Ai Image Generation Skills

AI video, image generation. 40+ models — Sora, Veo 3, Kling, Seedance, GPT Image, Hailuo, WAN. Text-to-video, image-to-video, text-to-image,image-to-image.

Pixwith Media Generation

Use This Skill When

Use this skill when the user wants to work with Pixwith through MCP to:

• generate images from text
• edit images with one or more reference images
• generate videos from text
• generate videos from a reference image
• inspect available Pixwith models
• check remaining Pixwith credits

Preconditions

On first install or first use of this skill, explicitly guide the user through setup before attempting any Pixwith tool call.

Use wording like:

"Before this skill can generate media, OpenClaw must be connected to Pixwith through MCP and you need a Pixwith API key.

• Step 1: Open OpenClaw MCP settings and add the Pixwith MCP server.
• Step 2: Go to https://pixwith.ai/api and create your Pixwith API key.
• Step 3: Paste that API key into your Pixwith MCP configuration in OpenClaw.
• Step 4: After setup is complete, come back here and I can help you list models, upload images, and create generation tasks."

If the user asks for a copyable MCP configuration, provide a ready-to-paste example using the Pixwith MCP endpoint and remind the user to replace the API key placeholder with their real key:

{
  "mcpServers": {
    "pixwith": {
      "url": "https://api.pixwith.ai/mcp",
      "headers": {
        "Api-Key": "YOUR_PIXWITH_API_KEY"
      }
    }
  }
}

If OpenClaw expects a different MCP configuration shape, adapt the same values to the host format instead of changing the endpoint or auth header.

Do not assume the user understands what MCP is. Describe it as the connection required for OpenClaw to use Pixwith.

If MCP tools are unavailable, tell the user Pixwith is not connected yet and repeat the setup guidance in plain language.

This skill assumes the Pixwith MCP server is already connected in OpenClaw and the following tools are available:

• list_models
• get_model_schema
• upload_image
• generate
• get_task_result
• get_credits

If these tools are unavailable, stop and tell the user the Pixwith MCP server must be configured before this skill can work.

Every request requires a valid Pixwith Api-Key. If authentication fails, send the user to https://pixwith.ai/api.

Core Workflow

1. Discover models

Call list_models first unless the user already provided a known model_id.

This step is mandatory because model availability is dynamic. Do not rely on static model lists in this skill package when choosing what to run.

• use type="image" for image jobs
• use type="video" for video jobs
• keep the returned model_id, name, and summary

2. Inspect the selected model

Always call get_model_schema(model_id) before generate.

Treat input_schema as the source of truth for:

• whether image_urls are allowed
• maximum image count
• valid option keys
• enum values
• defaults
• estimated time
• credit display

Treat min_credits from list_models only as a quick lower-bound cost hint:

• it is useful for fast filtering
• it is not the final price for the specific request
• for the final price, inspect input_schema.credits

Do not hardcode model parameters when schema data is available.

3. Check credits for expensive jobs

Call get_credits before expensive or long-running jobs, especially video jobs or higher-resolution image jobs.

If credits are insufficient, do not retry automatically. Tell the user to recharge at https://pixwith.ai/pricing.

4. Upload reference images when needed

Call upload_image when the image is local, private, temporary, or otherwise unreliable as a public URL.

Use the returned Pixwith-hosted image_url in generate.image_urls.

5. Create the task

Call generate with:

• input.prompt
• input.model_id
• input.image_urls only when allowed by schema
• input.options that conform to input_schema.options

Before submission:

• prompt must not be empty
• model must come from list_models
• image count must satisfy schema limits
• options must be schema-valid

6. Poll until terminal state

Call get_task_result until status becomes terminal.

• 1: processing
• 2: completed
• 3: failed

Recommended polling:

1. wait about 75% of estimated_time
2. poll once
3. if still processing, wait until around estimated_time
4. poll again
5. if still processing, poll every 10 seconds

Only expect result_urls when status is 2.

Guardrails

• Pixwith is asynchronous. Do not present generate as an immediate result.
• Pixwith tool responses are wrapped. Inspect code before trusting data.
• Do not guess mode support from the model_id prefix. Pass image_urls only when the selected schema allows or requires them.
• Some models are dual-mode under one model_id. For example, Flux MCP IDs use text-to-image without image_urls and route to Flux Kontext when image_urls is present.
• Prefer schema-valid retries over speculative retries.
• Relay backend error messages accurately.

References

• For authentication, wrapped tool responses, upload semantics, task states, and failure handling, read references/service-contract.md.
• For a non-authoritative model snapshot and common parameter patterns, read references/model-snapshot.md.

Examples

• For a text-to-image workflow, read examples/text-to-image.md.
• For an image-to-video workflow, read examples/image-to-video.md.