---
license: other
license_name: flux-1-dev-non-commercial-license
license_link: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md
task_categories:
- text-to-image
language:
- en
tags:
- flux
- lora
- peft
- personalization
- dreambooth
- llm-as-a-judge
- evaluation
pretty_name: LITHE evaluation images and judge outputs
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path:
    - metadata.csv
    - images/**/*.jpg
---

# LITHE evaluation images and per-image judge outputs

This dataset contains every generated image behind Table 1 (`tab:db30-pareto`) and the multi-seed
appendix tables of the LITHE paper. Each image comes with the raw output of the three LLM judges that
scored it (GPT-5.1, Claude Sonnet 4.6, Gemini 2.5 Flash): the integer score for each rubric dimension
and the free-text `reason` (and, on SC-40, `description`) fields the judge returned. With these files
the paper's judge-based numbers can be audited image by image, and the images can be rescored with a
different judge.

Code, prompts, judge prompt templates and the per-prompt score tables:
https://github.com/JianJiang1995/LITHE (`artifacts/`). Pretrained LITHE adapters:
https://huggingface.co/Fanjue/LITHE.

## Contents

| benchmark | images | methods x runs | prompts per concept |
|---|---:|---|---|
| DreamBooth-30 (`dreambooth30`) | 4,650 | 10 x 3, plus vanilla FLUX x 1 | 5 |
| CC101-20 (`cc101_20`) | 3,100 | 10 x 3, plus vanilla FLUX x 1 | 5 |
| SCIdentity (`scidentity`) | 3,100 | 10 x 3, plus vanilla FLUX x 1 | 5 |
| SC-40 (`sc40`) | 5,720 | 10 x 3, plus vanilla FLUX x 1 | 5 (DyLoRA and LoRA-FA: 3) |
| total | 16,570 | | |

Methods (`method`): `lora`, `lithe`, `dora`, `vera`, `adalora`, `qlora_nf4`, `dylora`, `lora_fa`,
`loraxs`, `tlora`, and `vanilla_flux` (FLUX.1-dev without an adapter). `seed` is the training seed of
the run: 42, 1 and 7 for LoRA and LITHE; 42, 43 and 44 for the other adapters; 42 for vanilla FLUX.
All adapters are rank 16 on FLUX.1-dev.

```
images/<benchmark>/<method>/seed_<seed>/<concept_id>__p<prompt_idx>.jpg
judges/<benchmark>.jsonl
metadata.csv
SHA256SUMS
```

The row set is identical to `artifacts/results/*_per_prompt.csv` in the code repository: one image per
(benchmark, method, seed, concept_id, prompt_idx) row, and every row there has an image here.

### Images

All images are FLUX.1-dev outputs at 1024 x 1024. They were re-encoded to JPEG (quality 92) from the
original output files, writing the pixels only, so no EXIF, XMP, ICC or other metadata is carried
over. The original files were JPEGs as well; `original_sha256` in `metadata.csv` is the SHA256 of the
original file as it was evaluated, and `sha256` is the SHA256 of the file in this dataset. Because of
the re-encoding, pixel values differ slightly from the evaluated files, so a rescore of these JPEGs is
close to but not bit-identical with what the judges saw.

No training or reference images are included. Images are generated outputs only.

* DreamBooth-30 and CC101-20 subjects come from the public DreamBooth and CustomConcept101 datasets;
  their reference images are available from the original authors.
* SCIdentity subjects are synthetic, rare animal-like individuals.
* SC-40 covers five axes (outfit, lighting, pose, painter style, anime style). The people shown in
  SC-40 images (in particular on the outfit and pose axes) are synthetic and were generated by
  FLUX.1-dev; they do not depict real persons.
* The SCIdentity and SC-40 training/reference images are not part of this release; their
  construction is described in the paper appendix.

### `judges/<benchmark>.jsonl`

One JSON object per image:

| field | meaning |
|---|---|
| `benchmark`, `method`, `seed`, `concept_id`, `prompt_idx` | row key, as in the code repository |
| `prompt_set` | prompt list the image was generated from (see `artifacts/results/README.md`) |
| `group` | SC-40 axis, `null` elsewhere |
| `prompt` | the generation prompt, including the trigger token |
| `image` | relative path of the image in this dataset |
| `rubric` | `identity` (SI, PF, VQ, SBD, CN), `sc40` (TF, PF, VQ), or `identity_legacy_si_pf_vq` (see below) |
| `judges` | one entry per judge: `gpt51`, `claude`, `gemini` |

Each judge entry has:

| field | meaning |
|---|---|
| `model` | GPT-5.1, Claude Sonnet 4.6 or Gemini 2.5 Flash |
| `scores` | integer 1-5 score per rubric dimension |
| `description` | SC-40 only: the judge's description of references vs. generated image |
| `reason` | the judge's justification text, verbatim |
| `raw_response` | the complete reply text as returned by the API (Claude and Gemini), including any Markdown code fence; `null` for GPT-5.1 |
| `followup_fields` | present when the first reply omitted these score fields; their values come from the single score-only follow-up request (Claude, 154 images) |

GPT-5.1 was run in JSON output mode and its reply was stored as the parsed JSON fields (`scores`,
`description`, `reason`); the reply string itself was not kept, so `raw_response` is `null` for
GPT-5.1. One Claude reply (`scidentity`, `loraxs`, seed 43, `glass_gill_axolotl_cat`, prompt 4)
contained scores but no reason; its `reason` is the empty string.

