps, guidance 1.5, the best FID above).
License
Weights: Apache-2.0. Training images: PD12M, public domain / CC0 (metadata CDLA-Permissive-2.0).