---
license: other
tags:
  - image-generation
  - class-conditional
  - imagenet
  - fd-loss
  - mmd
library_name: pytorch
---

# pMF-H · FD-SIM · ImageNet-256 · σ0.7 @ 4000 steps

Class-conditional ImageNet-256 generator (pMF-H lineage), fine-tuned with **FD-Loss**
(`mmdx` — Nyström-approximated MMD) on a 10-encoder gated-PID judge panel. This is the
**σ0.7 arm** (RBF bandwidth = 0.7 × median heuristic), continued from the converged
10-encoder gated-PID run (step 3499) to **step 4000**.

## This file

`model.pth` — **model weights only** (the non-EMA `model` state_dict).

| field | value |
|---|---|
| tensors | 654 (all `float32`) |
| size | 3.82 GB |
| opt-step | 4000 (`current_step`; `step`=3999, 0-indexed) |
| samples seen | 20,480,000 |
| source ckpt | `phase42-sigma07-from3499/step_0003999.pth` (model slice; EMA + optimizer + queues dropped) |
| resolution | 256×256, class-conditional (1000 ImageNet classes) |

## Loading

The file is a torch checkpoint with the weights under the `model` key — drop-in for the
MMDLoss/FD-Loss training repo's `load_from`:

```yaml
# config
load_from: /path/to/model.pth
```

Or directly:

```python
import torch
ckpt = torch.load("model.pth", map_location="cpu", weights_only=False)
state_dict = ckpt["model"]            # 654 tensors
model.load_state_dict(state_dict, strict=False)
print(ckpt["current_step"], ckpt["samples_seen"])  # 4000, 20480000
```

## Notes

- These are the **raw (non-EMA)** model weights. The EMA shadow weights and the optimizer
  state were intentionally excluded to keep this a slim inference artifact (the full 26.7 GB
  training checkpoint is resumable but not published here).
- Generation in the source project uses 1-step sampling at cfg 7.0.