krea

Generate images, videos, upscale/enhance images, and train LoRA styles using the Krea.ai API. Supports 20+ image models (Flux, Imagen, GPT Image, Ideogram, S...

Security Scan

The skill's code largely matches its stated purpose (wrapping the Krea.ai API), but the package metadata omits the KREA_API_TOKEN requirement and some helper functions (update checks / notifications / uploads) can contact additional endpoints or act on local files — review those parts before installing or supplying credentials.

Purpose & Capability

Name/description match what the code does: scripts call the Krea.ai API to list models, generate images/videos, enhance/upscale, and train LoRA styles. There are no unrelated credentials requested. However, the scripts require a KREA API token at runtime (via --api-key or KREA_API_TOKEN), but the registry metadata declared no required env vars/primary credential — this is an inconsistency.

Instruction Scope

Runtime instructions direct the agent/user to run Python scripts that: (a) fetch the Krea OpenAPI spec from https://api.krea.ai/openapi.json, (b) make POST/GET calls to the Krea API, (c) perform HEAD checks on arbitrary training-image URLs, (d) accept local file paths and resolve/upload them via ensure_image_url, and (e) download result URLs. These are expected for the stated purpose, but they also mean the skill will: access arbitrary external URLs supplied by the user, upload local files (possible PII/IP leakage), and make network requests beyond a single fixed endpoint. The SKILL.md does not call out all of these privacy/network implications explicitly.

Install Mechanism

No install spec is provided (instruction-only installer), which keeps risk lower. Scripts declare dependency on Python >=3.10 and the 'requests' library (via header comments). No archived downloads or external installers are used. Code writes caches and manifests to standard user locations (~/.cache/krea, .pipeline-state.json, output-dir).

Credentials

The code requires an API key (KREA_API_TOKEN environment variable or --api-key) via get_api_key(), but the skill metadata lists no required environment variables or primary credential. Asking for an API key is appropriate given the purpose, but failing to declare it in registry metadata is an incoherence that can mislead users about secrets the skill will need or persist.

Persistence & Privilege

always:false (good). The scripts create a per-user cache (~/.cache/krea/openapi_models.json), pipeline manifest files (.pipeline-state.json) and may write training manifests to an output-dir. This is expected for a CLI wrapper; nothing indicates the skill tries to modify other skills or escalate privileges. However, helper functions referenced (check_for_updates, send_notification, ensure_image_url) are invoked and their exact network/side-effect behavior is not shown in the truncated code — these should be reviewed because they can enlarge the network footprint (e.g., contacting GitHub or external notification endpoints).

Scan Findings in Context

[none] expected: No regex-based scan findings were reported. Absence of findings does not imply safety; the runtime instructions and helper functions perform network I/O and file uploads which require manual review.

What to consider before installing

This skill is broadly coherent with its description (it wraps the Krea.ai API) but has a few things to check before you install or run it: 1) The code expects a Krea API token (KREA_API_TOKEN or --api-key) even though the registry metadata doesn't declare it — do not supply your global secrets without verifying. Prefer passing --api-key for single runs instead of setting a long-lived env var if you have concerns. 2) The scripts can upload local files (ensure_image_url), perform HTTP HEAD checks on arbitrary URLs (train_style validation), download results, and cache the OpenAPI spec in ~/.cache/krea — review the ensure_image_url, check_for_updates, and send_notification helpers (their implementations are in krea_helpers.py) to confirm they don't contact unexpected endpoints or leak data. 3) Run list_models.py first (without your API key) to see what endpoints look like; inspect krea_helpers.py fully (the file is long and partially truncated in the listing) to confirm update/notification behavior. 4) If you want to reduce risk, run these scripts in an isolated environment (throwaway VM/container) and avoid uploading sensitive local files. If you can't review the helper functions, treat the skill as untrusted and avoid supplying high-privilege credentials.

