krea

name

krea

description

Krea API in Nebula — generate images with Krea's own Krea 2 model steered by reference images, moodboards, and custom-trained LoRA styles, plus search and train Krea styles. Activate when the user configures any krea-2-generate, krea-style-search, or krea-style-train node, or asks about Krea in Nebula. Sourced from docs.krea.ai and cross-checked against backend/handlers/krea.py + backend/data/node_definitions.json on 2026-06-04; see the Nebula audit guide at docs/api-guides/krea.md.

Krea Skill

When to use

• User configures any krea-2-generate, krea-style-search, or krea-style-train node.
• User wants a Krea 2 image (text-to-image), optionally steered by reference images, a moodboard, or a trained style.
• User wants to train a custom LoRA style/object/character on Krea and reuse its style_id.
• User wants to search the Krea style library to find a style_id to plug into a generation.
• User asks which Krea model/param to pick, about KREA_API_TOKEN, or about chaining styles into a generation.
• Boundary check: if the user asks for Krea video, image editing / image-to-image, upscaling, Krea 2 Turbo, or any third-party model behind Krea's gateway (Flux, Imagen, Ideogram, Nano Banana, etc.), this provider does not expose them — see Capability boundaries and route elsewhere.

Universal rules

1. Auth. Single bearer token sent as Authorization: Bearer <token>. Env var is KREA_API_TOKEN. (The handler also accepts a legacy KREA_API_KEY fallback, but KREA_API_TOKEN is the documented name — use it.) Missing key raises KREA_API_TOKEN is required. Get the key from the Krea dashboard → Developers → API Keys & Billing.
2. Base URL. https://api.krea.ai (hard-coded as KREA_BASE_URL in the handler). No version segment in the path.
3. Execution pattern (confirmed from the handler):
   • krea-2-generate → async-poll. POST /generate/image/krea/krea-2/{variant} returns {job_id, status}; the handler polls GET /jobs/{job_id} every 2 s (cap 300 polls = ~10 min) until terminal. If the submit response already reports status: completed, it skips polling.
   • krea-style-train → async-poll. POST /styles/train returns {job_id}; polled GET /jobs/{job_id} every 5 s (cap 300 polls).
   • krea-style-search → sync. GET /styles returns the list in one round-trip; no job.
4. Job lifecycle / status. Pending set: backlogged, queued, scheduled, processing, sampling, intermediate-complete. Terminal: completed, failed, cancelled. A failed/cancelled job raises with the job's error/message. An unrecognized status also raises (the handler is strict). Completed generate jobs expose result image URLs at job.result.urls; completed train jobs expose job.result.style_id.
5. Error / status codes. Non-2xx responses raise Krea <action> failed (<code>): <detail>, where detail is pulled from the JSON error/message. Known: 401/403 bad or missing token; 402 Krea balance depleted (the handler substitutes the message "Krea API balance is depleted" if the body is empty). Validation errors surface client-side before any call: missing prompt → Prompt input is required; missing training name → Krea style training requires a name; no training images → Krea style training requires at least one training image.
6. Input-URI rules (images). style_images / image_style_references / training images accept: public http(s):// URLs (passed straight through), data: URIs, /api/outputs/… paths from upstream nodes, and local file paths under the Nebula output root. Anything that is not already an http(s) URL is auto-uploaded to Krea via POST /assets, and the returned image_url is used. Paths outside the output root are rejected. So you can wire any upstream generator's image output directly into style_images — Nebula handles the upload.
7. Key gotchas.
   • resolution is 1K-only. It is an enum with a single option; any other value raises Krea 2 currently supports only resolution=1K.
   • variant has exactly two values: medium (default) and large. Krea 2 Medium Turbo is not exposed.
   • At most 10 image style references total across style_images + image_style_references + native-moodboard-derived refs combined; exceeding raises an error. Native-moodboard refs only fill the remaining slots after explicit refs.
   • At most one moodboard. Connecting more than one moodboard ID raises Krea 2 currently accepts at most one moodboard.
   • Style-training visibleWhen quirk: learning_rate and batch_size are only shown (and only sent) for flux_dev, flux_schnell, wan, wan22. They are hidden for qwen and z-image — don't set them for those models; they're ignored. Additionally, for qwen/z-image the handler coerces training_type: Default → Style.
   • The handler downloads the first result image to the run dir and returns a local path on the image port; the raw Krea job object is returned on the job port (useful for seeds/debugging and additional URLs).

