Midjourney V8.1 Alpha Complete Guide: Speed, 2K & Srefs (2026)

Midjourney V8.1 Alpha launched on April 14, 2026 at alpha.midjourney.com, and it fixes nearly every complaint users had about V8.0. Native 2K images are now the default without a separate upscaling step. HD generation runs 3× faster and 3× cheaper than V8.0. Standard resolution is 50% faster and 25% cheaper — as fast as V7 draft mode. Style references (--sref) are finally stable after being the defining weak point of V8.0. Image prompts with weights are back. The aesthetic drops the over-processed V8.0 look and returns to “the spirit of V7.” This guide covers every change in V8.1, when to use Raw mode versus Standard, how to get the most from the restored sref system, and what V7 holdouts should know before migrating.

V8.0 vs V8.1: What Changed and Why It Matters

V8.0 Alpha launched with a strong performance story: 4–5× faster generation than earlier versions and a new rendering architecture capable of native high resolution. But it shipped with three significant problems that alienated the Midjourney power user base.

First, the aesthetic. V8.0 defaulted to an over-polished, hyper-processed look that many users described as “AI-looking” in exactly the wrong way. The textures felt synthetic, the lighting too controlled. Photographers and designers who had built workflows around V7’s more grounded visual language found V8.0 outputs difficult to integrate into professional work.

Second, --sref instability. Style references — the system that lets you anchor a generation’s aesthetic to a reference image — became unpredictable in V8.0. Images would drift from the reference across a job, or apply the style inconsistently between panels in a grid. For users who relied on sref for brand consistency or character sheets, this was a workflow-breaking regression.

Third, image prompts were removed. The ability to supply reference images directly in your prompt, with adjustable weights, disappeared in V8.0. V8.1 restores all three.

Native 2K Resolution: What It Actually Means

Previous Midjourney versions generated at lower resolutions and used a separate upscaling model to reach print-quality output. V8 introduced a new rendering path that generates natively at 2K (2048×2048 base resolution) without a post-processing upscale step. V8.1 makes this the default behavior for HD jobs.

The practical implications:

• No double GPU cost for upscaling — you get 2K output from a single generation
• Sharper detail in the original render rather than sharpness added by a separate model, which means fewer upscaling artifacts
• Faster total workflow from prompt to final image, since you skip the separate upscale queue
• A --q 4 flag is available if you need additional coherence on complex prompts at the cost of slightly more GPU time

For the subscription economics: HD generation at 2K now costs 1.33 GPU minutes, down from considerably higher in V8.0. Standard resolution (non-HD) now costs under 1 GPU minute, which is a meaningful saving for volume workflows. At that rate, standard resolution in V8.1 is as fast as V7 draft mode — the mode V7 users turned to when they needed quick iteration.

Performance: Speed and Cost

V8.1 improves on V8.0 across the board:

• HD mode: 3× faster and 3× cheaper than V8.0
• Standard mode: 50% faster and 25% cheaper than V8.0
• Standard vs V7 draft: comparable speed, which means you can run V8.1 standard as your default iteration mode without GPU budget anxiety

For users on the Basic plan ($10/month) or Standard plan ($30/month) who were GPU-constrained in V8.0, V8.1 effectively gives you significantly more creative output per billing cycle. The HD job cost reduction is especially impactful for creators who shoot for print-quality resolution as their baseline.

The Aesthetic: Back to V7 Spirit

The V8.1 release notes describe the aesthetic direction as “a consistent and familiar aesthetic in the spirit of V7.” In practice, this means the model prioritizes photographic naturalism over computational polish. Skin textures feel grounded. Lighting has falloff and shadow variance rather than the controlled studio-light uniformity that characterized V8.0. Landscapes have atmospheric complexity rather than the hyper-saturated sharpness that read as “AI-generated” to trained eyes.

This is not a regression to V7’s specific look — V8.1 retains the resolution and coherence advances of the V8 architecture. It is a recalibration of the model’s default stylistic preferences toward outputs that blend more naturally into professional contexts: editorial photography, concept art, product visualization, and architectural rendering. If you found V7 commercially usable and V8.0 too synthetic, V8.1 is the version to evaluate.

Style References (--sref): Finally Reliable

The --sref (style reference) parameter lets you supply an image URL as an aesthetic anchor for your generation. The model extracts the stylistic qualities of the reference — color palette, brushwork, lighting treatment, compositional mood — and applies them to your prompt content while leaving the subject matter intact.

In V8.0, this system broke down in practice. The reference influence was inconsistent between images in a 4-panel grid, sometimes only affecting one panel strongly and drifting in others. For users creating character sheets, brand campaigns, or consistent illustration series, this made sref essentially unusable for production work.

V8.1 fixes this. Style references are described as “super-stable,” meaning the aesthetic anchor holds consistently across all panels in a grid and across successive generations with the same --sref URL. Moodboards — which combine multiple reference images to define a composite aesthetic — are also stable in V8.1.

Basic usage:

your prompt here --sref https://example.com/your-reference-image.jpg

For multiple references:

your prompt here --sref https://ref1.jpg https://ref2.jpg

The --sw parameter (style weight) controls how strongly the reference influences the output, from 0 (minimal) to 1000 (maximum). A value around 100 is a reasonable starting point for most style transfer use cases.

