# Flux2 Klein Prompting

> Expert prompt engineering for FLUX.2 [klein] image generation and editing model. Use this skill whenever the user wants to create prompts for FLUX.2 [klein], generate images, edit photos with the klein model, work with multi-reference image editing, or needs templates for T2I/I2I tasks. Trigger for any mention of: FLUX.2, flux klein, BFL API, image editing prompts, text-to-image prompts for FLUX, product mockups, poster generation, UI mockups, sticker packs, character design, seamless textures, or any request to write/improve/translate prompts for FLUX-family models. Also trigger when user asks about guidance_scale, inference steps, distilled vs base modes, or multi-reference workflows. Do NOT use for training a FLUX.2 Klein / Qwen-Edit LoRA (use flux2-lora-training), nor for reconstructing a prompt FROM an existing source image (use forensic-prompt-compiler); this skill is for authoring generation/edit prompts only.

- Skill: `anastasiyaw/flux2-klein-prompting` (Agent Skill)
- Install (CLI): `npx skillmds@latest add anastasiyaw/flux2-klein-prompting`
- Raw SKILL.md: https://api.skillmd.com/api/skills/anastasiyaw/flux2-klein-prompting/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Design & Media
- Author: AnastasiyaW (https://skillmd.com/u/anastasiyaw)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/anastasiyaw/flux2-klein-prompting

---


# FLUX.2 [klein] — Prompt Engineering Guide

## Core principle: prose, not tags

BFL guidance favors a concrete natural-language description over an ambiguous
bag of tags. State who/what is in the image, where, style, materials/light/camera
and — for editing — the properties to preserve. Prompt shape is a starting
point, not a quality guarantee; assess the result against the requested edit.

---

## Model variants quick reference

| Axis | Options | Notes |
|---|---|---|
| Size | 4B / 9B | 9B better for complex instructions; 4B fastest |
| Mode | Distilled / Base | Use the exact model-card or serving API settings; step/CFG values are implementation-specific |
| License | 4B Apache-2.0 / 9B Non-Commercial | Check before commercial use |
| Task | T2I / Edit (I2I) / Multi-reference | The current BFL image-editing guide documents up to 8 references via API (10 in playground); re-check the selected endpoint schema before use |

**9B uses Qwen3 8B text embedder** → solid multilingual support (Russian works natively).

---

## Prompt structure

### T2I (text-to-image)
1. **Subject** — who/what, key attributes
2. **Scene/context** — where, time of day, surroundings
3. **Composition** — framing, angle, background
4. **Light/materials** — source, softness, reflections, texture
5. **Style/genre** — photorealism, illustration, catalog, poster, UI
6. **Text in image** (if needed) — exact string in quotes + position/font

### Edit (I2I, no mask)
1. **Base anchor** — "This exact image but…"
2. **What to change** — object / background / text / color / material
3. **What to preserve** — face, lighting, style, perspective, brand elements
4. **Multi-reference** — reference by "image 2 / image 3", keep prompt concise

---

## Key rules

**Text in image** → always in straight quotes, specify position. Without this: garbled glyphs.
```
Заголовок: "ТОЧНЫЙ ТЕКСТ". Шрифт жирный гротеск, ровный кернинг. Других надписей не добавлять.
```

**Negatives → positives** → don't say "don't change X", say "preserve X"
```
❌ "не меняй освещение"
✅ "Сохрани освещение, перспективу и лицо"
```

**Multi-reference** → simplify text, use explicit indexing
```
"Возьми персонажа из image 2 и помести рядом с объектом из image 1."
```

Choose distilled or base only after the task’s quality/latency requirement and
the exact model card are known; neither is an automatic “preview” or “final”
mode.

---

## Ready-to-use templates (Russian)

### Photorealistic object
```
Фотореалистичная предметная фотография [объект] на [фон], ракурс [сверху/на уровне глаз/крупный план], мягкий студийный свет, реалистичные материалы и фактуры, аккуратные тени, высокая детализация. Без логотипов и водяных знаков.
```

### Product mockup / e-commerce
```
Каталожный product shot: [товар] в центре кадра, фон [описание], чистая композиция, цвет товара строго [HEX или словом], реалистичные отражения, нейтральный стиль, как для e-commerce.
```

### Logo / icon
```
Минималистичная иконка: [смысл/символ], плоский дизайн, 2–3 цвета, чёткий силуэт, без мелких деталей. Без текста.
```

### Character design
```
Персонаж: [кто], внешний вид: [рост/пропорции/одежда], выражение лица [эмоция], стиль [аниме/3D/иллюстрация], палитра [цвета], фон простой. Сохранить узнаваемость: [признак 1], [признак 2].
```

### Sticker pack (6 emotions)
```
Набор стикеров одного персонажа (6 штук): радость, злость, удивление, смущение, сон, восторг. Единый стиль, толстый контур, яркая палитра, прозрачный фон, без текста.
```

### Poster with readable text
```
Постер [стиль]. Вверху крупный заголовок: "ТОЧНЫЙ ТЕКСТ". Шрифт: жирный гротеск, ровный кернинг, читаемо. Ниже подзаголовок: "Ещё одна строка". Остальные надписи не добавлять.
```

### UI mockup (mobile)
```
UI‑мокап мобильного приложения [тематика]. 3 экрана в одной сетке. Читаемые заголовки на русском в кавычках: "[Экран 1]", "[Экран 2]", "[Экран 3]". Минималистичная дизайн‑система, много воздуха, аккуратная типографика, без лишнего декоративного шума.
```

### Seamless texture
```
Бесшовная текстура (seamless): [материал], равномерное освещение, без объектов, без текста, высокая детализация, натуральные вариации, без резких пятен.
```