Krea AI — Image, Video & Enhancement Generation

Generate images, videos, upscale/enhance images, and train LoRA styles using the Krea.ai API. Supports 20+ image models (Flux, Imagen, GPT Image, Ideogram, Seedream...), 7 video models (Kling, Veo, Hailuo, Wan), and 3 upscalers (Topaz up to 22K).

IMPORTANT: Do NOT invent model names. Run list_models.py to get the live list of models, CU costs, and accepted parameters from the Krea API's OpenAPI spec. All scripts resolve models dynamically from the spec — there are no hardcoded endpoint tables. Scripts also accept full endpoint paths from list_models.py --json output (e.g. --model /generate/image/google/imagen-4-ultra).

Usage

Scripts are in the scripts/ directory alongside this file. Run them with uv run from the user's working directory so output files are saved where the user expects.

Generate image:

uv run ~/.codex/skills/krea/scripts/generate_image.py --prompt "your description" --filename "output.png" [--model nano-banana-2] [--width 1024] [--height 1024] [--api-key KEY]

Generate video:

uv run ~/.codex/skills/krea/scripts/generate_video.py --prompt "your description" --filename "output.mp4" [--model veo-3.1-fast] [--duration 5] [--aspect-ratio 16:9] [--api-key KEY]

Enhance/upscale image:

uv run ~/.codex/skills/krea/scripts/enhance_image.py --image-url "https://..." --filename "upscaled.png" --width 4096 --height 4096 [--enhancer topaz-standard-enhance] [--api-key KEY]

Train a LoRA style:

uv run ~/.codex/skills/krea/scripts/train_style.py --name "my-style" --urls-file images.txt [--model flux_dev] [--type Style] [--trigger-word "mystyle"] [--api-key KEY]

List available models:

uv run ~/.codex/skills/krea/scripts/list_models.py [--type image|video|enhance]

Run a multi-step pipeline:

uv run ~/.codex/skills/krea/scripts/pipeline.py --pipeline pipeline.json [--api-key KEY]

Check job status:

uv run ~/.codex/skills/krea/scripts/get_job.py --job-id "uuid" [--api-key KEY]

Important: Always run from the user's current working directory so files are saved where the user is working.

Default Workflow (draft → iterate → final)

Goal: fast iteration without burning CU on expensive models until the prompt is right.

• Draft (cheap/fast): use --model z-image or --model flux-1-dev (3-5 CU, ~5s) for quick iteration

  uv run ~/.codex/skills/krea/scripts/generate_image.py --prompt "<draft prompt>" --filename "yyyy-mm-dd-hh-mm-ss-draft.png" --model flux-1-dev
  

• Iterate: adjust prompt, keep trying with cheap models

• Final (high quality): switch to --model gpt-image or --model nano-banana-pro

  uv run ~/.codex/skills/krea/scripts/generate_image.py --prompt "<final prompt>" --filename "yyyy-mm-dd-hh-mm-ss-final.png" --model nano-banana-pro
  

Available Models

Models, CU costs, and accepted body fields are fetched live from the Krea API's OpenAPI spec (/openapi.json). Run list_models.py to see what's currently available:

uv run ~/.codex/skills/krea/scripts/list_models.py                     # all models with params
uv run ~/.codex/skills/krea/scripts/list_models.py --type image         # image models only
uv run ~/.codex/skills/krea/scripts/list_models.py --json               # machine-readable

Short aliases (e.g. flux for flux-1-dev) are maintained for convenience. The scripts resolve them automatically via the spec. If a model isn't in the alias list, pass the full OpenAPI model ID or endpoint path.

