---
language:
- en
license: other
license_name: flux-output-conditions
license_link: LICENSE.md
task_categories:
- text-to-image
tags:
- synthetic
- characters
- greenscreen
- webdataset
- flux
- animation
size_categories:
- 1K<n<10K
pretty_name: Vivarium Greenscreen Characters — FLUX 9B
configs:
- config_name: default
  data_files:
  - split: train
    path: data/*.tar
---

# Vivarium warm-painted greenscreen characters

**500 fictional people ×10 natural facial expressions =5,000 portrait images**, each832×1,248 pixels. The selected aesthetic is the **ghibli-warm-painted** direction: restrained facial proportions, delicate drawn lines, warm painted shading and a solid chroma-green backdrop. Generation uses the **undistilled9B base**, the **rank64 character LoRA**, **50 steps**, real **CFG4**, and **BF16**, with no quantization or distilled checkpoint.

**Production snapshot:7/500 complete people;70/5,000 images uploaded.** Pilot and production validation are recorded in the linked metadata. Counts describe the snapshot; incremental shard commits can be newer.

The [base model](https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B) and [character LoRA with examples and instructions](https://huggingface.co/laion/FLUX.2-klein-base-9B-vivarium-loras) are linked separately. The compatible [background dataset](https://huggingface.co/datasets/laion/Anim-E-Ghibli-Backgrounds-FLUX-9B) contains wide warm-painted environments for these foreground characters.

## People, ages and intended distribution

| Age group | Intended ages | Male | Female | People | Images |
|---|---|---:|---:|---:|---:|
| `baby` | 9–24 months sampled | 22 | 22 | 44 | 440 |
| `preschool` | 3–5 years | 22 | 22 | 44 | 440 |
| `younger_schoolchild` | 6–8 years | 22 | 22 | 44 | 440 |
| `older_schoolchild` | 9–12 years | 24 | 24 | 48 | 480 |
| `younger_teen` | 13–15 years | 24 | 24 | 48 | 480 |
| `older_teen` | 16–18 years | 24 | 24 | 48 | 480 |
| `young_adult` | 19–29 years | 24 | 24 | 48 | 480 |
| `middle_adult` | 30–49 years | 22 | 22 | 44 | 440 |
| `older_adult` | 50–64 years | 22 | 22 | 44 | 440 |
| `younger_senior` | 65–79 years | 22 | 22 | 44 | 440 |
| `older_senior` | 80–95 years | 22 | 22 | 44 | 440 |

There are250 male and250 female prompt labels. Older schoolchildren, younger teens, older teens and young adults receive24 people per gender; the other age groups receive22. Each age/gender cell is exactly50% white European-ancestry in the prompt; its remaining people rotate across seven other groups. Overall allocation is:

| Intended prompt label | People | Images |
|---|---:|---:|
| `white_caucasian` | 250 | 2500 |
| `african_american` | 36 | 360 |
| `arab` | 36 | 360 |
| `latino` | 36 | 360 |
| `indian` | 36 | 360 |
| `japanese` | 36 | 360 |
| `southeast_asian` | 35 | 350 |
| `indigenous_american` | 35 | 350 |

These are **intended prompt attributes**, not ethnicity or gender inferred from rendered appearance, biological categories, or ground-truth measurements. Skin tones, hair, clothing and builds vary; no cultural costume is required to indicate a group. The distribution is deliberately selected coverage rather than population prevalence. Infants are sampled at9,12,15,18,21 or24 months and safely seated with the full clothed body visible. Other ages use a standing head-to-thigh composition. Teen prompts explicitly request natural adolescent proportions; late teens use near-adult faces with modest eyes and a defined jaw, while schoolchildren retain age-appropriate proportions without oversized doll-like eyes.

The full [class list](CLASSES.md), [taxonomy](metadata/taxonomy.json), [exact allocation](metadata/allocation.json), [500-person CSV](metadata/persons.csv), and [full prompt plan](metadata/person-plan.json) expose every attribute. No real people's photographs or external audio references were used.

