Midjourney Quality Parameter and Speed Modes: The GPU-Time Reference (2026)

TL;DR: --quality (--q) trades GPU time for detail. On V7 the values are 1, 2 and 4. On V8.1 and V8.2, Midjourney's own compatibility chart marks the quality parameter unsupported, along with Turbo mode. Cost on the current default model is set by SD, HD and Draft instead. Checked August 27, 2026.

What does the --quality parameter actually do?

It buys GPU time, and detail is what it buys with it. Midjourney's Quality article states the mechanism plainly: quality "controls how much GPU time is spent making your images", and that "can change how the details and textures look".

Three properties matter more than the number you pick.

It is scoped to the first grid. The parameter "only influences the first set of four images you create" and does not affect later variations, inpainting, outpainting or upscales. A --q 4 job followed by three variations is one expensive job and three ordinary ones.

It is multiplicative on cost. On V7, --q 2 and --q 4 "use 2 times or 4 times more GPU time". Those are the only cost multipliers Midjourney publishes for any parameter.

It has no promised effect on the image. Nothing in the documentation says higher quality produces a better picture, only more GPU time and different detail and texture. Those are not synonyms.

a hand-thrown ceramic bowl on raw linen, single window light --ar 4:3 --v 7 --q 1

Which --quality values work on which Midjourney version?

Different values on every model line, and none at all on the current default. This is the part most guides get wrong, by treating one version's range as universal.

The V8 row is not an omission. The Feature Compatibility chart in the Version article uses a check icon and a no-symbol icon in every cell, and the Quality Parameter row carries the no-symbol icon in the V8.1 and V8.2 column, the same icon used for Turbo Mode, Multi-Prompting and Character Reference. V8.2 has been the default model since July 24, 2026.

Two V7-specific rules are worth memorising because they cost money silently. There is no --q 3 setting, and Midjourney's Quality article says that if you use it, it "will automatically switch to --q 4". A typo therefore quadruples the bill rather than erroring. And --q 4 is not compatible with Omni Reference, so a --q 4 --oref prompt is not a premium job, it is a broken one.

If you are still working out which model flag you are on, the version flag reference covers --v and --niji separately.

What are Fast, Relax and Turbo, and what does each one cost?

Three GPU speeds, and the difference between them is which pool of compute you draw from and how fast it is handed to you. Midjourney's own analogy in the GPU Speed article is a phone plan: subscribing buys time on the GPUs, not a number of images.

Fast Mode spends the metered Fast time your plan grants each month. It resets at renewal and does not roll over. The GPU Speed article is blunt about it: "any unused time doesn't carry over to the next month."

Relax Mode spends nothing metered. Jobs join a queue and run as GPUs free up, with published waits of "0 to 30 minutes" that scale with your own recent usage relative to other subscribers.

Turbo Mode also spends Fast time, but at double the rate, in exchange for a separate high-speed GPU pool that the article describes as "up to four times faster than Fast Mode". It is documented as available on versions 5 and later, and as experimental, with the caveat that "its availability and costs might change anytime". If the Turbo GPUs are busy or your model does not support it, the job silently reverts to Fast.

Turbo's own speed claim has drifted: "3.5X faster at 2x the cost" at the V6 launch on March 7, 2024, "up to four times faster" in the help centre today.

You switch speed three ways. Append --fast, --relax or --turbo to a single prompt. Set a default in the settings panel next to the Imagine bar. Or, in Discord, use /fast, /relax or /turbo as standalone commands. Web and Discord settings are synced, which the GPU Speed article warns about directly, and which is the most common reason people burn Fast time they meant to save.

Does Relax mode have usage limits?

No metered cap, but three kinds of documented limit that catch people out.

The first is feature gating. Midjourney's GPU Speed article lists four exclusions together: permutation prompts, the repeat parameter, HD-resolution videos and Max Upscale, a legacy upscaler, are "not available while using Relax mode". The Permutations and Repeat articles say the same thing from the other direction, that both features "only work in Fast and Turbo Modes".

