1.3 · Generate, upscale, variations

Level 1 · Beginner: first Midjourney prompts

1.3Generate, upscale, variations

Goal: master the iteration cycle and stop rewriting the prompt when a simple variation is enough.
Estimated time: 11 min · self-paced reading · If you only have 5 min: read “The essentials in 30s” + the diagram.

The essentials in 30 seconds

  • Grid of 4 then 3 moves: U1-U4 (upscale = finalise in HD), V1-V4 (variations = refine the direction), re-roll (start over).
  • Two intensities: Vary (Subtle) for the almost-good image, Vary (Strong) when the direction is good but the rendering is not enough.
  • Funnel: re-roll → variations → vary subtle → upscale — vary if it is the execution, rewrite if it is the intent.

Running mini-case — Sonia finds her gem without rewriting

Sonia launches her terrace prompt and gets 4 proposals: no. 2 has the right composition but an approximate cup. Instead of rewriting, she clicks V2 (variations), then Vary (Subtle) when it is almost good, and finally U2 (upscale) for the HD deliverable. On a second try where all 4 images miss (wrong mood, wrong framing), she does not vary: she fixes the prompt. She learns the rule: 3-4 re-rolls on a good prompt before concluding, never persisting on a bad lead.

Sequel to cases 1.1-1.2 — Sonia moves from structure to iteration.

1. The basic cycle: grid, upscale, variations, re-roll

One generation produces a grid of four proposals from the same prompt. Three actions structure everything: Upscale (U1-U4 on Discord, equivalent on the web: enlarges in HD, it is the final version), Variations (V1-V4: new proposals inspired by a kept image, same spirit, different details) and Re-roll (relaunches everything from zero, with no link to the previous ones).

Upscale is not a simple enlargement: it is a new generation pass that adds details (textures, sharpness) — but it can slightly change the image and does not fix structural errors (hands, anatomy, text). Sonia upscales in HD only the truly kept image.

2. Intensities and the funnel method

Vary (Subtle) = light variations, perfect when the image is almost good; Vary (Strong) = strong variations, when the direction is good but not yet convincing. Added to these are Vary (Region), Pan and Zoom Out for frame retouching (section 4.1).

Funnel method: re-roll until you find a lead → variations to refine it → vary subtle to finish → upscale. The decision rule: execution problem (good composition, detail at fault) = vary; intent problem (all 4 off) = fix the prompt. And never judge an idea on a single image.

Common misinterpretation: “You must rewrite the prompt at every disappointment” or “upscale fixes everything”. No. Most execution disappointments are fixed with variations, and upscale adds details without fixing structural flaws — vary on the good lead, rewrite only when the intent is missed.
The funnel iteration cycle From the lead to HD without needless rewriting Re-roll Find a leadVariations Refinethe directionUpscale Finalise in HD
From the lead to HD without needless rewriting
Text version of the diagram (SEO / screen readers / translation): 1) Explore: Sonia does re-rolls until she finds an interesting lead (3-4 tries on a good prompt). 2) Refine: she varies (V2) then adjusts with subtle when it is almost good, without rewriting. 3) Finalise: she upscales (U2) the kept one in HD and downloads it. Lesson: funnel re-roll → variations → subtle → upscale; only a missed intent justifies rewriting.

The frantic U button (never do this)

“I upscaled all 4 images of each grid, just in case.”
Verdict: 4 HD files to sort, fast hours gone — upscale the keeper, period.

Moral: upscaling everything is like printing every blurry photo.

🔎 Concrete example — 🔎 Concrete example — Vary rather than rewrite

Execution vs intent — like Sonia on her cup.

/imagine a steaming cup of coffee on a wooden table, sunny terrace, watercolor, golden hour --ar 4:5
Good composition, faulty detail → V2 then Vary (Subtle) then U2.
All 4 off → fix the prompt. Poor randomness → re-roll.

Reading: varying keeps the direction found; rewriting is only useful when the intent is missed.

Lesson: 3-4 re-rolls often reveal a gem on a solid prompt.

Section vocabulary

Grid of 4
The four proposals generated from one prompt, among which you choose.

Simple example: The 4 terraces Sonia received, including the kept no. 2.

What it is not: a single image to judge definitively.

Reflex: choose the closest lead before any other move.

Upscale (U1-U4)
High-definition enlargement of the chosen image; produces the final usable version.