Raw Mode: When to Use It

Raw mode disables Midjourney’s automatic creative embellishment — the layer of model judgment that adds compositional suggestions, color grading, and atmospheric effects to under-specified prompts. In Raw mode, the model generates more literally from your prompt text, producing outputs that are closer to photographic realism with simple prompts and more controllable with complex ones.

Use Raw mode when:

• You are working with detailed, fully-specified prompts and the model’s aesthetic additions are conflicting with your intent
• You need photographic realism — raw-mode images trend toward photo-like output rather than illustrative or painterly aesthetics
• You are using --sref and want the style reference to take full control of the aesthetic without the model adding its own layer on top
• You are supplying image prompts (iweights) and want the image influence unfiltered by model style preferences

Use Standard mode (the default) when:

• You are prototyping and want the model to fill in details you have not fully specified
• You want cinematic, illustrative, or stylized outputs from short prompts
• Speed and iteration rate matter more than precise control

Image Prompts Return with Weights

Image prompts let you supply a reference image alongside your text prompt. The model blends the visual content of the reference into the generated output. The image weight parameter (--iw) controls the ratio: a higher value pulls the output closer to the image reference, a lower value gives more weight to the text prompt.

V8.1 restores this capability after it was removed in V8.0. Usage:

https://your-image-url.jpg your text prompt here --iw 0.5

Combining image prompts with style references is possible in V8.1: the image prompt provides subject and compositional reference, while --sref provides the aesthetic reference. This is the pattern to use for character consistency workflows, where you want the model to recognize a subject from a reference image but render them in a specific illustrative style.

The Prompt Shortener

V8 has a maximum prompt length. When you exceed it, the new Prompt Shortener activates automatically. Rather than truncating your prompt or throwing an error, the model runs a compression step that identifies the semantically critical elements of your prompt and rewrites it within the length limit while preserving intent.

For users who write highly detailed prompts — multi-clause scene descriptions with specific lighting, materials, camera angles, and color specifications — the Prompt Shortener means you can draft freely without manually counting tokens. The automatic compression is not perfect, but in practice it degrades gracefully and surfaces a summary you can review before generation.

Text in Prompts: Improved Accuracy

One of the long-standing limitations of AI image generation is rendering readable text. V8.1 improves text rendering accuracy, and introduces a convention: text wrapped in quotation marks inside your prompt is treated as a high-priority text-rendering instruction. For example:

a minimalist movie poster with the title "APEX" in bold sans-serif type

The quoted text receives focused attention from the text-rendering system, improving the probability of legible, accurately-spelled output. This is still not reliable enough for production use on long text strings, but for short titles, labels, and single words it has become significantly more usable in V8.1 than in any previous version.

Style Creator and Moodboards

The Style Creator tool generates a shareable style code from a set of images you curate. Upload or link a collection of reference images, and the tool distills them into a single --sref-compatible code. This code can be shared with collaborators or bookmarked, giving teams a stable reference anchor that does not depend on any single image URL remaining accessible.

Moodboards — multi-image style references constructed directly in the Midjourney web interface — are now stable in V8.1 after the V8.0 regression. You can build a moodboard from multiple images, weight their relative influence, and save it as a reusable aesthetic preset. For studios and brands managing visual consistency across multiple projects, moodboards are the practical upgrade path from manually managing --sref URLs.

Access: alpha.midjourney.com Only

V8 and V8.1 are only available on alpha.midjourney.com during the alpha period. They are not accessible in Discord, in the main Midjourney web app, or via the API. Creations made in the alpha environment do not appear in the main Midjourney gallery.

All Midjourney subscription tiers have access to the alpha. You sign in at alpha.midjourney.com with your existing Midjourney account credentials. GPU minutes spent in the alpha count against your main subscription allowance, so the cost accounting is unified. The V8 upscalers and edit/inpainting/outpainting tools are on the V8 roadmap and will ship as separate alpha updates before V8 graduates to general availability.

Pricing

Midjourney’s subscription tiers have not changed with V8.1:

• Basic — $10/month, 200 fast GPU minutes
• Standard — $30/month, 900 fast GPU minutes + unlimited relax-mode generations
• Pro — $60/month, 1,800 fast GPU minutes + stealth mode
• Mega — $120/month, 3,600 fast GPU minutes

With HD costing 1.33 GPU minutes and standard costing under 1 minute in V8.1, the Basic plan yields roughly 150 HD images per month at the new V8.1 rates — compared to significantly fewer at V8.0 rates for the same resolution. The cost improvement is real and meaningful for high-volume creators on lower tiers.

Who Should Migrate from V7 to V8.1

If you stayed on V7 through the V8.0 alpha because of the aesthetic issues or sref instability, V8.1 addresses both directly. The V7 holdout case was rational; the V8.1 case is now also rational. The most practical approach: run your 10 most representative prompts in V8.1, compare outputs against V7 at the same prompts, and let the results determine the answer. The speed improvement alone (standard V8.1 ≈ V7 draft mode) means you can iterate faster at comparable cost during the evaluation.

The exceptions: if your workflow depends on Discord-based generation (V8.1 is alpha.midjourney.com only), or if you use features not yet ported to V8 such as the full upscaler suite or inpainting, you will need to stay on V7 for those tasks. Those features are on the V8 roadmap but not yet shipped. Watch the updates.midjourney.com release page for the V8 upscaler and edit model announcements.