What Midjourney V7 Omni Reference carried between scenes
A KINTO Technologies designer used an existing image of a pink-haired virtual presenter to generate a new portrait, then reused the selected portrait in other scenes. The designer found a balance around --ow 200–400 for that character. Higher settings increased resemblance in the illustrated tests, while also holding on to more of the original rendering. The refreshed character appeared in a company introduction video and internal presentation materials. KINTO dates the write-up June 16, 2025, a day before V7 became Midjourney’s default. KINTO’s illustrated character workflow
When V7 users gained access
Midjourney began testing Omni Reference on the web and in Discord on May 1, 2025 at 23:03:26 UTC, already May 2 at 08:03:26 in Seoul. On June 17 at 04:46:33 UTC, 13:46:33 in Seoul, the company announced it had made V7 the default. That put an existing V7 feature on the default model path. The June announcement summarizes the subject-consistency capability without showing a paired example. Exact clock times come from the saved first-party publication records. The current Version article gives June 17 at calendar-date precision, as does the archived record of the default switch. Midjourney archive · May test announcement · June default-model announcement · Midjourney Version documentation
The V7 workflow takes one Omni Reference image and a text prompt. The prompt describes the new scene, and the image supplies a character, object, creature or vehicle to carry into it. On the web the image goes in the Omni Reference slot, while Discord uses --oref and an image URL. Current documentation gives --ow a range of 1–1,000 and a default of 100. It lists double the GPU time of a regular V7 job and restrictions around Draft Mode, Vary Region, Pan and Zoom Out. Those details describe the current documented workflow. Midjourney Omni Reference documentation
What carried through in actual work
The KINTO designer tried several weights on the same presenter brief. Settings around 100–200 sometimes produced an appealing rendering with weaker resemblance. At 800–1,000 the result more closely followed the source and the designer found it harder to get the treatment they wanted. After selecting a portrait, they used it to generate new compositions and backgrounds. Cropping a reference to the face gave the prompt more room to change clothing. Cropping closer to the eyes gave it more room to change hair. These were decisions for one character and one team’s design brief. KINTO weight examples and practical tips
By July 14, the maker of MARGHERITA had used Omni Reference to prepare a rear view from a front-view still of police characters. In a reply, the maker described keeping the reference weight at 25 or below for that shot and rerunning it to get the characters to face away from the camera. The completed short also used Midjourney animation and a separate finishing pipeline. The case documents a specific change of viewpoint and the extra attempts it took. Here it shares the source already used for the Video V1 analysis. MARGHERITA production thread and creator replies
On June 28, a Runway user asked whether Midjourney V7 could assemble two separately referenced people and a location. A reply pointed to Omni Reference, Style Reference and Image Prompts. In a later reply, the author said faces held up better than full bodies in their experience. The thread gives no matched inputs or settings for a comparison. The one-image limit also matters for a different reason. Multiple people can share one composite reference, but the operator cannot assign each of them an independent Omni Reference image in V7. June 28 comparison thread · Midjourney guidance for multiple characters
Where multi-image reference entered the product
The first V8.2 Edit Model test was announced at 23:32:15 UTC on August 27, 2026, or August 28 at 08:32:15 in Seoul. It offered up to four reference images in a separate edit and generation workflow. Current documentation points V8 users to that Edit Model and lists Omni Reference under V7. The V7 examples show how a designer reused a selected portrait and how a filmmaker iterated on a new view of a recurring character. Exact identity across camera angles, full-body details and separate subjects remained dependent on the particular generation. August 2026 Edit Model announcement · Current Edit Model documentation · KINTO’s production example