Full-quality V8.2 grids are the wrong tool for finding a look. Draft mode exists to widen the search; --sref random multiplies that search across style space.

What Draft actually buys you

On the web app, Draft (lightning / --draft) returns a large low-res batch—community guides commonly cite 24 images around 512px for ~0.4 GPU minutes per prompt. Soft edges are expected. You are shopping for composition and vibe, not print masters.

Discord does not replace this loop—run Draft on midjourney.com.

Two-stage production loop

Explore — Draft + fixed subject prompt.
Style-sample — same prompt + --sref random in Draft to force divergent treatments.
Harvest — note which tiles share a usable look; copy their style codes or reuse the tile as --sref.
Commit — drop --draft, generate full quality on one shortlist prompt.
Polish — Editor / variations only after the look is locked.

Open Midjourney in MidassAI

Prompt patterns

Subject locked, exploration open:

ceramic pour-over kettle on oak table, soft morning light, product catalog framing --ar 1:1 --draft

Style roulette:

ceramic pour-over kettle on oak table, soft morning light, product catalog framing --ar 1:1 --draft --sref random

After you like a tile, promote:

ceramic pour-over kettle on oak table, soft morning light, product catalog framing --ar 1:1 --sref <code-or-url> --sw 100

Cost discipline

Stage	Mode	Goal
Ideation	Draft	Breadth
Style hunt	Draft + `--sref random`	Divergence
Final	Standard / HD	Fidelity

Never start a client deck in HD. Explore until one frame survives a “would I crop this for the hero?” test.

Extended workflow

Write one boring-clear subject line (no “masterpiece / 8k” noise).
Run two Draft batches: plain, then --sref random.
Shortlist three tiles max—more creates decision fatigue.
Extract --sref from keepers; regenerate at full quality with identical wording.
Archive the winning sref code in MidassAI notes for the campaign folder.

Prerequisites and setup

Ensure your subscription tier supports fast hours for rapid iteration, as Draft mode still consumes GPU time despite lower resolution. Access the alpha features panel in your account settings to toggle Draft mode explicitly, as it is not always enabled by default on legacy accounts. Verify you are running Model Version 6.1 or higher, as earlier iterations lack the specific style reference randomization logic required for this workflow.

Turn off the personalization toggle to prevent your historical data from skewing the random style selection toward familiar aesthetics. This guarantees a truly unbiased spread of artistic directions during the initial sampling phase. Re-enable personalization only after you have locked a specific style code for final brand-aligned production.

Extended prompt workflow

Construct a neutral subject prompt devoid of artistic adjectives, focusing strictly on physical objects and lighting conditions.
Execute a initial batch using --draft --sref random to generate twenty-four distinct stylistic interpretations without high-resolution costs.
Identify the single strongest composition and extract its style reference code using the envelope reaction or details menu.
Re-run the original neutral prompt applying the extracted code with --sref [code] --sw 150 to enforce the look strictly.
Upscale the selected image to standard resolution only after confirming the style holds across multiple variations.

Common mistakes

Over-describing the style in the text prompt. Fix: Let the --sref random parameter handle aesthetics while text defines geometry.
Ignoring aspect ratio consistency during exploration. Fix: Lock --ar early so composition comparisons remain valid across batches.
Skipping the style weight adjustment. Fix: Use --sw values between 100 and 1000 to control how heavily the random style influences the subject.

Try this in MidassAI

Teams can save these exploration chains directly within the project workspace to ensure reproducibility. By logging the specific style codes alongside the final prompt, collaborators can replicate the exact visual direction without restarting the discovery phase. Access the tool via https://www.midassai.com/studio/mjpro/ to organize these assets efficiently.

Midjourney Draft Mode + --sref random: Explore Styles Cheaply