Pick the right node

Nebula node (display name)	Node ID	Endpoint / model	Key params
Krea 2	krea-2-generate	POST /generate/image/krea/krea-2/{variant} (Krea 2, medium/large) — async-poll	variant (medium/large), aspect_ratio (8 ratios), resolution (1K only), creativity (raw/low/medium/high), seed, style_reference_strength, style_id, style_strength, moodboard_id, moodboard_strength
Krea Style Search	krea-style-search	GET /styles — sync	filter (all/user/community/krea/shared/public/gallery), model, ids, user, liked, limit, cursor
Krea Style Train	krea-style-train	POST /styles/train (+ optional POST /styles/{id}/share/workspace) — async-poll	name (required), model (6 base models), training_type, trigger_word, max_train_steps, learning_rate, batch_size, generation_strength, share_with_workspace

* learning_rate / batch_size shown only for flux_dev/flux_schnell/wan/wan22.

Param reference

krea-2-generate (Krea 2)

Input ports: prompt (Text, required) · style_images (Image, up to 10) · image_style_references (Any, up to 10) · styles (Any, multiple) · moodboard (Any, max 1). Output ports: image (Image — local file path) · job (Any — raw Krea job object).

Param	Type	Range / enum	Default	Notes
variant	enum	medium, large	medium	Picks the path segment; large is higher quality / cost.
aspect_ratio	enum	1:1, 4:3, 3:2, 16:9, 2.35:1, 4:5, 2:3, 9:16	1:1	Validated against the exact set; anything else raises.
resolution	enum	1K	1K	Only value accepted.
creativity	enum	raw, low, medium, high	medium	How far Krea may deviate from the prompt/refs.
seed	integer	0 – 2147483647	random	Omit for random; set for repeatability.
style_reference_strength	float	0 – 1	0.5	Default strength applied to each connected style_images reference (per-ref override possible via image_style_references items).
style_id	string	—	—	Paste an existing Krea style ID directly (in addition to / instead of wiring the styles port).
style_strength	float	-2 – 2	1	Default strength for connected styles + the pasted style_id. Negative pushes away from the style.
moodboard_id	string	—	—	Reference an existing Krea moodboard by ID.
moodboard_strength	float	0 – 1	0.23	Strength for the Krea moodboard_id. (Note: this is the node default; a Nebula native moodboard's reference images default to 0.7 strength — separate path, see chaining.)

How the request body is assembled (from the handler): prompt, aspect_ratio, resolution, creativity are always sent; seed if set; image_style_references is the merged list of {url, strength} from style_images + image_style_references + native-moodboard refs; styles is a list of {id, strength}; moodboards is a list of {id, strength} (≤1). A Nebula native moodboard also appends a Nebula moodboard direction: <styleBrief> suffix to the prompt.

krea-style-search (Krea Style Search)

Input ports: none. Output ports: styles (Array — raw style objects) · text (Text — a summary listing up to 10 title — id [models] lines).

Param	Type	Range / enum	Default	Notes
filter	enum	all, user, community, krea, shared, public, gallery	all	Scope of the library to search.
model	string	e.g. flux_dev, qwen, z-image, wan	—	Filter by the model a style was trained on.
ids	string	comma-separated style IDs	—	Fetch specific styles by ID.
user	string	—	—	Filter by owner.
liked	boolean	—	false	Only styles you've liked.
limit	integer	1 – 1000	25	Page size.
cursor	string	—	—	Pagination cursor from a prior page.

Only non-empty params are sent as query params. The styles array (the API's items) feeds directly into a krea-2-generate styles port.

krea-style-train (Krea Style Train)

Input ports: images (Image, required, multiple — the training set). Output ports: style (Any — {kind: "krea_style", id, strength}) · style_id (Text) · job (Any — raw job).

