---
license: other
license_name: s-lab-license-1.0
license_link: LICENSE
library_name: pytorch
pipeline_tag: image-to-image
base_model: wusize/Harmon-0_5B
tags:
- harmon
- masked-autoregressive
- unified-multimodal
- component-extraction
---
# Harmon-0.5B complete visual MAR and VAE components

All **684 trained non-language tensor entries** selected from the [official fixed Harmon-0.5B checkpoint](https://huggingface.co/wusize/Harmon-0_5B/tree/481ab555997c222ca2794020ed4ad57e33b2707d): `mar.*`, `vae.*`, `proj_in.*`, and `proj_out.*`. The VAE, shared MAR encoder, masked decoder, learned positions/buffer, diffusion prediction/sampling head and both language-space projections are retained. Original tensor names, dtypes and values are preserved. The Qwen language model is excluded.

Install `requirements.txt` in a dedicated environment. The native timm blocks are tested with `timm==1.0.8`.

```python
from huggingface_hub import snapshot_download
import sys, torch
folder = snapshot_download('gatilin/harmon_0_5b_visual_mar_codec')
sys.path.insert(0, folder)
from component import load_model
visual = load_model(folder)
rgb = torch.zeros(1, 3, 64, 64)  # RGB NCHW in [-1, 1]
with torch.inference_mode():
    z = visual.encode(rgb)  # [B,H/16,W/16,16], native VAE mode * 0.2325
    reconstruction = visual.decode(z)
    tokens = z.flatten(1, 2)
    mask = torch.zeros(tokens.shape[:2])
    features = visual.mar.forward_mae_encoder(
        tokens, mask, visual.mar.fake_latent, image_shape=z.shape[1:3])
```

The native VAE scaling factor is **0.2325**. `decode` preserves caller input while retaining the original arithmetic. MAR's native masks, learned buffer/positions and diffusion head remain directly accessible. `proj_in` maps MAR features to the 896-dimensional language space; `proj_out` maps language hidden states back. Complete language-conditioned generation requires the trained Qwen backbone, original buffer reordering and masked generation schedule; this package is not a standalone text-to-image model.

`validation.json` records **18 passed CPU float32 checks**: all source tensor hashes, strict loading, native scale, square and rectangular VAE encode/decode, finite reconstructions, masked MAR encoder/decoder geometry, both projections, diffusion prediction/variance outputs, the original 100-step sampling head and classifier-free-guidance sampling. Sampling covers two token conditions, not a complete image-quality or VQA benchmark. Re-run with `python validate.py`.

Native files come from the pinned official HF repository, with their URLs and original hashes in `code_provenance.json`. The documented portability changes allocate sampling noise and timestep tensors on the input device instead of hardcoded CUDA. Native sampler equations and probability distributions are unchanged. The excluded full-model file remains a source reference.

**Non-commercial terms:** the official model card and project specify **S-Lab License 1.0**, reproduced in `LICENSE`; retain its conditions and disclaimer. This is not an Apache/MIT weight release. Original component attributions and licenses are retained separately where applicable. Source-range mapping, per-tensor SHA-256 and executed validation are included; the selected safetensors are an extraction, not the original full checkpoint file.