O que flux faz?
The FLUX image tool skill — pick the variant + license, instruct in scenes, lock references + edit instead of
re-rolling, evaluate deliberately, and label before publishing. The agent prompts (and can call the API where
connected); the human judges every image; WoopSocial publishes. (Ships with tools/integrations/flux.md.)
The POV: control is the product — and the license is the trap
FLUX's 2026 edge isn't just quality; it's control: multi-reference consistency (up to ~8–10 images in one call — a campaign-consistent character or product with no fine-tuning), in-context editing ("change the jacket, keep everything else"), hex-code brand colors as parameters, 32k-token scene prompts, and typography clean enough that ad headlines are production-viable. Two top-1% edges most users miss. (1) The license split is the trap: [dev] outputs are commercially usable, but self-hosting [dev] to serve commercial work (clients, a product) needs a paid BFL tier — the agency tier includes just 3 clients before per-client fees — and the dev license requires content filters or manual review, which BFL says it may verify at random. Apache-2.0 freedom lives in [schnell] and [klein] 4B; the hosted APIs are the easiest commercial path (license + signed provenance handled). (2) Edit, don't re-roll: a 95%-right image is one Kontext-style instruction from done — re-rolling throws away the 95%. And the craft shift: FLUX reads natural-language scene briefs, not tag soup — describe subject, light, mood, camera; put exact in-image text in quotes and verify every character.
Read these first
- image-prompt — the model-agnostic prompt craft + router above all image tools.
- brand-profile + design-and-templates — the system (colors → hex codes, type, logo rules).
The framework: PIXEL
(Depth: references/the-pixel-framework.md.)
- P — Pick the variant + license: hosted API = easiest commercial; [klein] 4B/[schnell] = Apache-2.0 self-host; [dev] = outputs OK / services need a paid tier + required filters; [max] for web-grounded; verify at bfl.ai.
- I — Instruct in scenes, not tags: natural-language multi-part briefs (32k context); hex codes for brand-exact color; exact text in quotes; prompt upsampling helps [dev].
- X — eXact references + edits: multi-reference locks characters/products (never an unpermitted likeness); in-context edits fix the 5% instead of re-rolling; chain small edits, watch drift; verify rendered text.
- E — Evaluate + iterate: the human judges every render; one variable per iteration; keep the seed when composition lands; side-by-side drift review on sets.
- L — License + label: rights confirmed for the variant actually used; AI-disclosure where required; never strip the signed provenance metadata; then WoopSocial publishes.
The reality (verify-quarterly)
FLUX.2 (Nov 2025, current flagship): 32B rectified-flow + Mistral-3 VLM, multi-reference (~8–10 images), 4MP,
clean small-size typography (the "text soup" era largely fixed), hex-color parameters, photorealism reducing
"the AI look"; [klein] (Jan 2026) = sub-second on consumer GPUs (~13GB VRAM for the 4B), 4B Apache 2.0
(the 9B klein is non-commercial); [max] adds real-time web-grounded generation; FLUX.1 Kontext = the
in-context editing line (a selectable partner model in Photoshop Beta's Generative Fill). Licensing (bfl.ai, attributed): dev outputs commercial-OK; self-hosted commercial services need paid
tiers (developer ~10K img/mo single-domain, not client work; agency ~100K/mo, 3 clients included); filters or
manual review required on [dev] with random verification stated; API applies cryptographically-signed
provenance metadata; non-removable CSAM/NCII filters on API. Runs via Playground → APIs (BFL/fal/Replicate/
Together/Cloudflare) → self-host (ComfyUI; FLUX.2 [dev] quantized on an RTX 4090). Attribute all;
verify-quarterly. Full detail: references/flux-2026-reality.md. The variant table, scene-prompt pattern,
consistency + edit workflows, and two worked examples: references/prompt-patterns-and-templates.md.
Honest scope (never violate)
- The agent prompts, plans references/edits, and calls the API where connected (exact human steps otherwise); the human judges every image (no fabricated "that looks great"; side-by-side set reviews); WoopSocial publishes finished exports — it does not generate or edit media.
- License spine: verify rights for the variant actually used; [dev]-as-a-service needs the paid tier + the
filter/review obligation; Apache-2.0 = [schnell]/[klein] 4B; high-stakes → counsel (not legal advice).
Likeness/IP/provenance: no unpermitted real-person likeness (real or AI lookalike), no cloned trade dress
or copyrighted characters, never strip provenance, AI-disclosure where required (EU AI Act; C2PA). Data
honesty: never invent stats (data viz → infographic-and-data-viz). Never fabricate benchmarks/prices/
capabilities. (Full scope:
references/scope-and-connections.md.)
Distinct from its siblings (route correctly)
flux (this) = the FLUX-specific lane (photorealism + typography + multi-reference + editing + open-weight control) · image-prompt = the model-agnostic router/craft (read first) · ideogram = graphic-design/ text-layout lane · Midjourney = distinct aesthetic lane (external — no skill in this library) · nano-banana = its documented lane ( strengths shift per release — test, don't trust leaderboards) · ai-image-editing = the edit router FLUX editing will serve · canva = the design workflow output drops into · ai-video / veo-3 / kling = video (BFL's video model is announced — verify before claiming).
Where this connects
Reads first: image-prompt + brand-profile + design-and-templates. Feeds: canva (layout/type over
imagery), thumbnail-design, quote-cards-and-text-graphics (backgrounds), carousel-writer,
pinterest-pin-design, before-after-and-transformation (honest visuals only). Publishes via: export →
scheduling-and-queue → WoopSocial. Tool file: tools/integrations/flux.md. Measure with: native +
analytics-and-reporting — never fabricated.
Definition of done
A FLUX workflow that is on-brand and on-license: the variant chosen with its actual rights verified (hosted API for easy commercial; Apache-2.0 [schnell]/[klein] 4B for free self-host; [dev] outputs-vs-service line respected with the filter/review obligation met; paid tiers for client/product self-hosting), prompts written as natural-language scene briefs with hex-exact brand color and quoted in-image text, consistency achieved by multi-reference (original/consented characters only) and near-misses fixed by in-context edits rather than re-rolls, every render human-judged with one-variable iteration and side-by-side set drift review, all rendered text character-verified, and the shipped image labeled honestly (AI-disclosure where required; provenance metadata intact) then published via WoopSocial; no unpermitted likeness, no cloned trade dress, no stripped provenance, no invented stats, no fabricated benchmarks/capabilities; and correctly distinguished from image-prompt, the sibling image tools, ai-image-editing, and canva.