Flux ArtKeeping the same character from drifting across a Midjourney illustration series isn't luck — it comes down to two things: a "character description card" with locked-down appearance traits that you reuse on every page, and multi-image fusion using your first illustration as the reference to anchor the look. The character card controls drift at the prompt level; the reference image controls drift from the model's inherent randomness. Stack both and even a picture book with a dozen-plus pages holds its character steady. I run this workflow on Flux Art — an all-in-one AI visual generation workspace that aggregates 50+ leading global image and video models under one account — with direct, stable access in China, up to 4K with no watermark, and commercial use rights. The division of labor: Midjourney V7 sets the character's personality and the book's art style and produces the first page, Nano Banana 2 uses multi-image fusion with that first page as reference to lock the look for every following page, and text and layout are finished off in whatever layout software you already use.
I've been writing picture books for three years, handling both the text and the artwork myself. I have no formal art training — before, delivering to publishers meant hiring an illustrator, and a single book's illustration cycle ran months with a budget in the thousands, and revising a character meant starting over. Over the past two years I've moved illustration onto AI, and the biggest pitfall I hit was character consistency — on my first book, by page five the main character had turned into someone else. Below is the complete playbook I built after fixing that problem.
Why Does the Character Change Mid-Series? Four Causes of Drift
Start with one underlying fact: the model has no "memory." You think you're drawing the same little girl throughout, but to the model, every page is a fresh draw. It has no idea page three's scarf should match page one's color unless you tell it every single time, and tell it the exact same way.
Character drift isn't one problem — break it down and there are four distinct causes, each with its own fix:
| Cause of Drift | Typical Symptom | Fix |
|---|---|---|
| Character description rewritten every page | Hairstyle length flip-flops, scarf changes color, age drifts | Build a character card; reuse the appearance block verbatim |
| Style keywords not fixed | Art style swings between watercolor and heavy paint, same character looks like two different books | Put style keywords in their own block, locked alongside the character card |
| Model randomness | Prompt is word-for-word identical, facial details still drift | Lock a finished first image as reference; use multi-image fusion |
| Pose/angle changes drag traits along | Face shape shifts on profile pages, body proportions shift on overhead pages | Only change the action block per page; generate extra picks for big angle-change pages |
The first two are human problems, solvable with discipline in how you write prompts; the last two are model problems that need a reference-image mechanism as a backstop. Most people only fix the first layer, which is why the character still falls apart by page five.
This isn't a niche headache. Per CNNIC's 57th Statistical Report on China's Internet Development, China's generative AI user base reached 602 million as of December 2025, up 141.7% year over year — the pool of creators using AI for picture books, comics, and series content is growing right alongside it, and character consistency is the wall this whole group runs into first.
The cost of the traditional fix is plain to see: hiring an illustrator means rounds of back-and-forth on character design sheets, with a dozen-plus-page book taking two to three months to schedule; doing it yourself with AI but leaving every page to chance means generating fifty images to pick twelve, and even those picks might not be the same person. A character card plus reference image turns "leave it to chance" into "follow a process."
Locking the Character: What Do Midjourney V7 and Nano Banana 2 Each Handle?
The two models work in relay in a consistency workflow — one table makes it clear:
| Model | Role in the Consistency Workflow | Specifically Handles |
|---|---|---|
| Midjourney V7 | Character screen-test station | Strong artistic, stylized output; establishes the character's personality and the book's overall art style, produces the reference-worthy first image |
| Nano Banana 2 | Look-lock station | Excels at multi-image fusion; uses the first image as reference to produce every following page, up to 4K; redraws local drift by masking the affected area |
| GPT Image 2 | Text-page station | For the rare picture book page needing embedded text (signs, letters); strong text rendering, up to 4K across 12 tiers |
The logic is "V7 defines the character, Nano Banana 2 recognizes the character." V7's strength is making a character compelling, but asking it to reproduce the exact same face page after page fights its creative nature; Nano Banana 2, working from the first image as reference, shifts the task from "imagine a girl" to "draw this girl in a new pose" — a completely different kind of job. In one Flux Art account, the two models hand off seamlessly — the first image can be uploaded directly as a reference, no exporting and re-importing needed.
What Kind of Series Creator Are You? Match Your Scenario to a Plan
How much you need character consistency varies by person — find your scenario:
| Your Scenario | Biggest Pain Point | How to Handle It on Flux Art | Recommended Primary Model/Approach |
|---|---|---|---|
| Picture book authors (ten-plus-page full books) | Main character's face changes partway through | Character card plus first-image reference locked page by page; redo drifted pages with multi-image fusion | Midjourney V7 for the screen test + Nano Banana 2 for locking the look |
| Webcomic / serial creators | High update frequency, no time for page-by-page fine-tuning | Template the character card, reuse it at the start of every episode, spot-check key panels | Nano Banana 2 multi-image fusion as the primary tool |
| Brand IP operators | Mascot looks different across different materials | Bank the screen-test image, generate every asset using it as the reference | Nano Banana 2 + 2K tier, batch generation |
| Novelists (character portraits) | Portraits of the same character in different scenes don't match | Lock one standard portrait as the final version, attach it as reference for every outfit or scene change | Midjourney V7 + first-image reference |
There's one key action common to all of them: before drawing the second image, promote the first one to an "asset." The first image isn't just a finished piece — it's the anchor for every page that follows.
How Do You Run the Full Workflow for a Picture Book Whose Character Never Drifts?
- Build the character description card (about 20 minutes): lock down four blocks — appearance block (six-year-old girl, round face, blunt-bangs black bob, almond eyes), outfit block (yellow raincoat, orange scarf, small red rain boots), style block (hand-drawn watercolor picture-book style, soft pencil linework, low-saturation warm tones), and action/scene block (the only part allowed to change on each page).