Simple example: Sonia's U2 delivering the cup in HD to the client.

What it is not: a fixer for hands or unreadable text.

Reflex: upscale only the true keeper.

Variations (V1-V4)
New proposals inspired by a kept image: same spirit, different details.

Simple example: Sonia's V2 keeping the composition and redoing the cup.

What it is not: a relaunch with no link to the previous images.

Reflex: vary when the execution is what sticks.

Re-roll
Full relaunch of the prompt to get four new proposals with no link to the previous ones.

Simple example: Sonia's 3 re-rolls before finding her lead.

What it is not: a correction of the prompt text.

Reflex: re-roll a good prompt before calling it bad.

In practice — Hands-on — Run the generate → upscale cycle

  1. Launch a structured prompt and watch the four proposals: which one is closest to your intent?
  2. If none fits, re-roll (or fix the prompt); if one lead is good, ask for Variations.
  3. Refine with Vary (Subtle) when the image is almost good, or Vary (Strong) if the direction is good but the rendering is not enough.
  4. Finish with an Upscale of the kept image, then download it.
You master the iteration cycle and stop rewriting the prompt when a simple variation is enough.

Ready-to-say script — iterate as a funnel (30s)

“I launch my /imagine, I look at the grid of 4: if no. 2 has the right composition I click V2, then Vary Subtle when it is almost good, then U2 for HD. If all 4 miss, I fix the prompt instead of varying. I do 3 re-rolls before judging.”

Say while clicking V then U live; show the subtle/strong difference.

Copy-paste template — vary-or-rewrite decision

IF good composition + faulty detail → V1-V4 then Vary (Subtle) then U.
IF all 4 off (framing, mood, element) → fix the prompt (missing block).
IF good prompt + poor randomness → re-roll x3-4.
FINAL → Upscale the keeper.

Almost-perfect cup

Like Sonia: V2 on the good composition, subtle for the cup, U2 in HD — delivered to the café without rewriting.

Missed mood

4 images off (evening instead of morning): fix golden hour to morning light instead of varying into the void.

Wide exploration

Good but soft direction: Vary (Strong) to explore boldly before tightening back to subtle.

Structural flaw

Hands or unreadable text enlarged by upscale: do not insist, move to targeted retouching (4.1) or Photoshop (4.4).

✓ Best practices

  • Follow the funnel: re-roll → variations → vary subtle → upscale, with no skipped step.
  • Vary when the problem is execution, fix the prompt if it is intent.
  • Relaunch a good prompt 3-4 times before concluding: 4 images are only a sample.

✗ Common pitfalls

  • Rewriting the prompt when a simple variation would have been enough.
  • Persisting with variations on a lead that was never right from the start.
  • Judging an idea on a single image or upscaling all 4 instead of the keeper.

Key takeaways

  • One generation produces 4 proposals; Upscale finalises, Variations refine, Re-roll restarts from zero.
  • Vary (Subtle) = fine adjustment; Vary (Strong) = wider exploration of the same direction.
  • Funnel method: re-roll → variations → vary subtle → upscale.
  • Frequent mistakes: rewriting the prompt instead of varying, or persisting on a lead that was never right.

48h Challenge — apply tomorrow

Launch 1 solid prompt: do 3 re-rolls, pick the best lead, ask for V + subtle, then U. Note at each step whether it was execution or intent.

Section exit checklist
  • I tell varying (execution) apart from rewriting (intent).
  • I chain re-roll → variations → subtle → upscale.
  • I upscale only the keeper and I download it in HD.
Self-rating 1-5: do I vary instead of rewriting when the composition is good?do I do enough re-rolls before judging?

Frequently asked questions

When should I edit the prompt rather than ask for variations?

If it is execution (right composition and mood, faulty detail) → variations. If it is intent (all 4 off: framing, mood, missing element) → fix the prompt. If the 4 are very different and uneven → the prompt is too vague, add constraints.

Does upscale really improve the image or just enlarge it?

It does both: a new generation pass that adds details (textures, sharpness), not a simple stretch. But it can slightly change the image and does not fix structural errors (hands, anatomy, text) — retouch them (4.1, 4.4). For very large format, a dedicated upscaler completes it.

Check your understanding

What is the effective iteration method?

Additional resources (optional, to go further)

Level 1 examFresh, harder questions to test what you've learned