That is the trap. Permutations and --repeat are the tools you reach for when you want volume, and volume is why you switched to Relax. Every permutation Midjourney expands is billed as its own Fast job: "using permutations can quickly consume your available GPU time."

Plan gating is the second limit. Relax for images is available on Standard, Pro and Mega, not on Basic. Relax for video is narrower still, Pro and Mega only, standard resolution only. Concurrency, repeat size and permutation size are capped separately per tier.

Third, Relax availability is a promise Midjourney has already qualified. When Relax arrived for the V8 alpha on March 21, 2026, the announcement said: "There is a chance we run out of servers. If we do, we may change availability or speed of relax mode."

How much GPU time does each job type cost?

Midjourney publishes a table, and it is more precise than any paraphrase of it. From the GPU Speed article, updated June 10, 2026, plus the Draft and Video figures from their own pages.

Video batch size is the largest single cost lever in the product. The Video article publishes --bs 1, --bs 2 and --bs 4 at 2, 4 and 8 minutes for standard resolution, and 7, 13 and 26 for HD. Dropping from the default batch of four to one takes an HD video job from 26 minutes to 7.

The GPU Speed article adds two rules of thumb without numbers: upscaling, nonstandard aspect ratios and older versions "might require more time", while variations or lower quality settings "often uses less".

Is HD the default in Midjourney, or SD?

Standard resolution, on the most recent dated evidence, and this has flipped twice inside 2026.

• April 14, 2026, the V8.1 Alpha post: "HD mode is now 3x faster and 3x cheaper. It's so cheap we're making it default for V8.1."
• April 30, 2026, the V8.1 Updates post: "To save compute during a server transition, we're making SD default for a bit. You can still switch to HD under settings panel or add --hd after your prompt."

Nothing published since has reversed that again, and the current Version article, updated July 24, 2026, describes HD as something you turn on, in the Version section of the settings panel or with --hd and --sd. Note the phrase "for a bit". It was framed as temporary and never formally ended.

The resolution figures themselves are stable and published. At the default 1:1 ratio, V8.2 makes 1024 by 1024 pixel SD images and 2048 by 2048 pixel HD images. The Upscalers article publishes exact dimensions for exactly four ratios, HD being double SD in each: 4:3 is 1232 by 928 and 2464 by 1856, 2:3 is 896 by 1344 and 1792 by 2688, 16:9 is 1456 by 816 and 2912 by 1632. Anything outside those four is calculated by someone, not published.

Three consequences worth writing on a sticky note:

HD narrows your aspect ratio ceiling. The Version chart's Max Aspect Ratio row reads 14:1 for V6 and V7 and 14:1 (4:1 for HD) for V8.1 and V8.2. A 12:1 banner is a standard-resolution job. The full aspect ratio reference has every ratio and pixel size.

HD cannot be upscaled further. Per the Upscalers article, HD images "generate at 2048 x 2048px automatically, and currently cannot be upscaled further". Upscaling an SD image reaches the same pixel dimensions for 2 minutes of GPU time, against the 0.5 minute premium of generating in HD directly.

Editing an HD image downgrades it. Per the Version article, using Pan, Zoom Out or Edit and Vary Region on an HD image "will downscale the resulting images to SD". Getting back to HD costs another upscale.

Does --draft mode still exist?

Yes, and on the V8 series it is the cheapest thing in the product. It is also the clearest documentation conflict Midjourney currently has open.

The Draft and Conversational Modes article, updated July 27, 2026, has a section headed "Draft Mode in V8.1 / V8.2" marked with a "new" flag. One Draft prompt on those models generates 24 images at 512 by 512 pixels for 0.4 minutes of GPU time, half the cost of one standard-resolution prompt returning four. You then click Vary or Remix on the keepers to re-render them at full resolution.

The same article's opening line says "Draft mode is compatible with Midjourney versions 7 and 8.1", omitting 8.2. And the Version article's compatibility chart, updated three days earlier, marks Draft Mode with the no-symbol icon for V8.1 and V8.2 together. Three Midjourney statements, three different answers, all live on August 27, 2026. The June 16, 2026 changelog announcing Draft for V8.1 is the tie-breaker on cost: "Draft jobs use half as many fast hours as V8.1 SD jobs (even though they generate 24 images!)".

