How to batch-generate ad variants in Midjourney

Recommendation (short):
Use one anchored reference image + a stable composition prompt, lock the seed, and programmatically vary only 2–3 controlled attributes (color palette, mood/style, CTA/mock copy) to create 30–50 coherent ad variants. Generate prompts via ChatGPT, run batches in Midjourney, then upscale/export final frames and add text in a design tool.

Decision criteria (pick your path):
- Budget: if you have a higher Midjourney plan or credits, generate more variations directly in MJ. If budget is tight, produce fewer high-quality bases then augment in Photoshop/Figma.
- Skill level: Discord and MJ param familiarity required. If you’re unfamiliar, use ChatGPT to produce permutations and follow the checklist below.
- Workflow stage: early creative exploration = broader chaos/stylize. Final A/B-ready assets = strict seeds, fixed aspect ratio, upscaled + text added offline.
- Team size/output quality: small solo teams should favor reproducibility (seeds + spreadsheet); larger teams can parallelize prompt dispatch or use automation tools.

Practical checklist (step-by-step):
1) Create an anchor/reference image: design or photograph a frame with exact composition you want (subject placement, negative space for CTA). Upload to Discord and copy that image URL. This locks composition.
2) Build a stable base prompt template (example below). Always include the reference URL first. Fix composition with --ar and lock seed with --seed so all variants align.
Sample template:
" | main subject: woman holding product centered, 3/4 view, clean background, negative space right for text; camera: 50mm, shallow DOF; lighting: soft window light; color palette: {{PALETTE}}; mood: {{MOOD}}; brand style: {{STYLE_ADJ}} --ar 4:5 --seed 12345 --quality 2 --stylize 50"
3) Decide which attributes to vary (limit to 2–3). Good options: color palette, mood/style adjective, accessory or clothing color. Avoid changing composition or camera terms.
4) Use ChatGPT to generate a matrix of permutations. Example: 6 palettes x 5 moods = 30 prompts. Generate CSV rows with prompt text + metadata (palette, mood, seed).
5) Batch generate: paste prompts into MJ (Discord) in batches. Use the same seed for a family of variants; if you want sub-variants, change only one parameter (e.g., --chaos 10 or swap palette word).
6) For consistency, use the same version of Midjourney and the same upscaler option across all winners. Use the grid output (4 images) then press V# to create close variations if needed, keeping the seed constant.
7) Upscale pipeline: pick best candidates, use MJ Upscale (or Uplight/upbeta for less artifacting), then export PNGs. Do not rely on MJ for crisp type or logos—add final CTAs in Photoshop/Figma (same font, spacing, color across variants).
8) Deliver and test: create ad sets grouping by the variable you’re testing (color vs. mood). Keep naming consistent (e.g., brand_palette_mood_seed.png).

Best-for / Avoid-if
- Best for: fast visual A/B tests where composition must be fixed but surface treatments vary (colors, moods, textures).
- Avoid if: you need precise, readable logos or exact text baked into images (use a design tool to add text), or strict brand-legal imagery that must be 100% controlled.

Practical tips
- Use low/no stylize for photoreal consistency (e.g., --stylize 10–50). Lock the same --seed across a set to preserve layout.
- Keep a spreadsheet mapping prompt -> variant metadata so you can analyze test performance.
- If you need automation beyond manual Discord input, generate prompts with ChatGPT and use a lightweight RPA or an integration (Make/Make.com) to post to Discord—mind rate limits and MJ terms.

If you want, I can: (a) make a CSV of 30 prompts given your brand adjectives and palette, or (b) give a refined prompt tuned for your reference image.

Compare Midjourney and Leonardo AI