--- title: "Train Your Brand Aesthetic — Krea 2 Open-Source: lock your look, ship the campaign" tier: hero format: brand-campaign theme: aesthetic-lora persona: creator, marketer, brand, fashion, real-estate, travel, educator duration: ~30s campaign render · ~6–8 min if you train a brand LoRA budget_usd: ~$0.25 for an 8-shot campaign · ~$0.05/look thereafter · ~$22 one-time to train a brand LoRA caps: ["krea-2-os", "krea-2-lora-training", "krea-2-lora", "krea-2-large", "bg-remove", "kling-v3-turbo-i2v", "ltx-q-i2v", "sonilo-t2m", "music", "ffmpeg-concat", "ffmpeg-export"] verbs: ["create_media", "submit_lora_train", "get_lora_train", "apply_lora", "attach_lora_to_project"] skills: ["krea-2-open-source", "lora-training"] showcases: - title: "ATELIER NORD · Krea 2 Open-Source brand campaign" href: "/chapters/train-your-brand-aesthetic-example.html" reliability: 4.3 # krea-2-os campaign is fully proven E2E (8/8 shots, ~3s each) +0.5; bg-remove + i2v finishing proven; the optional LoRA brand-lock cap is registered + meters jobs but full weight-retrieval through the SDK hits a known 202-polling follow-up — the recipe is wired across MCP/CLI/webapp and works the moment that lands --- # Train Your Brand Aesthetic — agent playbook ⭐ flagship > Paste this whole file (with the BRIEF filled in) into Claude cowork, chat, or Code. **This is the Krea 2 Open-Source showcase.** Krea 2 is the most aesthetic open-source image model — and the first open base on the network you can *train*. Lock a brand's signature look once, then generate an unlimited, on-brand campaign that stays coherent shot-to-shot. Fashion-led, but the exact same recipe ships a **real-estate**, **travel**, or **premium-education** campaign — just swap the BRIEF. ## What you'll get A cohesive, ready-to-publish brand campaign built on one aesthetic: - **8 hero campaign images** in a single locked look (`krea-2-os`) — hero portrait, full look, detail/texture, lifestyle, accessory still, atmospheric, back-look, beauty - **A clean e-comm packshot** — `bg-remove` on a product/detail still → transparent cutout - **A 5-second motion teaser** — the hero still animated via `kling-v3-turbo-i2v` (or `ltx-q-i2v`) - **A licensed soundtrack** — `sonilo-t2m` (commercial-clean) or `music` - **A platform-cut reel** — `ffmpeg-concat` the stills + teaser → `ffmpeg-export` to 9:16 / 1:1 / 16:9 - **(Optional) a brand LoRA** — train your signature look on `krea-2` once (`submit_lora_train base_model:"krea-2"`), then every future look is on-brand by construction via `krea-2-lora` **Time**: ~30s for the 8-shot campaign (Krea 2 Turbo renders in ~3s/shot). **Cost**: ~$0.25 for the campaign, ~$0.05/look after, ~$22 one-time if you train a brand LoRA. **Your attention**: one BRIEF + one approval gate. ## Why Krea 2 Open-Source | | What it gives you | |---|---| | **Most aesthetic open base** | Out-of-the-box taste — lighting, palette, composition, film grain — beats every other open image model. The deliverable *looks* art-directed before you touch it. | | **Open weights = trainable** | Unlike the hosted `krea-2` caps, the open `krea-2-os` (RAW) checkpoint is malleable. It's the 2nd trainable base on the network after flux — and the one to pick when *look* matters more than a photoreal identity. | | **One palette, infinite shots** | A text-to-image model has no cross-shot memory. A trained brand LoRA does: the campaign stays coherent across 8, 80, or 800 looks. | ## Tell the agent about the campaign ```yaml brand_name: # e.g. "ATELIER NORD" one_liner: # e.g. "Scandinavian minimalist fashion house" # The single most important field — your signature look, one sentence. # This is what makes 8 shots feel like one campaign. aesthetic: # e.g. "ivory, sand, warm stone & oat palette; soft diffused north-window # light; sculptural wool & linen; quiet luxury; fine film grain; # Vogue-cover restraint; medium-format." # 8 shots (defaults below work for fashion; swap subjects for any vertical) shots: - hero portrait close-up - full-length look - macro detail / texture - lifestyle / motion - accessory still life - atmospheric wide / negative space - back-look at a window - beauty close-up aspect: "4:5" # 4:5 editorial, 9:16 social, 16:9 hero banner train_brand_lora: false # set true to lock the look as a reusable krea-2 LoRA finish: # which finishing passes to run packshot: true # bg-remove a product still teaser: true # animate the hero into a 5s motion clip music: true # add a licensed soundtrack reel: true # stitch into a platform-cut reel ``` > **It generalizes.