# xj + MJHQ-30K — 4-model comparison samples

Inference outputs of 4 stage-2 text-to-image models on the combined
`xj_mjhq30k_prompts` set (30,954 prompts), one image per prompt per model:

| Model | CFG | Tar file |
|---|---|---|
| flux       | 6 | `t2i-ddt-en28d1152hd72-dn2d2048hd128-flux-vpred-t4-norepa-v0_ep-0000020_xj_mjhq30k_prompts_steps50_cfg6.tar` |
| flux2      | 7 | `t2i-ddt-en28d1152hd72-dn2d2048hd128-flux2-vpred-t4-norepa-v0_ep-0000020_xj_mjhq30k_prompts_steps50_cfg7.tar` |
| e2e-vavae  | 6 | `t2i-ddt-en28d1152hd72-dn2d2048hd128-e2e-vavae-vpred-t4-norepa-v0_ep-0000020_xj_mjhq30k_prompts_steps50_cfg6.tar` |
| langpe-l   | 7 | `t2i-ddt-n28_2d1152_2048hd72_128-rae-langpe-vit-l-vpred-t4-norepa-v0_ep-0000020_xj_mjhq30k_prompts_steps50_cfg7.tar` |

The 4 image folders are packed as **`.tar`** (no gzip — PNG is already
compressed) to keep file count low for fast cloning. Each tar extracts into a
folder named after itself, containing `00000.png … 30953.png` and matching
`.txt` prompt sidecars.

## Extract

```bash
for f in *.tar; do tar -xf "$f"; done
```

## Contents after extract

```
<this-repo>/
├── <model1_dir>/00000.png, 00000.txt, ...
├── <model2_dir>/...
├── <model3_dir>/...
├── <model4_dir>/...
├── selected/<NNNNN>/{flux,flux2,e2e-vavae,langpe-l}.png + prompt.txt + meta.json
└── viewer/   # browser-based 4-model comparison viewer (works once tars are extracted)
```

## Viewer

After extracting all tars, browse the 4-model comparison side-by-side:

```bash
python viewer/viewer.py            # default :8765, or --port N
# open http://localhost:<port>/test/xj_mjhq30k_inference_outputs/viewer/index.html
```

(The viewer expects to be served from a directory layout where the run folders
and `viewer/` are siblings, as in this repo.)

`selected/<NNNNN>/` holds curated picks — 54 prompts selected via the viewer's
Save button, each with the 4 models' images plus the prompt text and metadata.