## Stable person IDs and expression variants

`personality_id` and `character_id` are identical stable keys such as`person-0253`. A person has one outfit, identified by`outfit_id=outfit-01`, and ten sample IDs such as`person-0253-neutral` and`person-0253-angry`. The neutral/slightly friendly image is the **default**. This release does not generate outfit changes; later outfits can retain the person ID and add new outfit IDs.

| Expression ID | Intended mild facial cue | Images |
|---|---|---:|
| `neutral` | a relaxed neutral expression with a slight friendly smile | 500 |
| `joyful` | natural joy: an open smiling mouth, raised cheeks and bright happy eyes | 500 |
| `grateful` | warm gratitude: a gentle closed-mouth appreciative smile and softly relaxed eyelids | 500 |
| `playful` | gentle playfulness: a small lopsided mischievous grin and one eyebrow slightly raised | 500 |
| `angry` | mild controlled anger: eyebrows drawn inward and lowered, lips firmly pressed together | 500 |
| `afraid` | mild apprehension: raised inner eyebrows, slightly widened eyes and gently parted lips | 500 |
| `sad` | quiet mild sadness: raised inner eyebrows, softly sad eyes and downturned mouth corners | 500 |
| `annoyed` | mild annoyance: partly lowered eyelids, one brow lowered and a faint flat line of the lips | 500 |
| `pain` | brief mild discomfort: a slight wince, gently narrowed eyes and compressed lips, with no visible injury | 500 |
| `disgust` | mild distaste: a slight nose wrinkle, subtly raised upper lip and asymmetric downturned mouth | 500 |

All expressions request mild natural facial changes, with clearly specified mouth, brow and eyelid cues, rather than exaggerated caricatures. Joyful uses a natural open smile; grateful uses a softer appreciative smile. Pain means mild discomfort with no visible injury. Playful means an innocent lopsided smile. Fully clothed age-appropriate everyday outfits are used for all ages, including children and adolescents. Outfit styles range across casual layers, sportswear, formal clothing, practical workwear and rock-inspired denim; clothing avoids green and teal to support later chroma-keying.

The neutral image is generated first. **Each of the nine other images uses that same neutral image as its visual reference**, resized to416×624 using PIL BICUBIC, with the same seeded noise across the entire person group. Edits request the same identity, hair, skin, outfit, pose and framing and modify the face. The reference PNG's SHA256 and the reference sample ID are recorded. Reference conditioning encourages consistency; it is not a pixel-locking identity guarantee. Small pose, facial-detail or clothing changes can occur. The visual review records what was inspected rather than claiming all appearances were automatically verified.

An initial pilot preserved identity but made facial expressions too similar. It was rejected before public ingestion. Subsequent pilots use focused face-edit prompts instead of repeating the full person description and remove the generic expression-suppression negative clause from edits. The accepted quality report, when available, records the actual inspected samples.

## Prompt template and example

The shared style is:

```text
Studio Ghibli-inspired grounded coming-of-age animation, delicate expressive hand-drawn lines, rich warm watercolor-like painted shading, gentle luminous colors, restrained almond-shaped eyes and proportionate faces.
```

For a default portrait,`tools/plan_character_production.py` builds a complete age/heritage/hair/outfit description and applies the shared character template. For edits it uses a focused instruction, for example:

```text
vvar_char, edit the reference portrait. Make this exact same fictional 17-year-old person show mild controlled anger: eyebrows drawn inward and lowered, lips firmly pressed together. Change the mouth, eyebrows and eyelids so this mild natural emotion is clearly visible. Preserve exact identity, face shape, skin tone, hair, clothing, pose and framing. Studio Ghibli-inspired grounded coming-of-age animation, delicate expressive hand-drawn lines, rich warm watercolor-like painted shading, gentle luminous colors, restrained almond-shaped eyes and proportionate faces. Fully clothed, age-appropriate. Perfectly uniform solid bright chroma green background. No scenery, no text, no logo, no watermark.
```

Negative prompts exclude extra people, changed clothing, different hairstyles, text, logos and scenery. School-age and older prompts also discourage oversized doll eyes and chibi proportions. The full positive and negative prompts are in every sample's JSON and the actual positive prompt also has its own TXT member.