Draft on V7 behaves differently: four images at half the cost of a standard V7 prompt, with an Enhance button that re-runs a chosen draft at full cost. Draft on V8.1 and V8.2 is web-only, "not currently available in Discord".

Do --stylize, --raw and --exp change the cost?

Midjourney does not say they do, and that absence is the answer rather than a hedge.

The Stylize article documents a default of 100 and a range of 0 to 1000 on current versions, with settings-panel presets at 50, 100, 250 and 750. The Raw article describes turning off Midjourney's default aesthetic. Neither makes any cost claim at all, so whether they change GPU time is simply not published. Our longer look at --raw covers what that flag does to the image.

--exp is barely documented at all. Across all 105 articles in Midjourney's help centre, the string appears twice: as a row in the Version compatibility chart, marked available on V7, V8.1 and V8.2, and in a passing line of the Omni Reference article. There is no --exp article. Everything known about it comes from the April 30, 2025 changelog that introduced it:

• "An experimental image aesthetics parameter", running 0 to 100, with 0 as the default.
• "Similar to --stylize (and you can combine it with stylize)", making images "more detailed, possibly more dynamic, creative, and more 'tone mapped'".
• "The prompt accuracy will degrade as you go to higher values as well as the image diversity."
• Recommended values are 5, 10, 25, 50 and 100, most of the visible change falling between 5 and 50.

No GPU cost figure appears in that post or anywhere else. If you see a number for what --exp costs, it did not come from Midjourney. The same caution applies to --sref, where the only cost statement on record is a March 21, 2026 note that a new style-reference version is "4x faster/cheaper" than the previous one, selectable back with --sv 6. Our style reference guide covers the syntax.

Nineteen copy-paste prompts for controlling Midjourney cost

Each block differs only in its cost levers. Run them against your own account and read the usage counters afterwards, because that is the only ground truth for your plan.

V8.2, the current default. Quality is not a lever here, resolution and Draft are.

# 1. Cheapest exploration pass. Explicit SD, unmetered queue.
a brutalist concrete chapel at first light, sea fog pouring over the steps --ar 16:9 --v 8.2 --sd --relax

# 2. The same idea as a delivery pass. HD, metered, 1.3 min instead of 0.8.
a brutalist concrete chapel at first light, sea fog pouring over the steps --ar 16:9 --v 8.2 --hd --fast

# 3. Draft sweep. 24 images at 512px for 0.4 min. Web only.
a brutalist concrete chapel at first light, sea fog pouring over the steps --ar 16:9 --v 8.2 --draft

# 4. A ratio HD cannot do. 14:1 is SD-only on V8.1 and V8.2.
a panoramic salt flat at dusk, mirrored sky, single distant figure --ar 14:1 --v 8.2 --sd

# 5. Relax-safe volume. No --repeat, no permutations, so it will actually queue.
weathered fishing buoys stacked on a stone harbour wall, overcast --ar 4:3 --v 8.2 --sd --relax

# 6. Fast-only volume. --repeat is rejected in Relax mode.
weathered fishing buoys stacked on a stone harbour wall, overcast --ar 4:3 --v 8.2 --sd --fast --r 4

# 7. Permutation across ratios. Also Fast or Turbo only. Three separate billed jobs.
a lone red kayak on black volcanic sand --v 8.2 --sd --fast --ar {1:1, 4:3, 16:9}

# 8. Aesthetic ladder in Draft, 0.4 min per rung. Permutations need Fast or Turbo.
an abandoned tram depot reclaimed by ivy --ar 3:2 --v 8.2 --draft --fast --exp {0, 10, 50}

# 9. Stylize ladder. No published cost difference between the rungs.
an abandoned tram depot reclaimed by ivy --ar 3:2 --v 8.2 --sd --fast --s {50, 100, 250, 750}

# 10. Raw plus HD for a controlled photographic delivery.
a chef's hands folding pasta dough on flour-dusted marble --ar 2:3 --v 8.2 --hd --raw

V7, where the quality parameter is documented and does something.

