13:27
Midjourney
The video argues that reliable character consistency in Midjourney is achievable mainly through the Omni Reference feature combined with prompting for a unique, distinguishing look, and that a handful of secondary tricks (clothing-reference reuse, combining two characters into one reference image, switching aspect ratios, and pairing rough Midjourney character sheets with ChatGPT and Krea/Flux) extend that consistency across scenes, outfits, and even multiple characters, despite Midjourney currently supporting only one Omni Reference image at a time.
Character consistency is achieved via Midjourney's Omni Reference feature: drag a reference image directly into the prompt bar (no download/upload needed), write the new prompt, and set other settings to get a character that looks similar to the reference.
Prompt for a unique, distinguishing look (e.g. pink hair, specific clothing colors, body shape, haircut) so the character stands out and stays recognizable across generations.
For full-body shots, explicitly describe items like shoes and add words like 'standing' in the prompt, since Midjourney may otherwise omit footwear or seat the character.
Reuse the original character's Omni Reference image while prompting for a new outfit to get clothing consistency across a story's different looks and scenes, rather than prompting for outfits from scratch each time.
Midjourney currently only allows one Omni Reference image, so two characters can be combined into a single reference image as a workaround to put both characters in one scene; the video calls this 'not the ideal way' but says 'it does work.'
When combining two characters, you still must describe each character's distinguishing qualities in the prompt (e.g. orange sweater vs. leather shoes vs. white sneakers), and some detail 'spillage' between characters occurs (e.g. she also gets brown leather shoes).
The same combined-reference trick can be used to get clothing consistency for a specific character within a two-character image (e.g. designing a wedding dress) even though results aren't 100% matched (the skirt ends up longer, dragging on the ground).
Odd/non-human characters (e.g. a cat) can cause environmental or anatomical errors (posed like a human, duplicated objects); trying different aspect ratios can fix wrong elements, broken composition, and body-shape/proportion issues.
Photoshop's Generative Fill can be used after generation to extend a scene's composition and remove unwanted stray elements, such as a floating cup.
Adobe has moved Generative Fill from free testing to a paid model: 250 free generations/month, then paid tiers described as pricing that is 'kind of strange, kind of aggressive' — e.g. Firefly Standard at €11/month allows only 25 seconds of video (~55 cents/video) versus roughly 5 cents per image generation.
Character design sheets (rough sketches) generated directly in Midjourney can serve as a beautiful, interesting artistic starting point even when not fully detailed; a 3D or Octane render can be requested afterward for more detail.
Cropping a character out of a design sheet and uploading it to ChatGPT to get a Midjourney-style description is presented as a faster alternative to writing the prompt description manually.
Sending the cropped sketch/description through Krea (using the Flux model) produces a more detailed version of the character, described as 'not 100% the same character' but close enough to bring back into Midjourney for further variation.
Midjourney is shown sticking closely to fine details of an input reference image across entirely new scenes and perspectives (e.g. a staff's round top with two horns and hanging bells, leather necklaces, sword placement).
The overall workflow -- character, Omni Reference, and clothing/outfit variants -- is framed as enabling a short film or graphic novel built from consistent characters and scenes.
Omni Reference — Midjourney's character-reference feature, used by dragging a reference image directly into the prompt area, that keeps a character's appearance consistent across new generations. Apply: Drag the character image into Omni Reference, write the new prompt and other settings, and generate a scene where the character resembles the reference.
Unique-look prompting — A tip to give a character a distinguishing, describable trait (e.g. pink hair, specific clothing colors, body shape, haircut) so Midjourney can reliably reproduce it. Apply: Write a prompt calling out one or two standout features (e.g. 'pink hair, black leather jacket, blue jeans, white sneakers') before using the image as an Omni Reference.
Full-body shoe/standing prompting — A trick for full-body shots: explicitly describing shoes and adding 'standing' to the prompt ensures those elements appear, since Midjourney may otherwise seat the character or omit footwear. Apply: Add 'standing' and a shoe description to the prompt when a full-body shot with visible footwear is needed.
