wavespeed-batch

name	wavespeed-batch
description	Runs batch image or video generation across an A B variant axis tracking cost per variant for cheap fast experiments
version	0.1.0

Wavespeed Batch

Batch generator for cheap and fast A/B tests. Takes a base prompt and a variant axis (e.g., 5 different hooks, 3 different palettes), runs the matrix on Wavespeed, and returns the artifacts plus per-variant cost. Used to filter ideas before committing to a more expensive provider.

When to invoke

• Hook testing across 5 to 20 caption variants paired with the same image template.
• Palette or composition exploration on a quote card or product shot.
• Pre-flight before a Higgsfield render: pick the winning variant on Wavespeed first.
• Batch generation of carousel slides where each slide shares a template but differs in copy.
• Cost-conscious early ideation when the brief is still loose.

Inputs

• base_prompt: string. Shared prompt body across all variants.
• variant_axis: object. { field, values }. Field is the prompt slot that changes (hook, palette, subject). Values is the list of options.
• kind: string. image or video.
• aspect: string. 1:1, 9:16, 4:5, 16:9.
• model: string. Wavespeed model id (flux-schnell, sdxl-turbo, wan-video).
• count_per_variant: integer. Default 1. Useful when sampling multiple seeds per variant.
• seed: integer, optional. Base seed; per-variant seeds are derived deterministically.
• cost_cap_usd: number, optional. Stop and return partial results if the cap is hit.

Process

1. Validate inputs. Reject if variant_axis.values is empty or base_prompt is missing.
2. Expand the matrix: for each value in variant_axis.values, build a final prompt by injecting the value into base_prompt at the { field } placeholder.
3. For each final prompt, derive a per-variant seed from seed and the variant index for reproducibility.
4. Estimate cost per variant from the model's price-per-call table. Sum and check cost_cap_usd.
5. Submit the batch to Wavespeed in parallel with a sane concurrency cap (default 5).
6. Collect the artifacts. For each, record the variant value, the seed, the artifact path or URL, and the actual cost.
7. If a variant fails, retry once. If it fails again, mark it as failed in the result.
8. Append a summary line per variant to data/wavespeed-usage.jsonl.

Outputs

• results: array of objects. Each { variant_value, prompt, seed, artifact_ref, cost_usd, status }.
• total_cost_usd: number.
• successful_variants: integer.
• failed_variants: integer.

Examples

Example 1: hook test across 5 variants

Input: { base_prompt: "Editorial portrait of a woman, soft light, brand mood. Hook overlay: { hook }.", variant_axis: { field: "hook", values: ["Color is your edge", "Stop guessing your palette", "Your style starts here", "Discover your season", "Wear what works"] }, kind: "image", aspect: "1:1", model: "flux-schnell", count_per_variant: 1 } Output: { results: [5 entries with artifact paths], total_cost_usd: 0.12, successful_variants: 5, failed_variants: 0 }

Example 2: palette exploration on a quote card

Input: { base_prompt: "Quote card with text 'Sua cor te trai'. Palette: { palette }.", variant_axis: { field: "palette", values: ["beige and black", "navy and gold", "off-white and rust"] }, kind: "image", aspect: "1:1", model: "sdxl-turbo" } Output: 3 artifacts with cost summary.

Failure modes

• cost_cap_usd hit mid-batch: stop submission, return partial results, set failed_variants to the unattempted count, and surface a warning.
• All variants fail: surface the underlying provider error and refuse to write a misleading success log.
• variant_axis placeholder not found in base_prompt: throw a descriptive error pointing at the missing placeholder.
• Concurrency cap too high for the API: drop to a safe default (3) and warn.

Related skills

• gpt-image-prompt-builder: precise alternative when typography matters more than batch speed.
• higgsfield-prompt-builder: cinematic alternative for the winning variant after the Wavespeed test.
• topview-prompt-builder: UGC alternative when the variant axis is the avatar or scene.
• qa-tech-specs: validates each artifact against platform specs.
• video-prompt-builder: dispatcher that may delegate to this skill for batch hook tests.