flux-2-klein-prompting
Prompt Black Forest Labs' FLUX.2 [klein] — the subject/environment/style/technical hierarchy it processes in order, guidance-scale and step choices, negative prompts aimed at real failure modes, and the seed-locked one-variable-at-a-time loop its distilled checkpoints make fast enough to run. Which dials exist depends on base versus distilled. Use whenever the model id contains flux-2/klein or flux-2-klein (fal-ai/flux-2/klein/4b, fal-ai/flux-2/klein/9b, their /base, /edit and /lora variants, fal-ai/flux-2/klein/realtime, black-forest-labs/flux-2-klein-4b, black-forest-labs/flux-2-klein-9b) on generate_image, edit_image or a TextToImage node. Not for FLUX.2 pro, flex, max or dev.
- License 许可
- AGPL-3.0
- 最近更新
- 2026-11-15
- Stars
- ★ 635
- Forks
- 91
01 前言 Front Matter
- name
- flux-2-klein-prompting
- description
- Prompt Black Forest Labs' FLUX.2 [klein] — the subject/environment/style/technical hierarchy it processes in order, guidance-scale and step choices, negative prompts aimed at real failure modes, and the seed-locked one-variable-at-a-time loop its distilled checkpoints make fast enough to run. Which dials exist depends on base versus distilled. Use whenever the model id contains flux-2/klein or flux-2-klein (fal-ai/flux-2/klein/4b, fal-ai/flux-2/klein/9b, their /base, /edit and /lora variants, fal-ai/flux-2/klein/realtime, black-forest-labs/flux-2-klein-4b, black-forest-labs/flux-2-klein-9b) on generate_image, edit_image or a TextToImage node. Not for FLUX.2 pro, flex, max or dev.
02 资源文件 1
- 参考文档references/submission-contract.md3.4 KB
03 SKILL.md 正文
FLUX.2 [klein] → structure first, then iterate fast
klein is the small, fast member of the FLUX.2 family, at 4B or 9B. It ships
two kinds of checkpoint. The distilled ones (every route without /base) run
in 4 to 8 steps, which fal's guide puts under a second. The /base ones run
28 steps by default and carry the guidance, negative-prompt and acceleration
dials. The speed is the point of the distilled routes. When a generation takes
30 seconds, prompt refinement stalls; here you can change one clause and look
again, so the working method below is worth more than any single prompt.
Reach it with find_model for text_to_image or image_to_image, then
generate_image / edit_image. generate_image forwards only the prompt, a
negative prompt, the size and a quality string, and edit_image only the
images, the prompt, a negative prompt, a target size and a strength. Every other
field below exists only on the route's own node: find it with search_nodes,
read its fields with get_node_info, and run it with run_node. Each run is a
paid generation, so confirm the scope with the user before a long loop.
Prompt hierarchy
klein reads a prompt in this order, and content words (nouns) move the output more than modifiers do. Write it in the same order:
- Subject — concrete beats abstract. "A woman in her mid-30s with shoulder-length auburn hair" over "a person".
- Environment — lighting, atmosphere, spatial relationships. "First light filtering through morning mist" gives rendering cues that "morning scene" does not.
- Style and mood — direction that guides without overwhelming: "documentary photography", "studio lighting", "cinematic".
- Technical — composition ("rule of thirds"), depth of field ("shallow focus on the foreground"), palette ("warm earth tones").
Professional headshot of a male architect in his 40s, salt-and-pepper beard, black-rimmed glasses, charcoal blazer. Modern office background with architectural models visible but softly blurred. Natural window light from the left creating gentle shadows. Corporate photography style, sharp focus on the eyes, neutral grey backdrop.
Parameters
| Parameter | Where it exists | What it trades | Where to sit |
|---|---|---|---|
guidance_scale |
fal /base routes (default 5); Replicate -base calls it guidance (default 4) |
Prompt adherence against creative freedom | 2–4 for artistic interpretation, 5–8 for product photography and technical illustration |
num_inference_steps |
fal only: 4–50 on /base (default 28), 4–8 on distilled (default 4), 1–8 on realtime |
Quality against time | Low while developing a prompt, high for print-ready output |
acceleration |
fal /base routes |
Detail against throughput | regular for production, high for maximum speed |
negative_prompt |
fal /base routes |
See below | Failure modes only |
seed |
every route | Nothing — it removes noise from the comparison | Pin it while testing prompt variations |
On a distilled route the prompt and the seed are the whole instrument, and a
negative prompt sent through generate_image is dropped.