Param	Type	Range / enum	Default	Notes
name	string	—	—	Required. Style name.
model	enum	flux_dev, flux_schnell, wan, wan22, qwen, z-image	flux_dev	Base model the LoRA trains on.
training_type	enum	Style, Object, Character, Default	Style	For qwen/z-image, Default is coerced to Style.
trigger_word	string	—	—	Optional activation token to invoke the style in prompts.
max_train_steps	integer	1 – 2000	(API default ~1000)	Training step cap.
learning_rate	float	—	—	Shown only for flux_dev/flux_schnell/wan/wan22. Hidden + ignored for qwen/z-image.
batch_size	integer	≥ 1	—	Shown only for flux_dev/flux_schnell/wan/wan22. Hidden + ignored for qwen/z-image.
generation_strength	float	-2 – 2	1	Strength baked into the emitted style object (used downstream when this style feeds a generation).
share_with_workspace	boolean	—	false	If true, after training the handler calls POST /styles/{id}/share/workspace.

The submit body is {model, type, name, urls} plus optional trigger_word, max_train_steps, and (full-knob models only) learning_rate/batch_size. urls are produced by uploading every connected training image to /assets first.

Recipes

Recipe 1 — Straight Krea 2 text-to-image

1. Drop a krea-2-generate node.
2. Wire a text/prompt source into its prompt port (required).
3. Set aspect_ratio (e.g. 16:9), keep creativity: medium, optionally set a seed for repeatability. Leave variant: medium (or pick large for higher quality).
4. Run. The generated image lands on the image port; the raw job (with seed + all URLs) on job.

Recipe 2 — Style-referenced generation (look transfer)

1. Bring in one or more reference images — uploads, or the image output of any upstream generator node — and connect them into the krea-2-generate style_images port (up to 10).
2. Tune style_reference_strength (0–1, default 0.5) to control how hard the references bias the result.
3. Connect your prompt and run. Nebula auto-uploads any local/generated images to /assets and sends them as image_style_references. (A Nebula moodboard wired into the moodboard port additionally injects its representative images as references and appends a style-brief suffix to the prompt.)

Recipe 3 — Discover a style, then generate with it

1. Drop a krea-style-search node. Set filter: krea (or community), optionally model: flux_dev and limit. Run — read the text summary to pick a style, or pass the whole styles array onward.
2. Wire the styles array output into the krea-2-generate styles port (or copy a single style ID into the node's style_id param).
3. Set style_strength (-2–2, default 1), add your prompt, and generate on-brand images.

Recipe 4 — Train a custom LoRA style, then generate with it

1. Drop a krea-style-train node. Connect a set of training images into images and give it a name (required). Pick model: flux_dev and training_type: Style; optionally set a trigger_word. (Leave learning_rate/batch_size alone unless you've chosen a FLUX/Wan model and know the values; they're hidden for Qwen/Z-Image.)
2. Run it. When training completes it emits style_id (Text) and a style object on style.
3. Wire the style output into a krea-2-generate node's styles port (or paste the style_id into the style_id param), set style_strength, add a prompt, and generate. Optionally toggle share_with_workspace on the train node to make the style available to your whole API workspace.

In the nebula_nodes context

• Node IDs (provider krea): krea-2-generate, krea-style-search, krea-style-train. All three set envKeyName: KREA_API_TOKEN. In the palette, Krea 2 and Krea Style Train sit under image-gen; Krea Style Search is an analyzer node.
• Handler: backend/handlers/krea.py. Entry points: handle_krea_generate, handle_krea_style_search, handle_krea_style_train. Shared helpers: _poll_krea_job (polls /jobs/{id}), _upload_asset / _image_value_to_url (the /assets upload path), _collect_image_style_references, _collect_styles, _collect_moodboards.
• Node defs: backend/data/node_definitions.json (keys krea-2-generate, krea-style-search, krea-style-train). The visibleWhen block on learning_rate/batch_size is what drives the hide-for-Qwen/Z-Image UI behavior.
• Ports & chaining:
  • krea-2-generate consumes prompt (Text, required), style_images (Image ×10), image_style_references (Any ×10), styles (Any), moodboard (Any ×1); emits image + job.
  • krea-style-search consumes nothing; emits styles (Array) + text (Text). The styles array plugs straight into krea-2-generate.styles.
  • krea-style-train consumes images (Image, required); emits style + style_id (Text) + job. The style output (a {kind:"krea_style", id, strength} object) plugs into krea-2-generate.styles; or route the style_id Text into the style_id param.
  • Internal resource shapes the handler recognizes on input ports (not user-facing nodes, but useful to know when an agent emits them): krea_image_style_reference ({kind, image, strength}) on image_style_references; krea_style ({kind, id, strength}) on styles; nebula_moodboard on moodboard. A nebula_moodboard is treated specially — its representative images become style references (default strength 0.7, filling only remaining slots) and its styleBrief is appended to the prompt — whereas a raw moodboard_id string/object goes into the moodboards body field for Krea-side lookup.
• How outputs render: the image port is a downloaded local file served via /api/outputs/…; the canvas image preview renders it. The job port carries the raw Krea job JSON (seeds, all result URLs, status) for debugging or fan-out.

