Flux ArtWhen a Midjourney character's hand comes out mangled, you don't need to scrap the whole image and start over. Four steps, in order, will fix it: first "lighten the load" on the hand in your prompt (specify one hand, a natural pose, and where the other hand goes), then batch-generate 4 images and cull the bad ones, pick the best composition and inpaint just the hand, and if the same composition keeps breaking, switch models as a fallback. I run this entire workflow on Flux Art — an all-in-one AI visual generation workspace that aggregates 50+ top global image and video models under one account — with direct, stable access with no extra network setup, up to 4K with no watermark, and commercial use allowed. The division of labor is simple too: Midjourney V7 handles the first-draft mood and illustration style, Nano Banana 2 handles inpainting the hand and full-image fallback regeneration, and text layout before delivery is finished off in whatever software you're already comfortable with.
I've been taking portrait illustration commissions for four years — mostly social media avatars, couple keepsake illustrations, and character portraits for novels. When clients review a piece, the part they scrutinize most is, first, the face, and second, the hands. A broken face gets caught immediately, but a broken hand often only surfaces after delivery, called out in a comment on the client's social feed — and that's the kind of rework that stings the most. Over four years I've turned the patterns of hand failures and how to fix them into a fixed routine, and this article lays out all of it.
Why Does AI Keep Getting Hands Wrong? Four Types of Hand Distortion
Let's start with the root cause. A single hand has more than twenty movable joints, and the same hand can appear in near-infinite poses in a photo, often half-hidden by a sleeve, a prop, or the other hand. What a model learns about "hands" from its massive training set is a pile of incomplete, highly variable, mutually contradictory samples — and when we write prompts, we almost never describe the hands at all, so the model is left guessing based on probability. Faces don't have this problem, because facial structure is stable and well-represented in training data — a front-facing face is just a front-facing face.
Worth emphasizing: hand distortion is a shared challenge across all image generation models, not a Midjourney-specific flaw. V7 is widely recognized as strong on character mood and illustration style, and it's what I use to start most of my stylized commissions — but hands remain a high-risk area regardless. The difference is only in how often it breaks and in what way.
Based on four years of rework logs, broken hands fall into four categories, and each one calls for a different first move. Working through the table below saves far more time than blindly regenerating:
| Distortion Type | Typical Symptom | High-Risk Scenario | First Fix |
|---|---|---|---|
| Wrong count | Six fingers, four fingers, or an extra hand out of nowhere | Hands overlapping, multiple people in frame | State "one hand" in the prompt and specify where the other hand goes |
| Fused structure | Knuckles bunched together, adjacent fingers merged | Fine motor actions like gripping, pinching, clasping | Simplify the pose: switch to relaxed hanging, resting flat, or cupping from below |
| Hand-object clipping | Fingers passing through a cup handle, a phone "grown" into the palm | Hands holding props | Use action words implying larger contact surfaces, or upload a reference image of the prop |
| Proportion and orientation | Hand too large or too small, wrist twisted the wrong way, awkward pointing | Close-up hand shots, extreme perspective compositions | Pull back the shot to shrink the hand's visual weight, or inpaint the hand separately |
Tolerance for hand flaws has clearly dropped in the past couple of years. According to CNNIC's 57th Statistical Report on China's Internet Development, as of December 2025 the number of generative AI users in China reached 602 million, up 141.7% from December 2024. Clients themselves are heavy users now — they've seen good output, and a six-fingered image no longer slides by.
Now for the pain of the old-school fix. Broken hands used to leave only two options: go into photo editing software and liquify/clone-stamp your way through it by hand, or sketch out a new hand from scratch and blend it in — either way, rework on a single image was billed by the hour, and the math never worked out against commission rates. The current approach solves the problem at the generation stage instead: rewrite the description, generate more variants, inpaint locally — a full round takes minutes and costs credits, not hours.
Fixing Hands: What Do Midjourney V7, Nano Banana 2, and GPT Image 2 Each Handle?
