Midjourney V8 vs V7: The Prompt Anatomy That Changed

Midjourney’s V8 family changed the contract between you and the model. V7 was an interpreter: it read your prompt as a suggestion and embellished toward its own taste. V8 is an executor: it follows your specification far more literally, with --raw as the dial for how much house taste remains – a contract that rewards precision and punishes vibes. Meanwhile the parameter surface shifted under the prompt box – quality values gone, multi-prompts gone, seeds now approximate, upscalers replaced by native HD, and one reference type that silently runs on the old model. This is the migration map I keep next to the prompt box, verified against Midjourney’s own version documentation.1

TL;DR

• The mental model flipped. V7 interpreted and beautified; the V8 family executes far more literally, with --raw as the literalism dial. “Dramatic lighting” was a V7 prompt; “single overhead key light, no fill, hard shadows” is a V8 prompt.
• The prompt hierarchy still rules: subject first, then details, context, style, technicals, parameters. Early tokens carry the most weight in both generations – leading with style words was always the amateur tell, and V8’s literalism makes it costlier.
• Removed or changed from V7: --q quality values (not supported), multi-prompts with :: (not V8-compatible), byte-identical seeds (now “99% identical”), the separate upscaler step (replaced by native 2048px --hd), and Niji/Turbo (not V8 paths).1
• Still works: --no, --raw, style references, image prompts and weights, moodboards, chaos, stylize, weird, --exp, personalization.1
• The trap: --oref Omni Reference on a V8 job is annotated “(Uses V7)” – the parameter is accepted, but V7 renders the job, so you get V7’s look while believing you are in V8.1
• The production gotcha: editing an HD render (Pan, Zoom Out, Vary Region) silently downscales it to SD. Upscale last, after all edits.1
• Full parameter-by-parameter coverage lives in the Midjourney guide; this post is the V7-to-V8 delta.

The Regime Change: Interpreter to Executor

The single most useful thing to internalize about the V8 family (V8.1, default June 10 to July 23, 2026; V8.2, default since July 2412) is that it stopped guessing on your behalf. V7 would take “beautiful cinematic portrait” and supply its own lighting scheme, palette, and mood – often gorgeously, always opinionated. V8 hands the opinion back to you.

The practical consequences:

Specify like a cinematographer, not a mood board. “Single overhead key light with no fill, hard shadows” beats “dramatic lighting.” “35mm film photograph, grain, Kodak Portra 400 palette” beats “vintage look.” Named anchors still work and work better – “Annie Leibovitz portraiture,” “Roger Deakins cinematography” – because they compress a precise specification into two words.

--raw is the dial for how much house taste survives. With Raw on, Midjourney’s default aesthetic interpretation drops away almost entirely; without it, V8.2 in particular applies a bolder house style than V8.1’s calmer default.2 If a pipeline was tuned against V8.1’s look, V8.2’s “more creative, bold, sophisticated, and edgy” default reads as a regression until you re-tune.2

The hierarchy matters more, not less. The prompt anatomy that worked in V7 – subject, subject details, context, style/mood, technical, parameters, in that order – survives intact, because token position still maps to influence. What changed is the penalty for violating it: V7 would rescue a style-first prompt with its own judgment; V8 renders your disordered priorities faithfully.

What V7 Had That V8 Does Not

The compatibility chart is the part of the docs everyone skips and then pays for.1 The deltas that actually bite:

And the one that deserves its own paragraph: --oref and --ow are accepted on V8 jobs but annotated “(Uses V7)” – the render is executed by V7.1 Nothing warns you. If your omni-reference outputs look subtly unlike your other V8.2 work, they are not V8.2 work.

What V8 Gives Back

The trade is real, not punitive. The V8 family brings roughly 4-5x faster standard jobs, native 2048px output behind --hd (1.3 GPU minutes against 0.8 for SD), and Draft Mode’s 24-image exploration grids at half the fast-hours of a standard job (on V8.1 and, per the July 27 docs revision, V8.2 – see the FAQ).13

Style references got a generational upgrade of their own: the --sv 7 style-version, which became the baseline during the V8 Alpha, runs about 4x faster and 4x cheaper than the previous system, and moodboards default to it on the same terms.4 Old codes are not stranded – --sv 4 and --sv 6 still interpret V6-era codes – but regenerating references against the current system beats compatibility mode.

V8.2’s headline is Personalization: profiles read your accumulated taste “much better,” new profiles draw on a larger image pool, and – the migration relief – your V7 profiles work in V8.2 unchanged. There is still no Global V8 profile; the global slot remains your V7 profile.5

Three Gotchas That Cost Real GPU Minutes

The HD edit round-trip. Pan, Zoom Out, or Edit/Vary Region on an HD image downscales the result to SD – the docs say so plainly, and the workflow implication is strict ordering: compose and edit first, upscale to HD as the final step.1 An edit-heavy pipeline that upscales early pays for HD twice.

