/venice-create-image

name: venice-create-image
description: End-to-end image creation workflow with Venice. Use when a user wants finished images, not API docs - "make me a logo", "generate a hero image", "create 4 variations". Covers turning a brief into a prompt, picking a model and its aspect ratios from the live catalog, style presets and style references, variants and seeds, iterating with edit and multi-edit, upscaling, background removal, saving files to disk, and cost. Uses the Venice MCP tools (venice_image_generate, venice_image_edit, venice_image_multi_edit, venice_image_upscale, venice_image_remove_bg, venice_image_styles, venice_list_models, venice_model_details) when available, or POST /image/generate and the /image/* edit endpoints otherwise.

Create images with Venice

A workflow for going from a brief to finished image files. It sits on top of the reference skills: [`venice-image-generate`](../venice-image-generate/SKILL.md) and [`venice-image-edit`](../venice-image-edit/SKILL.md) document every field.

| Step | Venice MCP tool | REST endpoint | |---|---|---| | Find models and their limits | `venice_list_models` (`type: "image"`), `venice_model_details` | `GET /api/v1/models?type=image` | | List style presets | `venice_image_styles` | `GET /api/v1/image/styles` | | Generate | `venice_image_generate` | `POST /api/v1/image/generate` | | Edit one image | `venice_image_edit` | `POST /api/v1/image/edit` | | Combine several images | `venice_image_multi_edit` | `POST /api/v1/image/multi-edit` | | Upscale 2x-4x | `venice_image_upscale` | `POST /api/v1/image/upscale` | | Cut out the background | `venice_image_remove_bg` | `POST /api/v1/image/background-remove` |

Prefer the MCP tools when they are connected: they return images inline so you and the user can review them. Use the REST endpoints when you need the files on disk (see [Save files](#save-files)).

Use when

• The user wants images produced, reviewed and refined in the conversation.
• You need several on-brand variations to choose from.
• A generated or supplied image needs edits, a transparent cutout, or a higher resolution.

Workflow

1. Pin down the brief

Ask only for what changes the result: subject, style or mood, where it will be used (which sets the aspect ratio), how many options, and any text that must appear in the image. Turn the answers into one concrete prompt: subject first, then setting, lighting, composition and style. Put things to avoid in `negative_prompt`, not in the prompt.

2. Pick a model

Call `venice_list_models` with `type: "image"`. Each entry includes capabilities, traits and pricing. Then call `venice_model_details` with the chosen `model_id` and `type: "image"` to read its constraints: supported aspect ratios and resolutions, whether it takes `negative_prompt`, steps, and its style-reference limit.

• Without a `model`, the MCP tool uses its default (`flux-2-pro`).
• Text rendering, photorealism and stylized art vary a lot between models. Pick by the brief, and say which model you used.
• Do not guess `aspect_ratio` or `resolution` values; use the ones `venice_model_details` lists. `/image/generate` doesn't reject an unlisted `aspect_ratio`: most models silently fall back to their default ratio, so a wrong value gives you the wrong shape rather than an error.

3. Choose a look

• **Style preset:** `venice_image_styles` lists names for `style_preset`. Offer two or three that fit the brief.
• **Style references:** to match existing artwork, pass `style_references` as a list of objects, not plain URLs: each is `{ "image": "<URL, data URI or raw base64>", "strength": 0.1-1 }`. Only models whose details report `supportsStyleReferences` take them, up to their `maxStyleReferences`.

4. Generate

Call `venice_image_generate` with `prompt`, `model`, `aspect_ratio` (or `width`/`height`), and optionally `style_preset`, `negative_prompt`, `seed` and `variants`.

• `variants` (1-4) returns several options in one call. Each variant is billed as one image.
• Only the first variant uses your `seed`. On models that honor `seed`, regenerating with the same `seed` and `variants: 1` reproduces a result. Some models ignore it, including GPT Image, Recraft, Nano Banana Pro and Grok Imagine.
• Pass `safe_mode: true` unless the user explicitly wants otherwise and the model allows it. The API defaults it to on, but `venice_image_generate` and `venice_image_edit` send `safe_mode: false` when the argument is left out.

5. Review and refine

Show the results and ask what to change. Then pick the smallest operation that gets there:

| Request | Tool | Notes | |---|---|---| | Change one thing ("make the sky orange", "remove the cup") | `venice_image_edit` | Pass `image_url` and an instruction-style `prompt`. | | Put elements from several images together | `venice_image_multi_edit` | Pass `image_urls` and a `prompt` describing the composition. | | Transparent PNG for a logo, sticker or product | `venice_image_remove_bg` | | | Print or large-display size | `venice_image_upscale` | `scale` 2-4. Keep `creativity` low to stay faithful. | | Start over with a different idea | `venice_image_generate` | Change the prompt, not just the seed. |

Edit and multi-edit use their own models, listed by `venice_list_models` with `type: "inpaint"` (default `firered-image-edit`).

Image inputs must be reachable by Venice: a public `https://` URL, or for edit and multi-edit, a `data:` URL. Through the MCP tools, upscale and background removal need a public `http(s)://` URL. Local file paths are not accepted. The REST upscale endpoint is different: it takes a raw base64 `image` in JSON or a multipart file, not a URL.

6. Deliver

Tell the user which model, prompt, seed and style you used so they can regenerate the image, and save the final files where they asked.

Save files

The MCP tools return images inline, which is right for review but does not write files. To save to disk, call the REST endpoint from a shell with