** Real-estate: `aesthetic:` "warm golden-hour interiors, wide-angle, natural oak & limestone, lifestyle staging"; shots become rooms. Travel: "sun-drenched film-stock postcards, kodak portra, golden coast light"; shots become destinations. Education: "clean isometric explainer plates, soft pastel, generous whitespace"; shots become concept panels. Same pipeline, new BRIEF. ## The pipeline (what the agent runs) ### Step 1 — Lock the look, render the campaign (proven, ~30s, ~$0.25) For each of the 8 shots, call `create_media` with **one shared aesthetic prefix** + the per-shot subject: ``` create_media({ action: "generate", model_override: "krea-2-os", aspect_ratio: "4:5", prompt: ". " }) ``` The shared prefix is what binds the campaign. Keep it **byte-identical** across all 8 calls; vary only the trailing subject clause. (Want a touch more fidelity on the hero frames? Swap `krea-2-os` → `krea-2-large`.) ### Step 2 — (Optional) Lock the look as a reusable brand LoRA (~6–8 min, ~$22) Once you love the campaign, freeze it into a LoRA so *every future look* is on-brand by construction: ``` # MCP submit_lora_train({ zip_url: <8–16 on-brand images>, trigger_word: "ATELIERNORD", base_model: "krea-2", steps: 500 }) get_lora_train({ job_id }) # poll to done → weights_url attach_lora_to_project({ lora_id, project_id }) # now every create_media on this project is on-brand # or one-shot: apply_lora({ weights_url, trigger: "ATELIERNORD", prompt: "a new SS26 look, full length", model: "krea-2-lora" }) ``` ``` # CLI livepeer lora train --zip --trigger ATELIERNORD --base krea-2 --steps 500 livepeer lora train-status livepeer lora apply --trigger ATELIERNORD --prompt "a new SS26 look" --model krea-2-lora ``` ``` # Webapp: right-click canvas → Train LoRA → base = "Krea 2 (Open-Source)" → # /lora apply → every free-form generate is now on-brand ``` > **The base is the inference base.** A `krea-2` LoRA must run on `krea-2-lora` (`fal-ai/krea-2/turbo/lora`); the surfaces auto-route it for you. Train a **style/aesthetic** LoRA on `krea-2`; train a tight **face/product-identity** LoRA on `flux-dev` and stack them. > > **Honest status (2026-06-24):** the `krea-2-lora-training` cap is registered, accepts jobs, and meters them. Full weight-retrieval through the SDK currently hits a known 202-polling follow-up (the orch returns HTTP 202 for async training and the SDK's status poller doesn't yet treat 202 as "still running"). The recipe above is wired across MCP + CLI + webapp and lights up the moment that SDK fix lands. Until then, Step 1 (the `krea-2-os` campaign) is the fully-proven path — and it already delivers a cohesive, on-brand deliverable. ### Step 3 — Finish into a deliverable (proven) ``` bg-remove : create_media({ action:"animate", model_override:"bg-remove", source_url: }) # packshot kling-v3-turbo-i2v: create_media({ action:"animate", model_override:"kling-v3-turbo-i2v", source_url:, prompt:"subtle cinematic push-in", duration:5 }) sonilo-t2m : create_media({ action:"music", model_override:"sonilo-t2m", prompt:"minimal ambient piano, quiet-luxury, 20s" }) ffmpeg-concat : stitch the 8 stills + teaser → one MP4 ffmpeg-export : scale+crop to 9:16 (TikTok) / 1:1 (IG) / 16:9 (hero banner) ``` ## Reference render (this playbook, run for real) The showcase page below is built from a real run of Step 1 + Step 3 on `sdk.daydream.monster`: **ATELIER NORD** — 8 `krea-2-os` shots (~3s each), one packshot, one motion teaser, one reel. See [`/chapters/train-your-brand-aesthetic-example.html`](/chapters/train-your-brand-aesthetic-example.html). Durable assets: `gs://storyboard-uploads/showcases/train-your-brand-aesthetic/`. ## Tips - The **aesthetic sentence** is 80% of the result. Spend your effort there, not on per-shot prose. - 8–16 images is the sweet spot for a style LoRA; use *diverse subjects, one consistent look*. - For a recurring **person/product** (a house model, a hero SKU), train a `flux-dev` identity LoRA and stack it on top of the `krea-2` aesthetic LoRA. - Keep `trigger_word` unique and made-up (e.g. `ATELIERNORD`), never a real word.