Rubric dimensions. Identity benchmarks: `SI` subject identity, `PF` prompt following, `VQ` visual
quality, `SBD` subject-background disentanglement, `CN` class naturalness. SC-40: `TF` target
fidelity, `PF` prompt following, `VQ` visual quality.

Vanilla FLUX is scored by GPT-5.1 only. On DreamBooth-30 and CC101-20 its images were generated from
the `vanilla_sks` prompts and scored with an earlier GPT-5.1 rubric that returned SI, PF and VQ plus
`description` and `reason` (`rubric = identity_legacy_si_pf_vq`); on SCIdentity and SC-40 it uses the
same rubrics as the adapters.

Each judge entry is the call whose scores enter the paper tables, taken from the same result files
as `artifacts/results/*_per_prompt.csv`, and `raw_response` is the reply of that same call. For 706
images (GPT-5.1 44, Claude 642, Gemini 20 judge entries) the evaluation logs also hold earlier or
partial re-runs with the same scores and different text; those are not included.

### `metadata.csv`

One row per image: the row key, `prompt`, `file_name` (image path), `width`, `height`,
`original_sha256`, `sha256`, `rubric`, all per-judge rubric scores and averages (`gpt51_*`,
`claude_*`, `gemini_*`, `llm_avg`) and the embedding metrics (`dreamsim`, `clip_i`, `clip_t`), with
the same values as `artifacts/results/*_per_prompt.csv`. Columns that do not apply to a row are
empty. `artifacts/results/aggregate.py` in the code repository reproduces Table 1 and the appendix
tables from these values.

## Judging protocol

GPT-5.1 was called through the OpenAI chat-completions API at temperature 0 with JSON output mode,
image detail `high`, and `max_completion_tokens` 300 (identity) / 400 (SC-40). Claude Sonnet 4.6 and
Gemini 2.5 Flash were called through an OpenAI-compatible endpoint at temperature 0 with the same
prompt text and image order, without JSON mode; when a reply omitted a score field, the missing
fields were requested once more with a short score-only follow-up on the same images.

The identity rubric shows the judge the subject's first training image (`REFERENCE:`) and the
generated image (`GENERATED:`). The SC-40 rubric shows two reference images of the concept and the
generated image. Per-image LLM-avg is the mean of the three judges' dimension means.

## Rescoring with another judge

The judge prompt templates and a minimal judge script are in `artifacts/judge/` of
https://github.com/JianJiang1995/LITHE. The script talks to any OpenAI-compatible chat-completions
endpoint, configured through `OPENAI_API_KEY` and `OPENAI_BASE_URL`:

```bash
git clone https://github.com/JianJiang1995/LITHE && cd LITHE/artifacts/judge
export OPENAI_API_KEY=...            # and OPENAI_BASE_URL for a non-OpenAI endpoint
# identity benchmarks: one reference image (the subject's first training image)
python judge.py --rubric identity --model <judge-model> \
    --image <dataset>/images/dreambooth30/lithe/seed_42/db_cat__p0.jpg \
    --ref <dreambooth>/cat/00.jpg --prompt "a photo of a <v_sub_001> cat" --class-name cat
# SC-40: two reference images of the concept
python judge.py --rubric sc40 --model <judge-model> --image <image> --ref <ref1> --ref <ref2> \
    --group anime_style --concept candy_glass_cyber_cel --prompt "<prompt from the jsonl>"
```

Take `prompt` from `judges/<benchmark>.jsonl`, the class word and trigger from
`artifacts/prompts/<benchmark>.csv`, and pass `--no-json-mode --detail ''` for endpoints that reject
JSON mode. Reference images are needed for the identity and target-fidelity dimensions and are not
part of this dataset (see Images above). Prompt following and visual quality can be rescored from the
prompt and image alone with a rubric of your own.

## License

The images are outputs of FLUX.1-dev and are released under the FLUX.1 [dev] Non-Commercial License
(https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md). The judge texts are
model outputs released for research use under the same non-commercial terms. DreamBooth and
CustomConcept101 subjects remain subject to the terms of their original datasets.