Midjourney: variations vs upscale (when to use each)

After you generate a grid in Midjourney you get two button families that sound similar but do different jobs: Variations (V1–V4 on Discord, Vary Subtle/Strong on the web) create new versions of the same concept; Upscale (U1–U4 on Discord, Upscale Subtle/Creative on the web) separates one image from the grid and enlarges it to HD resolution (~2048 px on the long edge with V8.2). Mixing them wrong burns GPU minutes: you upscale a shot you still dislike, or you Vary ten times on an image that was already done. Official docs: https://docs.midjourney.com/hc/en-us/articles/32692978437005-Variations and https://docs.midjourney.com/hc/en-us/articles/32804058614669-Upscalers. In 18 minutes you learn a repeatable flow: explore with variations, lock composition, upscale once.

1. Steps 1

   Generate a test grid with a clear shot

   Open the Midjourney web app (Create) or Discord with /imagine. Write a concrete prompt — not a mood: «matte ceramic mug on wooden table, three-quarter shot, soft window light from left, no hands --ar 4:5». Generate a 2×2 grid. Do not press any U or Vary button yet. Scan the four cells and note which one best keeps subject, angle and light. If none work, the prompt is the problem, not variations: go back to midjourney-prompts-que-funcionan and tighten the shot first.

   What you should see: A fresh 2×2 grid. Four V buttons (V1–V4) and four U buttons (U1–U4) under the grid on Discord, or the Creation Actions panel on the web. No action taken yet.

2. Steps 2

   Understand the fork: explore (V) vs deliver (U)

   Variations build a new grid from one chosen cell: they change details or rearrange the concept while keeping the base idea. Upscale separates one cell and enlarges it — on V8.2 from ~1024 px to ~2048 px on the long edge with Subtle or Creative (docs: https://docs.midjourney.com/hc/en-us/articles/32804058614669-Upscalers). Golden rule: if you still hesitate between the four cells, you are exploring → use V, not U. If you already know the winner and only need pixels to export or retouch, you are delivering → use U. Upscale early then vary still works, but costs more: each upscale uses GPU.

   What you should see: A mental or written note: «explore = V / Vary» vs «deliver = U / Upscale». Grid not upscaled yet.

3. Steps 3

   Explore with V1–V4 or Vary Subtle

   Pick the winning cell (e.g. #2). On Discord press V2; on the web open Creation Actions and choose Vary (Subtle). Subtle makes small changes — ideal when you like the composition but want to refine a detail (shadow, texture, subject placement). Variations docs: https://docs.midjourney.com/hc/en-us/articles/32692978437005-Variations. Generate a new grid and compare: did the axis you care about improve or did only noise change? Write one line on what moved («V2 Subtle: right-edge light reads better»).

   What you should see: A second grid derived from cell 2. Composition similar to the original with one different detail. Your note on what improved.

4. Steps 4

   When to switch to Vary Strong (or V3/V4 with Strong mode)

   If Subtle does not move the shot enough — you want another layout of the same concept, a different angle or a bolder read — use Vary (Strong) on the web or enable Strong Variation Mode on Discord and press V on the chosen cell. Strong keeps the core idea but changes the image more visibly. Use it sparingly: two or three Strong rounds are enough to explore one direction. If you are still lost after Strong, the brief is not locked: rewrite the prompt, do not chain Strong forever.

   What you should see: A Strong grid clearly different from the previous Subtle pass, but still recognizable as the same job (same mug, same set).

5. Steps 5

   Lock composition before you touch Upscale

   To «lock» means you can say in one sentence why this image wins: «cell 2 of the third grid: three-quarter angle, left light, clean table edge». If you cannot articulate it, keep Vary Subtle on that cell or return to the prompt. Do not press U1–U4 or Upscale until you pass this test. This step generates nothing new: it is a criteria pause. Save the exact prompt (parameters included) in a note — you will reuse it.

   What you should see: One written sentence that justifies the winning image. Prompt copied to a pad. Grid without upscale.

6. Steps 6

   Upscale Subtle vs Creative (U + Upscale on Discord, Creation Actions on web)

   With composition locked, separate the winning cell. On Discord: U2 (if #2 won). On the web: Upscale (Subtle) or Upscale (Creative) in Creation Actions. Both double dimensions (~2× on V8.2). Subtle enlarges with minimal changes — pick this almost always for export to design or Runway. Creative adds or changes detail while enlarging; it sometimes fixes hands or faces but can also invent things. You can repeat Creative with different results; each step uses GPU (see https://docs.midjourney.com/hc/en-us/articles/33329329805581-Modifying-Your-Creations). One Subtle upscale for delivery; reserve Creative for when Subtle left a local defect.

   What you should see: A single enlarged image (~2048 px on the long edge), not a grid. File ready to download or take into Runway.

7. Steps 7

   Full session flow and when to stop

   Recap the flow you just practiced: (1) generate grid with a shot prompt, (2) pick a cell, (3) Vary Subtle once or twice, (4) Vary Strong only if exploration needs it, (5) lock with a written sentence, (6) Upscale Subtle once. If after upscale you need to change the whole background light, do not chain Creative five times: return to the prompt or use regional inpaint. For motion, take the upscaled JPG into Runway (runway-primer-clip) — do not upscale again «just in case». If the next step is copy, the still is done; write text in ChatGPT or Claude, not in Midjourney.

   What you should see: A folder with the final upscaled JPG, saved prompt and a flow note (how many V rounds, one U). Session closed with no orphan undecided grids.