Model selection guidance

Map user requests for images:

• "fast", "quick", "cheap" → flux-1-dev or z-image
• "high quality", "best" → nano-banana-pro or gpt-image
• "text in image", "typography" → ideogram-3
• "photorealistic" → seedream-4 or nano-banana-pro
• No preference → nano-banana-2

Map user requests for video:

• "fast" → hailuo-2.3
• "cinematic", "high quality" → veo-3.1
• "with sound", "with audio" → veo-3 with --generate-audio
• No preference → veo-3.1-fast

Enhancers: topaz-standard-enhance (faithful upscaling, default), topaz-generative-enhance (creative enhancement), topaz-bloom-enhance (adding creative details).

Image Generation Parameters

Param	Description	Default
--model	Model ID or alias (run list_models.py)	nano-banana-2
--prompt	Text description (required)	—
--filename	Output filename (required)	—
--width	Width in pixels (512-4096)	1024
--height	Height in pixels (512-4096)	1024
--aspect-ratio	Aspect ratio (1:1, 16:9, 9:16, 4:3, 3:2, etc.)	—
--resolution	1K, 2K, 4K (nano-banana models)	—
--seed	Seed for reproducibility	—
--image-url	Input image URL or local file path for image-to-image	—
--style-id	LoRA style ID to apply	—
--style-strength	LoRA strength (-2 to 2)	1.0
--batch-size	Number of images (1-4)	1
--steps	Inference steps, 1-100 (flux models)	25
--guidance-scale	Guidance scale, 0-24 (flux models)	3
--quality	low/medium/high/auto (gpt-image)	auto
--output-dir	Output directory	cwd
--api-key	Krea API token	—

Video Generation Parameters

Param	Description	Default
--model	Model ID or alias (run list_models.py)	veo-3.1-fast
--prompt	Text description (required)	—
--filename	Output filename (required)	—
--duration	Duration in seconds	5
--aspect-ratio	16:9, 9:16, 1:1	16:9
--start-image	URL or local file path for image-to-video	—
--end-image	End frame URL (kling only)	—
--resolution	720p, 1080p (veo only)	720p
--mode	std, pro (kling only)	std
--generate-audio	Generate audio (veo-3 only)	false
--output-dir	Output directory	cwd
--api-key	Krea API token	—

Enhancement Parameters

Param	Description	Default
--enhancer	Enhancer ID (run list_models.py --type enhance)	topaz-standard-enhance
--image-url	Source image URL or local file path (required)	—
--filename	Output filename (required)	—
--width	Target width (required)	—
--height	Target height (required)	—
--enhancer-model	Sub-model variant	Standard V2
--creativity	1-6 (generative) or 1-9 (bloom)	—
--face-enhancement	Enable face enhancement	false
--sharpen	Sharpening 0-1	—
--denoise	Denoising 0-1	—
--scaling-factor	Upscaling factor 1-32	—
--output-format	png, jpg, webp	png
--output-dir	Output directory	cwd
--api-key	Krea API token	—

LoRA Training Parameters

Param	Description	Default
--name	Style name (required)	—
--model	Base model: flux_dev, flux_schnell, wan, qwen, z-image	flux_dev
--type	LoRA type: Style, Object, Character, Default	Style
--urls	Training image URLs (space-separated)	—
--urls-file	Text file with one URL per line	—
--trigger-word	Trigger word to activate the LoRA in prompts	—
--learning-rate	Learning rate	0.0001
--max-train-steps	Max training steps	1000
--batch-size	Training batch size	1
--timeout	Polling timeout in seconds	3600
--skip-validation	Skip URL HEAD-check validation	false
--output-dir	Directory to save training manifest	—
--api-key	Krea API token	—

Training requires 3-2000 images. The script validates all URLs before submitting. Training takes 15-45 minutes. On completion, the style ID is printed to stdout and a training-manifest.json is saved if --output-dir is set.

Use the style ID with --style-id in generate_image.py:

uv run ~/.codex/skills/krea/scripts/generate_image.py --prompt "mystyle product on white background" --style-id "style_abc123" --model flux-1-dev --filename "branded.png"

API Key

Scripts check for API key in this order:

1. --api-key argument (use if user provided key in chat)
2. KREA_API_TOKEN environment variable

If neither is available, the script exits with an error message.

