When the client brand board lands with logo lockups, talent faces, packaging angles, and a hero still already approved, and the fifth asset—the seasonal product SKU that must share the same wardrobe light and shelf world—arrives after you have already spent Midjourney Edit Model’s four reference slots on character, product, environment, and type treatment, the debate over which model looks prettier stops mattering.

You are out of reference budget.

Small and mid creative agencies already run Midjourney and Flux on the same desk. The Slack ping is rarely “which model is more beautiful.” It is “we need six Meta variants, three Stories crops, and a carousel that still reads as the same talent and pack.” The board lives in Figma or a PDF deck for the client walkthrough. That deck is presentation. It is not the system of record for locked brand refs. Every generator only sees what you attach as multi-reference input. The fifth asset that does not fit the fourth Midjourney ref slot turns Midjourney vs Flux into a reference-budget fight.

Midjourney’s Version docs state that V8.2 has been the default since 2026-07-24, and that Edit Model replaces Omni Reference, Character Reference, and Retexture (https://docs.midjourney.com/hc/en-us/articles/32199405667853-Version). The Edit Model article is explicit: written-instruction edits, generate new images using up to four reference images, and an Editor that supports inpaint plus outpaint (https://docs.midjourney.com/hc/en-us/articles/48495453462797-Edit-Model).

For an agency brand kit, that four-reference ceiling is the real constraint. Character refs, pack shot, shelf world, and a type or logo plate often fill the slots before the seasonal SKU or a second talent arrives. Prompt adherence can still look strong inside those four. Batching across formats does not invent a fifth slot. Treat Midjourney as the beauty winner without mapping the kit to Edit Model’s reference ceiling, and the team regenerates continuity instead of shipping variants.

Black Forest Labs’ FLUX.2 overview frames multi-reference as a tiered budget, not a single number (https://docs.bfl.ml/flux_2/flux2_overview). FLUX.2 [klein] sits at up to four references—the same rough ceiling agencies already feel on Midjourney Edit Model. [max], [pro], and [flex] move to up to eight references via the API and up to ten in the playground. [dev] is recommended at a maximum of six. The same overview speaks in terms of up to ten sources in framing language. That spread matters when a client kit carries talent, product family, environment, typography, and campaign props in one pass.

FLUX.2 [flex] is specialized for typography and text rendering and for small details, covering text-to-image and image-to-image paths (https://docs.bfl.ai/api-reference/models/generate-or-edit-an-image-with-flux2-%5Bflex%5D). Agencies that live on pack shots with readable labels and on-image claims often need that flex lane more than another beauty A/B. Multi-reference headroom without typography control still burns hours on illegible type.

The operational question is whether this week’s kit shape fits Edit Model’s four-reference ceiling or needs FLUX.2’s higher multi-reference tiers. A two-ref beauty still—face plus pack—can live inside Midjourney Edit Model with written-instruction edits and Editor inpaint or outpaint for crop fixes. A six-ref retail world—talent, two SKUs, shelf, logo plate, seasonal prop—pushes toward FLUX.2 [pro], [max], or [flex] on API budgets that allow eight, or playground paths that allow ten, with [dev] around six.

Match the tool to the kit shape before you score tiles. Prompt adherence on a single hero does not predict batching across a brand kit. Character refs that drift after the fourth attachment do not get fixed by a prettier model name. The reference ceiling is the production constraint. Beauty is the demo metric.

Agencies burn afternoons exporting the same brief into Midjourney and Flux, then picking the sharper tile for the client deck. The deck updates. The locked refs do not. Figma frames and PDF boards stay presentation surfaces. Drive folders named final_v4 still hold unversioned PNGs with no character identity, no universe, and no reusable edit history.

A beauty-contest A/B answers which generator won one still. It does not version the brand kit or keep character refs aligned for multi-format crops. It does not close Idea to Create to Produce to Publish to Measure to Improve when performance notes arrive and the same talent must reappear. Without a kit system of record, every Midjourney vs Flux round restarts the reference budget from zero.

Havincy is built for that system of record: persistent characters and universes, image generation plus editing and retouch, multi-format production, Director for directing the work, and social publishing with analytics so the loop can run Idea → Create → Produce → Publish → Measure → Improve. Keep the brand kit—character refs, universe, approved stills, edit stack—outside the presentation deck and inside production, so Midjourney Edit Model’s four-reference ceiling and FLUX.2’s multi-reference tiers become routing decisions instead of weekly rediscovery.

When the fifth asset arrives, you already know whether this kit needs four refs or eight. When the client asks for another Stories crop, you edit against locked characters and universes instead of reattaching orphan PNGs. When Measure shows which variant held, Improve feeds the next Create pass without rebuilding the kit from a Figma export.

No. Past four references, FLUX.2 [max], [pro], and [flex] offer higher multi-reference budgets on the API and playground paths documented by Black Forest Labs, while [klein] stays near four and [dev] is recommended around six. That headroom helps when the kit needs more simultaneous sources. It does not automatically beat Midjourney Edit Model on every crowded brief. If the brief can be sequenced—lock character and pack first, then edit in the seasonal SKU with written instructions and Editor inpaint—four Midjourney refs can still ship. Flux is stronger when the kit requires concurrent multi-reference, especially when [flex] typography sits inside the same pass.

Edit Model is enough when the brand kit fits four reference images, when written-instruction edits cover the revision language, and when inpaint or outpaint in the Editor can finish crops without a new identity. Many social packs and ad variants live here: one talent, one hero pack, one environment, one type treatment. Midjourney V8.2 as default since 2026-07-24, with Edit Model replacing Omni Reference, Character Reference, and Retexture, is a coherent stack for that shape. It stops being enough when the fifth locked asset must sit in the same generate pass as the first four, or when typography-heavy pack work needs FLUX.2 [flex]’s text-rendering lane.

More Midjourney or Flux seats raise parallel generation. They do not raise the reference ceiling on a single generate call, and they do not version characters or universes. If the kit lives only in Figma boards and unlabelled Drive dumps, extra seats multiply unversioned beauty contests. Fix the kit system of record first—persistent characters, universes, edit stack, multi-format outputs, then Publish and Measure—so each seat spends reference budget against the same locked board instead of inventing a new one.