Flux 2 Klein

Flux 2 Klein — Pro Pack on RunComfy

runcomfy.com · 9B model · 4B model · GitHub

Black Forest Labs' Flux 2 Klein (the distilled, low-latency variant of Flux 2) hosted on the RunComfy Model API — no API key, async REST.

npx skills add agentspace-so/runcomfy-skills --skill flux-2-klein -g

When to pick this model (vs siblings)

Flux 2 Klein's distinct strength is latency-first creative iteration: sub-second feedback enables live art-direction sessions and rapid product visualization that batch-style models can't sustain. Pick it when iteration speed matters more than ceiling resolution.

If the user said "Flux 2 Klein" / "BFL Klein" / "flux klein" explicitly, route here regardless. If they said "Flux 2" generically, ask whether they want Klein (fast) or Pro (max quality) before defaulting.

Prerequisites

1. RunComfy CLI — npm i -g @runcomfy/cli
2. RunComfy account — runcomfy login opens a browser device-code flow.
3. CI / containers — set RUNCOMFY_TOKEN=<token> instead of runcomfy login.

Endpoints + input schema

Two variants, same endpoint shape, same prompt grammar.

blackforestlabs/flux-2-klein/9b/text-to-image

The fidelity-first variant. Use for polish / final output.

blackforestlabs/flux-2-klein/4b/text-to-image

The latency-first variant. Sub-second 4-step inference. Use for live iteration / concepting.

Same field set as 9B. Default steps is effectively 4 — the variant is built for that step count.

Reference images (both variants)

Up to 4 simultaneous reference images are supported on the same endpoint for style transfer / guided composition. The exact field name in the JSON body is documented on the model's API tab — pass it through the CLI verbatim. Reference-image use enables editing-style workflows without a separate /edit endpoint.

How to invoke

Fast concepting (4B, sub-second):

runcomfy run blackforestlabs/flux-2-klein/4b/text-to-image \
  --input '{"prompt": "<user prompt>"}' \
  --output-dir <absolute/path>

Polish / final (9B, ~25 steps):

runcomfy run blackforestlabs/flux-2-klein/9b/text-to-image \
  --input '{
    "prompt": "<user prompt>",
    "steps": 25,
    "width": 1024,
    "height": 1024
  }' \
  --output-dir <absolute/path>

Wide-format poster:

runcomfy run blackforestlabs/flux-2-klein/9b/text-to-image \
  --input '{"prompt": "<user prompt>", "width": 1536, "height": 864}' \
  --output-dir <absolute/path>

The CLI submits, polls every 2s until terminal, then downloads any *.runcomfy.net / *.runcomfy.com URL from the result into --output-dir. Stdout is the result JSON. Stderr is progress.

For pipe-friendly usage:

runcomfy --output json run blackforestlabs/flux-2-klein/4b/text-to-image \
  --input '{"prompt":"..."}' --no-wait | jq -r .request_id

Prompting — what actually works

These are model-specific patterns that empirically improve output quality.

Subject-first declarative grammar. The structure Flux 2 Klein was trained on is "Subject + action + scene + style + lighting + camera + quality". Front-load the subject; trail with directives. Example: "A vibrant hummingbird mid-flight sipping nectar from a bright pink hibiscus, iridescent feathers in morning sun, soft bokeh tropical garden, macro photography, razor-sharp detail, cinematic lighting".

Specificity wins over flowery language. "4k product photo, softbox lighting, reflective table, 35mm, f/2.8" guides predictably. "A really pretty product image" doesn't.

Step-count by phase.

• Concepting: 4–8 steps on the 4B variant — sub-second feedback for live exploration.
• Refinement: 8–15 steps still on 4B, locking in subject + framing.
• Polish: ~25 steps on the 9B variant — texture, microdetail, fine typography.

Multi-reference alignment. When passing reference images, keep their aesthetics aligned. Mixing a watercolor + a photoreal + a 3D render in the same call confuses the editor. Pick one consistent visual register across all refs.

Conditional edits: state what stays, then what changes. "Same composition and lighting as reference, but change the background from beach to mountain studio." This pattern holds composition stable.

For text rendering (Klein has the 8B Qwen3 embedder, decent but not GPT Image 2 territory): add "crisp typography, high-contrast label" and bump steps to ~25 if the text comes out soft. For heavy in-image text or multilingual rendering, route to GPT Image 2 instead.

Anti-patterns:

• Don't conflict adjectives. "minimalist + ornate" cancels.
• Don't exceed ~512 tokens. The model degrades, doesn't truncate gracefully.
• Don't ask for 4K — the model's resolution ceiling is ~2K.
• Don't ask for ultra-wide (>16:9) — the model crops.

Where it shines

Sample prompts (verified to produce strong results)

From the model page (BFL example):

A vibrant hummingbird mid-flight sipping nectar from a bright pink hibiscus
flower, iridescent emerald and sapphire feathers catching the morning sun,
soft bokeh tropical garden background, macro photography, razor-sharp
detail, cinematic lighting

Product-photo pattern:

A matte ceramic mug on a reclaimed-wood table, soft northern window light
from the left, shallow depth of field, 50mm prime, f/2.0, neutral
background, e-commerce ready, 4K product photography

Brand-consistent pair (multi-ref):

Same composition and lighting as the reference image, but the bottle
label is now blue with white sans-serif typography reading "AURA";
keep the bottle silhouette, table, and shadow exactly as in the reference

Limitations

• Resolution ceiling ~2K — for higher native res, route to Seedream 5.
• Aspect ratio cap 16:9 — extreme wide/tall ratios get cropped.