Clothing-change via Omni Reference — Reusing the original character's Omni Reference image while prompting for a new outfit, so redesigned clothing stays visually consistent with the established character across a story. Apply: Feed the original character image back into Omni Reference and prompt for the new outfit (e.g. a wedding dress) instead of designing it from scratch each time.
Two-character combination trick — A workaround for Midjourney only supporting one Omni Reference image: combining two characters into a single reference image so both can appear together, described as 'not the ideal way' but functional. Apply: Create one input image containing both characters, describe each character's distinguishing qualities in the prompt, and expect some detail spillage between them.
Aspect-ratio switching — Trying different aspect ratios to get different Midjourney outputs, said to be able to fix wrong elements, broken composition, and body-shape/proportion issues. Apply: When a render has an environmental or anatomical problem, regenerate at a different aspect ratio to get a cleaner result.
Photoshop Generative Fill cleanup — Using Adobe Photoshop's Generative Fill (Gen Fill) to extend a scene's composition and remove unwanted stray elements after Midjourney generation. Apply: After generating the base image, use Generative Fill to extend the canvas and paint out mistakes such as a duplicated or hovering object.
Character design sheets — Generating rough, sketch-style character design sheets directly in Midjourney as an artistic starting point, which can look 'beautiful' even though not fully detailed. Apply: Prompt Midjourney for a character design sheet, then optionally follow up with a prompt for a 3D or Octane render for more detail.
ChatGPT-assisted prompt description — Cropping a generated character image down to just the character and uploading it to ChatGPT to get a Midjourney-style prompt description of that character. Apply: Crop the character out of the design sheet, upload it to ChatGPT, and ask for a Midjourney-style description to reuse in further generations.
Krea + Flux re-rendering — Running a rough Midjourney character sketch through Krea (using the Flux model) to produce a more detailed, somewhat varied version of the character. Apply: Send the cropped sketch/description into Krea with Flux to generate a more detailed render, then bring the result back into Midjourney for further variations.
The two-character combination trick is an unofficial workaround for a hard platform limitation (one Omni Reference image only), not a supported feature, and it visibly leaks details between the two characters rather than cleanly isolating them.
Aspect ratio switching is presented as a general-purpose fix lever unrelated to prompt content itself -- capable of correcting proportions, composition, and even wrong elements, essentially by re-rolling the generation space rather than editing the prompt.
Rough Midjourney character design sheets are valued as their own aesthetic ('beautiful interesting artistic design'), not just as an intermediate step toward a fully rendered result -- the imperfection of the sketch is treated as a feature worth keeping as a starting point.
The Krea/Flux re-render pass is explicitly not meant to preserve 100% fidelity to the sketch; the resulting variation (e.g. added axolotl-like details) is treated as a usable creative bonus rather than a defect to eliminate.
Adobe's Generative Fill pricing is framed as a case study in inconsistent per-use AI costs: per-image pricing (~5 cents) is judged reasonable, while per-second video pricing (~55 cents/video under the €11 Standard plan) is judged excessive for what is described as sometimes just 'a little detail' edit.
Midjourney's consistency isn't just broad character resemblance -- the video highlights it replicating specific ornamental minutiae (an exact staff shape, a specific number of horns, hanging bell shapes, leather strap placement, where a sword is stuck to the body) across different scenes and camera angles.
«Wow, this works way better than expected.»
— 00:00
«When you prompt for a character I would suggest to you that you prompt for a unique look for that character.»
— 01:15
«it's not the ideal method, but it does work and you can get some really great results from that.»
— 04:53
«However, the pricing is kind of strange, kind of aggressive maybe in some parts.»
— 07:40
«this is a very powerful tool for artists and indie filmmakers.»
— 08:52
«It's not 100% the same character, but that's okay because I can play around.»
— 10:15
«look on how really close Midun is sticking to the details of the input image.»
— 11:20
Reception
Viewers appreciate the topic and find value, but criticize the video for lacking step-by-step walkthrough detail and specificity in explanations.
A practical, tip-forward walkthrough of Midjourney's Omni Reference workflow that offers concrete, testable tricks (unique-look prompting, clothing-reference reuse, aspect-ratio switching, the two-character workaround) while candidly noting where those workarounds fall short of perfect fidelity.
13:27