The HD aspect cap. Aspect ratios run to 14:1 normally but only 4:1 under HD.1 A cinematic banner at 8:1 quietly cannot be an HD render.

The New Profile version trap. Personalization profiles are created against your current default version setting – so set the version you want before clicking New Profile, or you build a profile against the wrong model.5

The Migration Checklist

1. Re-read your standard prompts and replace every mood word with its specification – lighting, lens, film stock, palette.
2. Once you have written the spec yourself, add --raw so house style does not fight it; drop it where V8.2’s taste is welcome.
3. Replace --style raw with --raw everywhere – the rename is the most common migration stumble.
4. Delete --q from templates; move quality-tier workflows to V7 explicitly.
5. Split any :: multi-prompts into single specifications.
6. Audit for --oref – decide per use whether V7’s rendering is acceptable, since that is what you are getting.
7. Re-order edit-then-upscale pipelines so --hd (or the upscale) comes last.
8. Regenerate high-use --sref codes against --sv 7; keep --sv 4/--sv 6 only for archival codes.
9. Set your default version before creating new Personalization profiles.

The full guide covers every parameter and mode in current detail, including the Draft Mode compatibility conflict between Midjourney’s own docs pages. (An earlier snapshot, Midjourney V8 Prompting, captures the V8.0 Alpha as it stood in April 2026; for migration decisions, this post is the current map.)

FAQ

What is the main difference between Midjourney V7 and V8?

The interpretive stance. V7 treated prompts as suggestions and applied heavy default aesthetics; the V8 family executes prompts literally, so precision in lighting, medium, and composition language matters far more. Mechanically, V8 adds ~4-5x speed and native 2048px HD, while dropping quality tiers, multi-prompts, exact seed reproduction, and the separate upscaler step.1

Do V7 style references and personalization profiles work in V8?

Personalization: yes – V7 profiles are compatible with V8.2, and your Global V7 profile works across the family.5 Style references: old V6-era codes need --sv 4 or --sv 6 compatibility mode, while the V8 baseline is the faster, cheaper --sv 7 system – regenerate frequently-used codes rather than living in compatibility mode.4

Why do my V8 images look different from V7 with the same prompt?

Because V8 stopped supplying the taste your prompt left unspecified. V7 filled gaps with its own aesthetic; V8 renders what you wrote. Replace mood words with specifications (“single overhead key light, no fill” rather than “dramatic lighting”), and use --raw to control how much of Midjourney’s remaining house style applies.12

Does Draft Mode work on V8.2?

Draft Mode reached V8.1 in June 2026, and as of an August 2026 re-check the docs now lean yes on V8.2 too: Midjourney revised the Draft Mode article on July 27 with explicit V8.1/V8.2 specifications (24 images at 512x512px, 0.4 GPU minutes per prompt, web-only), though the compatibility chart still lags with Draft un-checked on the merged column. Verify with one draft prompt on each version before building a pipeline; the --sref random sampler rides the same feature.13

1. Midjourney, Version documentation – the default-model lineage (V7 to June 9, 2026; V8.1 June 10-July 23; V8.2 from July 24), the merged V8.1/V8.2 compatibility column, quality/multi-prompt/Niji/Turbo non-support, seeds “(99% identical),” --oref “(Uses V7),” the HD edit downscale note, and the 14:1 vs 4:1 HD aspect caps. ↩↩↩↩↩↩↩↩↩↩↩↩↩↩
2. Midjourney, Version 8.2 – the July 24, 2026 default release: “an update focused on aesthetics, image quality, and Personalization”; “more creative, bold, sophisticated, and edgy”; reduced low-quality outliers. ↩↩↩↩
3. Midjourney, Draft Mode for V8.1 – 24-image draft grids at half the fast-hours of a V8.1 SD job, June 16, 2026. Draft Mode’s V8.2 status: Draft & Conversational Modes, revised July 27, 2026 with explicit V8.1/V8.2 specifications (24 images, 512x512px, 0.4 GPU minutes, web-only), while the compatibility chart still un-checks Draft on the merged V8 column. ↩↩
4. Midjourney, Relax Mode for V8 Alpha (March 2026) – the --sv 7 style-version baseline: about 4x faster and 4x cheaper than the prior system, supporting --hd, --p, --stylize, and --exp, with moodboards defaulting to it; --sv 4 and --sv 6 retained for legacy codes. ↩↩
5. Midjourney, Personalization documentation – V7 profile compatibility with V8.2, the absent Global V8 profile, larger image pool for new V8.2 profiles, and profile creation against the current default version setting. ↩↩↩