Preflight + Common Failures

Preflight:

• command -v uv (must exist)
• test -n "$KREA_API_TOKEN" (or pass --api-key)

Common failures:

• Error: No API key → set KREA_API_TOKEN or pass --api-key
• 402 Insufficient credits → top up compute units at https://krea.ai/settings/billing
• 402 This model requires a higher plan → model needs a paid plan upgrade at https://krea.ai/settings/billing
• 429 Too many requests → concurrent job limit reached; scripts auto-retry up to 3 times with backoff
• Job failed → check prompt for content moderation issues, try different wording

Filename Generation

Generate filenames with the pattern: yyyy-mm-dd-hh-mm-ss-name.ext

• Timestamp: current date/time in yyyy-mm-dd-hh-mm-ss (24h format)
• Name: descriptive lowercase text with hyphens (1-5 words)
• Extension: .png for images, .mp4 for videos

Examples:

• Prompt "A cyberpunk cat" → 2026-03-31-14-23-05-cyberpunk-cat.png
• Prompt "waves on a beach" → 2026-03-31-15-30-12-beach-waves.mp4

Prompt Handling

For generation: Pass user's description as-is to --prompt. Only rework if clearly insufficient.

For image-to-image: Use --image-url with the source image and describe the desired transformation in --prompt.

For video from image: Use --start-image with the source image and describe the desired motion/action in --prompt.

Preserve user's creative intent in all cases.

Output

• Scripts download the result and save it to the current directory (or --output-dir)
• Script outputs the full path to the generated file
• Do not read the image/video back — just inform the user of the saved path
• If --batch-size > 1, files are saved as name-1.png, name-2.png, etc.

Examples

Quick draft image:

uv run ~/.codex/skills/krea/scripts/generate_image.py --prompt "A serene Japanese garden with cherry blossoms" --filename "2026-03-31-14-23-05-japanese-garden.png"

High quality final:

uv run ~/.codex/skills/krea/scripts/generate_image.py --prompt "A serene Japanese garden with cherry blossoms, golden hour lighting" --filename "2026-03-31-14-25-30-japanese-garden-final.png" --model nano-banana-pro --resolution 4K

Image-to-image edit:

uv run ~/.codex/skills/krea/scripts/generate_image.py --prompt "transform to watercolor painting style" --filename "2026-03-31-14-30-00-watercolor.png" --image-url "https://example.com/photo.jpg" --model nano-banana-pro

Generate video:

uv run ~/.codex/skills/krea/scripts/generate_video.py --prompt "A majestic eagle soaring over snow-capped mountains at sunrise" --filename "2026-03-31-15-00-00-eagle-mountains.mp4" --model veo-3 --duration 8 --generate-audio

Upscale image to 4K:

uv run ~/.codex/skills/krea/scripts/enhance_image.py --image-url "https://example.com/photo.jpg" --filename "2026-03-31-15-10-00-upscaled.png" --width 4096 --height 4096 --enhancer topaz

Train a LoRA style:

uv run ~/.codex/skills/krea/scripts/train_style.py --name "acme-brand" --model flux_dev --type Style --trigger-word "acmestyle" --urls-file brand-images.txt --output-dir output/acme-brand

List models:

uv run ~/.codex/skills/krea/scripts/list_models.py --type image

Pipelines (Multi-Step Workflows)

For multi-step workflows (generate → enhance → animate, fan_out branching, template variables, parallel execution, resume, dry-run), see PIPELINES.md.

Quick example:

uv run ~/.codex/skills/krea/scripts/pipeline.py --pipeline '{"steps":[{"action":"generate_image","prompt":"a cat astronaut","filename":"cat"},{"action":"enhance","use_previous":true,"enhancer":"topaz-standard-enhance","width":4096,"height":4096,"filename":"cat-4k"}]}'