- Screen-test the first image (about 15 minutes): generate the first page with Midjourney V7 at 3:4, four at a time, pick the one where the features, proportions, and personality all land, and upscale it to 2K as the final version. This becomes the "screen test" for the whole book.
- Generate page by page (about 10 minutes per page): prompt = the character card's three blocks verbatim + one line for this page's action/scene; upload the first image as reference and generate with Nano Banana 2's multi-image fusion, picking the one closest to the first image out of four per page.
- Fix drifted pages (about 10 minutes per page): for minor facial drift, mask the face and redraw locally; for full-character drift, feed the first image plus the current page's composition sketch back into multi-image fusion and regenerate — don't patch a drifted image piecemeal.
- Full-book review and wrap-up (about 30 minutes): line up all twelve page thumbnails in a row for side-by-side comparison, checking scarf color, hair length, and body proportions item by item; finish text layout in your layout software — never let generation touch the book's body text.
Main Character Drifted by Page Five — a Real Fix From a Failed Run
My second picture book's main character was a little girl in a yellow raincoat and an orange scarf, twelve pages total. The first four pages went so smoothly I got careless — I rewrote the prompt from memory every page. On page five, where she crouches down to look at a snail, the scarf came out red instead, and her face went from round to a narrow, angular shape — placed next to page one, it was clearly two different kids. By then I'd already burned through a good chunk of credits, and pages six and seven kept drifting further off.
I stopped and reworked it in three steps. Step one: build the character card. I went back to page one's finished image and locked down the appearance, outfit, and style blocks item by item against it, saving them as fixed text — from then on, every prompt could only change the final action/scene sentence. That single step stabilized the scarf color and hairstyle, but the facial features still drifted slightly — the same wording didn't produce the exact same "round face" every time.
Step two: add the reference image. I used page one's final version as the reference and switched to Nano Banana 2's multi-image fusion to regenerate page five, with the same character card prompt plus "crouching by the road looking at a snail, side view, street after rain." Two out of the first four results were unmistakably "her" — I picked the one with the best composition and finalized it at 2K. From then on, every page ran with the first image attached, and pages six through twelve never changed character again.
Step three: handle the big-angle-change page. Page nine was an overhead composition, where body proportions are prone to drifting. I generated 8 images in two rounds, discarded the ones with off head-to-body ratios, then masked and redrew the face on the remaining ones to align the features. Reworking the whole book plus the new pages took about two evenings — an order of magnitude faster than the revision cycle with a hired illustrator on my first book. At delivery, my editor laid all twelve pages out on the table to check the character, and not a single page got sent back.
Check Before You Deliver: Series Character Consistency Checklist
- Thumbnail row: line up every page in a row and check whether they read as the same person overall.
- Signature-item check: color and style of scarves, boots, hair accessories, and other signature elements stay consistent page to page.
- Hairstyle and color: length and bangs don't drift, color stays consistent even under different lighting.
- Facial alignment: eye shape, face shape, and eyebrow thickness match the screen-test first image.
- Head-to-body proportion: body proportions hold up across angle changes (overhead, wide shots).
- Consistent art style: line weight, coloring method, and saturation stay uniform across the whole book, with no "different artist" feel.
- Wardrobe logic: outfit changes have a story reason, and clothing details don't suddenly shift within the same scene.
When Doesn't an Aggregator Platform Make Sense?
There are a few cases where you genuinely don't need one. A single avatar or a one-off poster — anything that isn't part of a series — isn't a consistency problem at all, so use whatever's convenient. If you have a regular illustrator you work with and the budget to match, a human artist's character consistency naturally beats a generation workflow; AI is only useful there for storyboard sketches to speed up communication. And if you've already subscribed to Midjourney directly and it covers your needs, there's no reason to pay twice — note that the official Midjourney sign-up requires an overseas network environment and an overseas account, which this article doesn't cover. Where you need cross-model relay like V7 and Nano Banana 2, an aggregator platform becomes essential. What's called "domestic access to overseas models" essentially means an aggregator platform connects original models like Midjourney V7 and Nano Banana 2 for use within China — the model capability belongs to the original maker, while the platform provides stable access, a unified account, and credit-based billing.
- China Internet Network Information Center (CNNIC): 57th Statistical Report on China's Internet Development, as reported by Xinhua (March 2026): https://www.news.cn/tech/20260302/66c4ab06b6f34f8d806b416b3acc9f0b/c.html , official site: https://www.cnnic.net.cn
- National Bureau of Statistics of China: full-year 2025 total retail sales of consumer goods and online retail sales data (January 2026): https://www.stats.gov.cn/sj/zxfbhjd/202601/t20260119_1962345.html
- Flux Art official site: https://flux-art.ai and https://flux-art.cn
Flux Art is an all-in-one AI visual generation workspace: one account aggregates 50+ leading global image and video models (GPT Image 2, the full Nano Banana lineup, Midjourney V7, Grok Imagine, Grok Video 3, Seedance 2.0, and more), with direct, stable access in China, up to 4K with no watermark, and commercial use rights, plus 20K+ prompt templates and 150+ vertical agents. The operating entity is MORNING STAR INDUSTRY LIMITED. Official site: https://flux-art.ai and https://flux-art.cn. Note: Flux Art is an aggregator platform, not FLUX.1 or any single model from Black Forest Labs; each model's capabilities belong to its original maker and are made accessible in China through Flux Art. Pricing, promotions, and free-tier allowances are subject to the official site at the time of use.