### Edit: replace / recolor / swap background
```
Это то же изображение, но: [что изменить]. Сохрани: [освещение / перспектива / лицо / композиция / стиль]. Сделай результат фотореалистичным и согласованным по теням и отражениям.
```

### Edit: add text to sign/label
```
Это то же изображение, но добавь на [табличку/вывеску] точный текст: "[ТЕКСТ]". Сохрани стиль таблички, фон и освещение. Текст должен быть читаемым. Больше текста не добавляй.
```

### Edit: multi-reference character swap
```
Это то же изображение, но возьми персонажа из image 2 и помести рядом с персонажем из image 1. Сохрани реалистичные тени, масштаб и общую атмосферу сцены.
```

---

## API parameters

### Recommended defaults (from official BFL HF Spaces)
| Mode | Settings | Use for |
|---|---|---|
| Distilled / Base | Read the exact checkpoint or API reference | Choose from the requested fidelity, latency and cost constraints |

### BFL API constraints (klein endpoints)
- Endpoint fields, size limits, reference-image count and result-URL lifetime
  are API-version facts. Read the current endpoint reference before constructing
  a request; do not infer them from a local pipeline or another FLUX endpoint.

### Available API fields (klein)
Do not maintain a local fixed field list. Copy the request schema from the
current BFL API reference for the selected endpoint; multi-reference capacity,
endpoint names and optional fields change independently of this skill.

---

## Python: BFL API (async polling)

Illustrative request shape only: copy the current selected-endpoint example from
BFL before use. This snippet does not establish that its endpoint or fields are
currently supported.

```python
import os, time, requests

BFL_API_KEY = os.environ["BFL_API_KEY"]

# 1. Create task
create = requests.post(
    "https://api.bfl.ai/v1/flux-2-klein-4b",
    headers={"x-key": BFL_API_KEY, "Content-Type": "application/json"},
    json={
        "prompt": 'Это то же изображение, но добавь на вывеску текст "ОТКРЫТО". '
                  "Сохрани фон, освещение и перспективу. Больше текста не добавляй.",
        "input_image": "https://example.com/your-image.png",
        "seed": 42,
        "output_format": "png",
    },
    timeout=60,
)
task = create.json()

# 2. Poll until ready
while True:
    time.sleep(0.5)
    data = requests.get(task["polling_url"], headers={"x-key": BFL_API_KEY}).json()
    if data["status"] == "Ready":
        print("Done:", data["result"]["sample"])  # signed URL
        break
    if data["status"] in ("Error", "Failed"):
        raise RuntimeError(data)
```

## Python: local Diffusers

Illustrative pipeline shape only. The shown step/guidance values are not a
recommended default; verify the installed pipeline and selected model card.

```python
import torch
from PIL import Image
from diffusers import Flux2KleinPipeline

pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-4B", torch_dtype=torch.bfloat16
).to("cuda")

# T2I
image = pipe(
    prompt='Постер "КОФЕ". Жирный гротеск, ровный кернинг, без других надписей.',
    height=1024, width=1024,
    guidance_scale=1.0, num_inference_steps=4,
    # A CPU generator makes comparisons more stable across GPU runs.  It is
    # not a cross-version or cross-hardware reproducibility guarantee.
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]

# Edit (I2I)
base = Image.open("input.png").convert("RGB").resize((1024, 1024))
edited = pipe(
    prompt="Это то же изображение, но замени фон на светлую кухню. "
           "Сохрани объект, освещение и перспективу.",
    image=[base],
    height=1024, width=1024,
    guidance_scale=1.0, num_inference_steps=4,
).images[0]
```

---

## Troubleshooting

| Problem | Cause | Fix |
|---|---|---|
| Garbled text / glyphs | Text not quoted explicitly | Exact string in quotes; say "no other text" |
| Blurry / artifacts | Selected configuration may be too aggressive for the task | Compare a small, pinned parameter change on the accepted visual criteria; do not assume a universal step count or mode |
| Style drift in edit | Missing preservation clause | Always add "Сохрани: свет/лицо/композицию" |
| Multi-reference "soup" | Overloaded prompt + conflicting refs | Simplify text; use "image 1 / image 2" indexing |
| Wrong resolution | Input not multiple of 16 | Pre-resize input to ×16, ≤4MP |

---

## Iteration workflow

1. Write scene in prose (one paragraph)
2. Create a controlled candidate using parameters supported by the selected
   checkpoint or API, and record them with the output.
3. Fix the seed when the runner supports it and record the model revision,
   Diffusers/PyTorch version, hardware, dtype and scheduler. A seed makes a
   controlled comparison within that recorded stack; it is not a cross-version
   or cross-hardware reproducibility guarantee. Pick 1–2 directions against
   the task's actual criteria.
4. Refine prompt: add specifics, quote text, remove filler adjectives.
5. Edit iterations: one change per step, state what must be preserved, and keep
   only variants that pass the requested fidelity checks.

---

## Quality metrics (for A/B testing)
- **CLIPScore** — prompt↔image alignment (reference-free)
- **FID** — realism vs real image distribution
- **Human rating** — separate scales for: (a) prompt adherence, (b) quality/realism, (c) text readability, (d) preservation of unchanged parts in edit

---

## Official sources
- Model page: https://bfl.ai/models/flux-2-klein
- Prompting guide: https://docs.bfl.ml/guides/prompting_guide_flux2_klein
- Image editing guide: https://docs.bfl.ai/flux_2/flux2_image_editing
- API reference 4B: https://docs.bfl.ai/api-reference/models/generate-or-edit-an-image-with-flux2-%5Bklein-4b%5D
- HF model card 4B: https://huggingface.co/black-forest-labs/FLUX.2-klein-4B
- HF model card 9B: https://huggingface.co/black-forest-labs/FLUX.2-klein-9B