How to Improve Midjourney Results With Variations and Upscales

Most Midjourney generations do not land perfectly on the first try. The grid gives you four starting points, and from there you use variations and upscales to get closer to what you want. This iterative loop is where the real work happens: choosing which direction to push, how much change to allow per pass, and when to lock in a result at full resolution.

This guide covers the variation and upscale tools available in Midjourney, how they behave differently across model versions, and a workflow for making one change per pass so you stay in control of the output.

How variations work in Midjourney

When you select an image from the grid (U1 through U4 to upscale, or V1 through V4 for variations), you are telling Midjourney to generate new versions based on that specific image. Variations keep the general composition, subject, and mood but introduce changes to details, poses, colors, and environment elements.

The amount of change depends on which variation mode you use. There are two primary modes available after selecting an image.

Vary Subtle

Vary Subtle makes small adjustments. The composition, subject position, and overall feel stay nearly identical. Use this when an image is 90% right and you just need minor refinement to details like facial features, texture quality, or small background elements.

a ceramic vase with crackle glaze on a wooden shelf, soft window light, product photography --ar 4:5 --stylize 200

After generating and selecting U2, clicking Vary Subtle might adjust the glaze pattern slightly, shift the wood grain, or refine the light falloff. The vase shape, shelf, and overall composition remain the same.

Vary Strong

Vary Strong introduces more significant changes. The subject and general concept stay, but Midjourney may alter poses, camera angles, background elements, and color choices more freely. Use this when you like the concept but want to explore different interpretations.

a lone astronaut standing on a red desert planet, massive ringed planet in the sky, cinematic lighting, sci-fi concept art --ar 16:9

Vary Strong on this image might change the astronaut's pose, alter the ring orientation, shift the desert color, or adjust the sky composition. The core concept (astronaut on alien world) stays intact.

Upscaling: resolution and detail

Upscaling in Midjourney increases the resolution of a selected image. In earlier model versions, upscaling also added fine detail that was not present in the grid preview. In V8, the grid images are already generated at higher quality, so upscaling primarily increases pixel dimensions for use in print or large display contexts.

Upscale behavior in V8

V8 generates grid images at higher internal resolution than previous versions. The upscale button still increases the final output size, but you will notice less dramatic quality improvement compared to V5 or V6 where grid previews were noticeably softer.

Separately from the Upscale button, V8 supports --hd as a generation parameter. Append --hd to request an HD job (about 2048px) instead of the default standard-resolution job (about 1024px). Use --hd when you need higher pixel dimensions for print or close crops from the start.

a detailed botanical illustration of a passionflower, scientific accuracy, white background, fine linework --ar 3:4 --hd

The --hd flag tells Midjourney to generate that prompt as an HD image. It does not replace selecting a grid cell and using Upscale when you only need a larger export of an existing result.

The one-change-per-pass workflow

The most effective way to use variations is to make one intentional change per pass. If you try to fix the composition, lighting, and color palette all at once, you lose track of what is working and what is not. A disciplined single-variable approach keeps your iteration focused.

Step 1: Generate the initial grid

a cozy cabin interior at night, warm fireplace light, bookshelves, a sleeping cat on an armchair, hygge atmosphere --ar 3:2 --stylize 300

Review all four results. Pick the one with the best overall composition and mood, even if details are not perfect.

Step 2: Use Vary Subtle for detail fixes

If the composition is right but the cat looks odd or the bookshelf has artifacts, use Vary Subtle. This preserves everything else while giving Midjourney another chance at the problematic details.

Step 3: Use Vary Strong for direction changes

If the image is close but you want a different camera angle or lighting direction, use Vary Strong. Accept that more elements will shift and pick the best result from the new set.

Step 4: Upscale when satisfied

Only upscale when you are happy with the image at the current detail level. Upscaling locks in the result and consumes additional compute. If you know you need HD resolution from the start, include --hd on the generation prompt rather than treating --hd as a post-upscale switch.

Combining variations with prompt adjustments

You are not limited to blind variation passes. Between variations, you can also re-run the prompt with small adjustments. This gives you control that pure variation cannot.

• If lighting is consistently wrong across variations, adjust the lighting descriptor in the prompt and regenerate.
• If the subject pose never quite works, add a more specific pose descriptor before running variations on the new grid.
• If colors drift in unwanted directions, add a palette constraint (muted earth tones, monochromatic blue) and regenerate before varying.

When to use Vary Subtle vs Vary Strong

• Use Vary Subtle when the image is nearly finished and you want to preserve 95% of the composition while fixing one or two small details.
• Use Vary Strong when you like the concept but want Midjourney to explore alternative interpretations of the same idea.
• Use Vary Subtle after Vary Strong to lock down a new direction once Strong gives you something interesting.
• Avoid chaining multiple Vary Strong passes without reviewing. Each pass drifts further from the original, and after 3-4 strong variations you may lose the qualities you started with.

Practical examples of iterative refinement

Portrait refinement

portrait of a middle-aged woman with silver hair, warm studio lighting, shallow depth of field, editorial photography --ar 4:5 --stylize 150

Generate grid. Select U3 for best expression. Vary Subtle twice to fix a slightly awkward hand position. Upscale the final selection for delivery, or regenerate with --hd when you need HD resolution from the start.

Environment concept iteration

ancient overgrown temple in a jungle canyon, volumetric god rays, moss and vines, concept art for a video game --ar 16:9 --stylize 400

Generate grid. U1 has the best depth but the temple structure is too simple. Vary Strong to get more architectural detail. From the Strong results, pick the most interesting structure. Vary Subtle to clean up any artifacts. Upscale final.

Tips for efficient iteration

• Start with a well-written prompt. Variations refine but cannot rescue a vague starting point.
• Decide before each variation pass what single thing you want to change.
• Keep a reference of which image in the grid you selected so you can backtrack if a variation path goes wrong.
• Do not upscale too early. Once upscaled, you lose the option for further grid-level variations on that image.
• Budget your Fast hours. Each variation is a full generation and consumes the same compute as a new prompt.

When a variation path stalls, go back two steps instead of forcing more Strong varies. Reopen the prior grid, pick a different cell, or rewrite only the lighting phrase and regenerate once. The goal is a short decision loop: choose, change one thing, compare, keep or revert. That habit saves Fast hours and produces cleaner series for Design and Video handoffs.

Related guides

• For inpainting and region-specific editing, see /blog/how-to-use-the-midjourney-editor-for-remix-vary-and-editing.
• For writing stronger starting prompts, see /blog/writing-better-midjourney-image-prompts.