## WebDataset layout and loading

Every sample contains three adjacent members:

```text
person-0253-neutral.png
person-0253-neutral.json
person-0253-neutral.txt
```

Images are written directly into TARs. There are no individually uploaded PNGs. A production shard normally holds **two complete person groups,20 images**, with10-image first-worker/pilot shards and smaller last-worker shards. Keeping all expressions of a person together makes grouping convenient. [Shard SHA256/size/count manifest](metadata/shards.json), [inference records](metadata/inference-records.jsonl), [progress](metadata/progress.json), and [sample index](metadata/sample-index.json) describe the release snapshot. Use`personality_id` for splitting: random splitting by image can leak the same fictional identity across train and evaluation.

Install the lightweight reader dependencies and stream with Hugging Face Datasets:

```python
from datasets import load_dataset
ds = load_dataset("laion/Vivarium-Greenscreen-Characters-FLUX-9B",
                  split="train", streaming=True)
sample = next(iter(ds))
image, metadata, prompt = sample["png"], sample["json"], sample["txt"]
print(metadata["personality_id"], metadata["emotion"], image.size)
```

The card config selects only`data/*.tar` so index files are not mistaken for training examples. An alternate standard-library TAR/HTTP reader avoids extracting image files:

```bash
python tools/read_webdataset.py pilot-person-0253.tar \
  --repo laion/Vivarium-Greenscreen-Characters-FLUX-9B --limit 10
```

## Inference and reproduction

Install`requirements.txt`, obtain access to the linked base model, and authenticate locally using your own Hugging Face account. The generator uses pinned model and LoRA revisions listed in [provenance](metadata/provenance.json); no access token is included in the dataset or code. The default example generates one late-teen person and all ten expressions directly into one TAR:

```bash
python tools/infer_character_tar.py --person person-0253 \
  --plan metadata/person-plan.json --output person-0253-emotions.tar
```

`--base-path` can point to an already downloaded base checkpoint.`--no-compile` disables compilation when needed. The default inference helper fuses the LoRA and QKV projections, compiles repeated transformer blocks, batches the positive/negative CFG branches and caches text embeddings. These optimizations preserve50-step real CFG4 sampling; there is no denoising approximation cache. The pipeline is BF16 and expects a CUDA GPU with sufficient memory. Published seeds and revisions enable matching the generation recipe, but bitwise reproduction across GPU/runtime versions is not promised.

## Timing, validation and provenance

Production uses eight independent A100-SXM4-80GB workers, batch size one per GPU. Neutral generation and reference edits have different workloads; their measured means are reported separately in [performance](metadata/performance.json). Reference edits include reference VAE encoding in the timed call. Positive prompt preencoding, initial compilation, loading, PNG encoding, TAR writes and uploads are excluded from those inference timings. The pilot shared one GPU with background generation and used CPU offload, so its timings are excluded from dedicated-worker statistics.

Every closed shard is verified against remote SHA256 and byte size after upload. The final integrity report fully decodes all PNGs, checks adjacent PNG/JSON/TXT grouping, unique planned IDs, prompt/seed equality, image checksums and reference linkage, complete ten-emotion person groups, and the exact planned distributions. [Integrity report](metadata/integrity-validation.json) and [visual review](metadata/quality-review.json) are published when their respective checks are complete.

The character adapter was trained on the previously selected Vivarium greenscreen data and chosen at training step1,500, rank/alpha64, strength1.0. This release's500 people are newly planned synthetic identities. Their count distribution is independent of the original Anim-E procedural prompt row frequencies. Generation terms are documented in [LICENSE.md](LICENSE.md) with the full [model license](MODEL_LICENSE.md).