# 11. Baseline. --q 1 is the V7 default; stating it makes the comparison honest.
a 1970s reading room, oak shelving, dust in a shaft of afternoon light --ar 3:2 --v 7 --q 1

# 12. Double GPU time.
a 1970s reading room, oak shelving, dust in a shaft of afternoon light --ar 3:2 --v 7 --q 2

# 13. Quadruple GPU time. Experimental when it shipped in May 2025.
a 1970s reading room, oak shelving, dust in a shaft of afternoon light --ar 3:2 --v 7 --q 4

# 14. The expensive typo. There is no --q 3; V7 promotes it to --q 4.
a 1970s reading room, oak shelving, dust in a shaft of afternoon light --ar 3:2 --v 7 --q 3

# 15. Documented incompatibility. --q 4 does not combine with Omni Reference.
a portrait of the same character in a rain-soaked alley --v 7 --q 4 --oref https://your-image-url.png

# 16. V7 Draft. Four images at half the standard V7 cost, then Enhance the keeper.
a 1970s reading room, oak shelving, dust in a shaft of afternoon light --ar 3:2 --v 7 --draft

# 17. Turbo on V7. Double Fast time, separate GPU pool, unsupported on V8.1 and V8.2.
a 1970s reading room, oak shelving, dust in a shaft of afternoon light --ar 3:2 --v 7 --turbo

# 18. Relax on V7. Unmetered, but the queue is 0 to 30 minutes.
a 1970s reading room, oak shelving, dust in a shaft of afternoon light --ar 3:2 --v 7 --relax --q 2

Video, where the numbers get large.

# 19. Batch 1 instead of the default 4. SD video drops from 8 min to 2 min.
# Video needs a starting frame image; add it in the Imagine bar before submitting.
slow push in on rain running down a diner window at night --motion low --bs 1

What Midjourney does not publish

Writing this list down beats filling the gaps with plausible numbers, because every one of these is currently answered confidently by pages that made the answer up.

• What --q 2 does on V8.2. The chart says unsupported. No page says whether it errors, is ignored, or falls through.
• Any GPU cost for --exp, --stylize, --raw, --chaos, --weird or --sref. Not one figure, anywhere.
• The multiplier for nonstandard aspect ratios. "Might require more time" is the whole statement.
• A current per-version Turbo speed figure. "Up to four times faster" today, "3.5X faster" at the V6 launch.
• Any service level for Relax queue waits. Zero to thirty minutes is a description, not a commitment.
• --preview. Introduced in the June 16, 2026 changelog to test early model builds, absent from all 105 help-centre articles.
• Whether HD is default again. The last dated statement is the April 30, 2026 reversal to SD.

How to check all of this yourself

The parameter surface here moved four times in 2026 alone, so the durable skill is verification, not memorisation.

Read the compatibility chart, not the parameter article. The Feature Compatibility chart in the Version article is the only page stating per-version availability for quality, Turbo, Draft, HD and --exp side by side. Where it disagrees with a parameter article, check the update date on both.

Use /info in Discord. It reports Fast time remaining as X of Y, plus your Fast, Turbo and Relaxed usage totals. It is the only view showing what your account actually spent, rather than what a table says it should have.

Watch for the settings panel changing itself. Midjourney's August 21, 2026 alpha-site changelog lists three fixes in this category: "Vary now respects your settings instead of quietly running in HD", "removing the HD bubble no longer silently changes your speed", and "Relax no longer flips back to Fast". Alpha bugs, but they show where surprise GPU spend comes from.

Fetch the docs properly. docs.midjourney.com returns 403 to most automated fetches, which is not a dead link. The Zendesk Help Center JSON API behind it answers plainly, and updates.midjourney.com serves the dated changelog posts the help centre summarises later and sometimes contradicts. Every figure here was read from one of those two, on August 27, 2026.

Rather than rebuilding a prompt every time a default flips, keep the parameter block as a saved template with version, speed and resolution as variables. It is the same discipline that makes an image model comparison worth running rather than guessing at.