Capability boundaries — what Krea's API can do that Nebula does NOT expose

Nebula deliberately scopes this provider to Krea's own Krea 2 model + the style system. Do not promise the following through a Krea node; route the user to the appropriate other Nebula provider (or tell them it's out of scope here):

• Krea 2 Medium Turbo (/generate/image/krea/krea-2-medium-turbo) — not exposed; only medium + large variants.
• Image-to-image / editing — Flux Kontext, SeedEdit, Seedream 4, etc. (/generate/image/bfl/flux-1-kontext-dev, …) — not exposed.
• Video generation — Veo 2/3/3.1, Kling, Runway Gen-3/4/4.5/Aleph, Hailuo, Seedance, Wan, Ray 2, LTX, Grok Imagine (/generate/video/{provider}/{model}) — not exposed. (Use Nebula's Runway / FAL / Replicate providers for video.)
• Upscaling / enhancement — Topaz, Topaz Bloom, Topaz Generative (/generate/enhance/topaz/…) — not exposed.
• Krea's third-party image-model gateway — Flux, Flux 1.1 Pro/Ultra, Imagen 3/4, Ideogram 2/3, Nano Banana / 2 / Pro, Qwen 2512, Z-Image, ChatGPT Image, Luma, Runway Gen-4, Seedream — not exposed through Krea. (Many of these have their own Nebula nodes; use those.)
• Node apps (saved Krea workflows, POST /node-apps/{id}/execute) — not exposed.
• Style CRUD beyond train + share: get a single style (GET /styles/{id}), update (PATCH /styles/{id}), shareable link, remove-from-workspace — not exposed. Sharing exists only as the share_with_workspace toggle inside krea-style-train, not as a standalone node.
• Asset management — list/get/delete assets — not exposed (upload is used internally only, no user-facing asset node).
• Job management — list jobs (GET /jobs), delete (DELETE /jobs/{id}) — not exposed as nodes (polling-by-id only). As of 2026-06-05, a poll cancellation fires a best-effort DELETE /jobs/{id} so the in-flight Krea job stops upstream instead of running to completion.
• Moodboard creation/listing — Krea has no public create/list moodboard endpoint; krea-2-generate can only reference an existing Krea moodboard_id, or consume a Nebula-native moodboard object.
• Webhooks — job-completion callbacks exist in the API but Nebula uses polling instead.
• 3D generation — Krea has a user-facing 3D feature but no documented public API endpoint; not available here.

Overall Nebula exposes roughly ~10% of the Krea API surface — by design, this provider is "Krea 2 + styles," not Krea's full multi-model gateway.

Sources

• Nebula audit guide: docs/api-guides/krea.md (node table, params, gap table, sourcing).
• Ground truth: backend/handlers/krea.py, backend/data/node_definitions.json (keys krea-2-generate, krea-style-search, krea-style-train).
• Official docs:
  • https://docs.krea.ai/llms.txt — full documentation / endpoint index
  • https://docs.krea.ai/api-reference/introduction.md — base URL, async job model
  • https://docs.krea.ai/api-reference/krea/krea-2-medium.md / krea-2-large.md / krea-2-medium-turbo.md — Krea 2 generation endpoints
  • https://docs.krea.ai/api-reference/styles/search-styles.md, train-a-custom-style-lora.md, share-a-style-with-your-workspace.md — styles & training
  • https://docs.krea.ai/api-reference/assets/upload-an-asset.md — asset upload
  • https://docs.krea.ai/api-reference/general/get-a-job-by-id.md — job lifecycle / polling