README.md

---
license: other
license_name: research-use-sdxl-openrail-derived
license_link: https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md
pretty_name: Improving CA-LoRA — CA measurements and generated images
tags:
  - lora
  - sdxl
  - cityscapes
  - concept-aware-lora
  - synthetic-images
  - reproducibility
  - research
task_categories:
  - text-to-image
size_categories:
  - 1K<n<10K
---

# Improving CA-LoRA — CA measurements and generated images

The two large evidence sets behind the reproduction-and-extension study of **CA-LoRA
(Concept-Aware LoRA)** on SDXL / Cityscapes:

1. **`measurements/`** — the raw head-granularity concept-attribution tensors for a 13-timestep
   sweep (backs **Table 2** and **Figure 1** of the report).
2. **`generated/`** — every image that was scored for the main results (backs **Table 3** of the
   report).

The adapters that produced the images live in
[`chs35/improving-ca-lora-checkpoints`](https://huggingface.co/chs35/improving-ca-lora-checkpoints).

---

## `measurements/tsweep_sdxl_head/` — 39 files, 191 MB

Head-granularity CA measurements over the timestep grid
`[1, 41, 81, 121, 201, 301, 401, 481, 601, 701, 801, 901, 981]` — 13 timesteps × 3 files:

| file | what it is |
| --- | --- |
| `t<NNNN>.safetensors` | the full measurement at that timestep (256 samples) |
| `t<NNNN>_half0.safetensors` | first split half |
| `t<NNNN>_half1.safetensors` | second split half |

The `half0` / `half1` pair is what the **split-half agreement** column of Table 2 is computed from
(Spearman 0.956 for the primary ews metric); the full files carry the CA magnitudes and direction
statistics that rank the timesteps and drive Figure 1's per-head heatmaps.

Each file carries a `unit_granularity` metadata field (`head`), which the loading code verifies —
head- and module-granularity measurement products cannot be mixed. Measurements were taken at
`batch_size = 1` with `crop="center"`, and seeds are derived so that pairing holds across concept
axes and across timesteps.

---

## `generated/` — 11 runs × 900 PNG + manifest, ~12 GB

The exact 1024×1024 images scored for **CMMD** and **CLIP-Score** in Table 3. One directory per
evaluated arm:

```
generated/<RUN>/
  in_domain/   00000.png … 00499.png   (500 images)
  foggy/       00500.png … 00599.png   (100)
  night-time/  00600.png … 00699.png   (100)
  rainy/       00700.png … 00799.png   (100)
  snowy/       00800.png … 00899.png   (100)
  manifest.json
```

Filenames are a **global index 00000–00899 across the whole run**, not per-condition counters — the
condition directory a file sits in is implied by its index range, as laid out above. 900 images per
run, 9,900 in total.

`manifest.json` records, for every single image: its `index`, the full `prompt` actually used, the
sampled `class_names`, the `condition`, the `seed` and `batch_seed_index`, and the relative `path`.
It also records the generation protocol (guidance scale 5.0, 25 steps, `EulerDiscreteScheduler`,
1024², fp32, seed 0) and which adapter checkpoint was loaded. Per-image scores can therefore be
recomputed exactly from these files.

### Runs, and the report rows they back

| directory | selection criterion | axis | Table 3 row |
| --- | --- | --- | --- |
| `A_t81_style` | t = 81 (the original paper's timestep) | style | "t = 81 (paper) / style" |
| `A_t81_viewpoint` | t = 81 (the original paper's timestep) | viewpoint | "t = 81 (paper) / viewpoint" |
| `B_top3_style` | multi-t, top-3 timesteps [41, 1, 81] | style | "multi-t [41, 1, 81] / style" |
| `B_top3_viewpoint` | multi-t, top-3 timesteps [41, 1, 81] | viewpoint | "multi-t [41, 1, 81] / viewpoint" |
| `C_t41_style` | t = 41 (this study's top-ranked timestep) | style | "t = 41 / style" |
| `C_t41_viewpoint` | t = 41 (this study's top-ranked timestep) | viewpoint | "t = 41 / viewpoint" |
| `D_t1_style` | t = 1 | style | "t = 1 / style" |
| `D_t1_viewpoint` | t = 1 | viewpoint | "t = 1 / viewpoint" |
| `E_random` | **control**: random 2% of heads, no CA | — | "random 2% (no CA)" |
| `E_all` | **control**: all attention projections, no CA selection | — | "full attention (no CA)" |
| `control_base` | **control**: base SDXL, no adapter at all | — | "0% control (base SDXL)" |

The four weather conditions are what the per-condition CLIP-Score columns and the weather-mean
column of Table 3 (and the gap analysis of Table 4) are computed over; `in_domain` is the CMMD set.

---

## ⚠️ Not included: the Cityscapes CMMD reference set

CMMD in Table 3 is measured against **500 centre crops of the Cityscapes `val` split**. Those
reference images are **not distributed here, or anywhere** — the Cityscapes license forbids
redistribution of the imagery. No real Cityscapes image appears in this repository.

The reference set is rebuildable from your own Cityscapes download: take the **first 500 images of
the `val` split in index order**, at **1024×1024**, **centre-cropped** and un-mirrored — the same
framing the training data was read with, so that CMMD measures a gap in content rather than a
difference in cropping. The exporter is `scripts/evaluate.py` (`export_reference`) in the code
repository, and the relevant config knobs (`eval.reference_split`, `eval.reference_num_images`,
`eval.seed`) are in the released `config.yaml`. Obtain Cityscapes from
<https://www.cityscapes-dataset.com/>. Every other number in the report is reproducible from the
artifacts here without it.

---

## License and intended use

**Research use.** The PNGs are **synthetic images generated by SDXL-derived adapters fine-tuned on
Cityscapes** — they are model output, not photographs, and they inherit the
[OpenRAIL++ use restrictions](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md)
of the SDXL base model, which apply to this data as they do to the model that produced it. Because
the adapters were fine-tuned on Cityscapes, users should also respect the
[Cityscapes terms](https://www.cityscapes-dataset.com/license/). Again: **no real Cityscapes images
are included** in this repository.

These images exist to make the report's metrics auditable and re-scorable. They are not a
general-purpose street-scene dataset, and they carry the domain and quality limitations discussed in
the report's Limitations section.

## Verifying integrity

The code repository ships `results/sha256_measurements.txt` and `results/sha256_generated_images.csv`
(`path,bytes,sha256`), keyed by the same repo-relative paths used here.