Aspect ratio follows the deployment: landscape for presentations and web, portrait for social and mobile, square for profiles and balanced compositions.
Technique
- Emphasis is natural language, not weights. fal's guide says there is no weight syntax. "Prominently featuring", "with particular attention to" and "especially detailed" are how you raise an element's priority.
- Negative prompts target failure modes, not quality (
/baseonly). For portraits: "distorted features, unnatural proportions, extra limbs". For landscapes: "oversaturated colours, artificial lighting, lens distortion". A generic quality list buys nothing. - Text renders well when specified. Give the copy, the font class, the
colour, the placement and the capitalisation: "a white coffee mug with the
text 'GOOD MORNING' in bold sans-serif black letters, centred on the mug".
Copy that belongs to the design, in other words — a title or a super on a
still headed for a timeline goes on as a text clip afterwards
(
caption-titles), not into the render. - Multi-reference conditioning is the family's edge. "The subject from the first image wearing the jacket from the second image, photographed in the environment from the third image."
The five mistakes
- Overloaded prompts. fal's guide says that past roughly 100 words the model starts dropping clauses. Every word should be doing something.
- Vague style references. "Make it look good" is not a direction. "Shot on Hasselblad medium format" or "impressionist painting technique" is.
- Missing composition. Prompts describe the subject thoroughly and never say where it sits. Add "centred composition", "rule of thirds with the subject on the left vertical", or "bird's eye view".
- Conflicting instructions. "Photorealistic portrait" and "watercolour painting style" in one prompt commits to neither.
- Neglected lighting. Lighting is what makes an image dimensional. Always name it: golden hour, studio softbox, rim lighting creating a silhouette.
The loop klein is for
Start with subject, environment and style. Generate. Pin the seed. Change one element — add lighting detail, refine the subject, move the composition — and generate again. One variable at a time is what tells you which clause drove which change; a distilled route's turnaround is what makes that affordable.
Keep the prompts that worked. A library of formulas per use case, with the guidance and step values that suited each, is the output of this loop as much as the images are.
Adapted from fal's FLUX.2 [klein] prompt guide: https://fal.ai/learn/devs/flux-2-klein-prompt-guide
04 同仓库更多 Skills
humanizer
Rewrite AI-sounding text so it reads like the writer without changing what it says. Use when editing or reviewing prose for AI tells: not-X-but-Y contrasts, one-line closers, staged openers, forced triads, dashes everywhere, inflated claims, sales language, stock AI words, bold labels, or filler.
nodetool
Use the NodeTool MCP server to build, validate, run, and debug visual AI workflows, and to work with assets, collections, and jobs. Use when the user mentions NodeTool or asks for a workflow graph, a node, or a media generation run on their NodeTool server.
wan-2-6-prompting
Prompt Alibaba's Wan 2.6 video — text-to-video with a global style line plus timing-bracketed shots, and image-to-video that describes only what changes over time. Covers the shot-mode switch per provider, the prompt budget, negative prompts, motion intensity wording, audio input and prompt expansion. Use whenever the model id names a Wan 2.6 video route (alibaba/wan-2.6/text-to-video, alibaba/wan-2.6/image-to-video, wan-video/wan-2.6-t2v, wan-video/wan-2.6-i2v, wan-video/wan2.6-i2v-flash) on generate_video, animate_image or a TextToVideo node. Not for Wan 2.5 or 2.7, or the Wan 2.6 image routes.
video-workflow
Build a reusable NodeTool video workflow for batches or repeated inputs. For a single finished piece, use a storyboard skill.
video-clone
Analyze a reference video and recreate its shot structure with a new product or cast in NodeTool.
video-audio-continuity
Keep sound continuous across a multi-scene piece cut from generated video — why one clip per scene hard-cuts the audio at every boundary, when to write all the scenes into a single generation instead, the sound-design vocabulary that survives a provider's copyright filter, and how to check the length, the aspect ratio and the mix of what came back. Use when a piece has more than one scene and the video model writes its own audio (Seedance 2, Veo 3, MiniMax H3, Kling 2.6 and later), or when a cut's sound drops out at the shot changes.