The three models aren't competing in the hand-fixing workflow — they're three stations on an assembly line:
| Model | Role in the Hand-Fixing Workflow | Specifically Handles |
|---|---|---|
| Midjourney V7 | First-draft station | Produces the first version of character mood and illustration style — sets the tone for style, lighting, and emotion |
| Nano Banana 2 | Refinement and fallback station | Inpaints the framed hand region; when the same composition keeps failing, regenerates the whole image using the draft as a reference, up to 4K |
| GPT Image 2 | Complex-instruction station | When a hand gesture has a specific requirement (a particular gesture, holding a specific object in a specific pose), it works through the instruction directly — 12 precision-and-resolution combinations, up to 4K |
The logic behind this division of labor is playing to each model's strengths. V7's strength is style and mood — making it regenerate repeatedly hoping for a good hand burns through credits fast without guaranteeing a win. Nano Banana 2's inpainting only touches the framed selection, so the original composition, lighting, and face all stay intact. GPT Image 2 follows text instructions more faithfully, so a specific ask like "left hand making an OK sign" has a higher success rate when handed to it. On Flux Art, all three models share one account and one credit pool — switching between them is just a dropdown menu, no re-uploading images back and forth.
What Kind of Portrait Creator Are You? Match Yourself to a Plan
Broken hands hurt different people in different ways. Find yourself below and copy the approach:
| Your Scenario | Most Painful Step | How to Do It on Flux Art | Recommended Primary Model / Approach |
|---|---|---|---|
| Commissioned illustrator (avatars, character portraits) | Client zooms in on hands during review | Generate 4 low-tier drafts with V7 and cull, upscale the finalist to 2K, inpaint the hand separately | Midjourney V7 + Nano Banana 2 inpainting |
| Content creator (article illustrations) | No time to retouch every image | Avoid hand close-ups in the composition entirely; keep hands out of frame or tucked in pockets for half-body shots | Midjourney V7, batch generation at a low tier |
| E-commerce operator (hands holding a product) | Hand clips through the product | Upload a white-background product photo as reference; describe the pose as "supporting from below, palm flush against the product" | Nano Banana 2 multi-image fusion |
| Small studio (group posters) | Wrong hand count when multiple people share a frame | Generate each person separately and composite afterward; avoid hand-holding or overlapping shoulder poses | GPT Image 2 + post-production compositing |
One rule applies across all four rows: whatever you can avoid at the prompt stage, don't leave for the retouching stage. When hands aren't the selling point of the image, letting them fall outside the frame is a zero-cost fix.
What's the Full Workflow for a Portrait Whose Hands Hold Up to a Zoom-In?
- Write a prompt that "lightens the load" on the hand (about 5 minutes): beyond describing the subject, give the hand its own dedicated sentence — "one hand naturally holding the cup, the other hand resting on the table edge." Pick the aspect ratio for your use case (3:4 for avatars, 1:1 for covers), and test the composition at a low resolution tier first to save credits.
- Batch-generate and cull (about 10 minutes): generate 4 at a time, and on the first pass only look at the hands, not the face — count fingers, check the knuckles, check the contact surfaces. If the hand fails, cull it immediately; don't get attached to a good composition.
- Upscale and regenerate the finalist (about 5 minutes): take the prompt that passed the cull, upscale it to the 2K tier unchanged and regenerate a round, then pick the best composition and mood among the qualifying results as your finalist candidate.
- Inpaint the hand (about 10 minutes): if the finalist candidate still has minor hand flaws, frame the hand plus a small margin around it, and write a prompt describing only that hand — "a hand naturally holding a ceramic cup, five fingers clearly defined, distinct knuckles" — leaving the rest of the image untouched.
- Switch models as a fallback + zoom in for a final check (about 10 minutes): if two rounds of inpainting still don't work, switch to Nano Banana 2 and regenerate the full image using the finalist candidate as a reference; finally, zoom the image to its original delivery size and check every single finger before handing it off.
A Portrait Holding a Coffee Cup Had Broken Fingers — Here's How I Actually Fixed It
Last month I had a job: the client wanted a warm-toned illustrated avatar of "a woman by a café window, cupping a mug," for social media. I used Midjourney V7 at a 3:4 aspect ratio, low tier, generating 4 images at once. The first batch failed across the board: one had six fingers on the left hand, one had both hands' knuckles fused into a blob, one had fingers passing straight through the cup handle, and only one had a barely acceptable hand but a mediocre composition.
Step one: rewrite the pose description. "Both hands cupping the mug" is a classic high-risk pose — two hands, ten fingers, and interaction with an object on top of that. I changed it to "one hand naturally holding the cup, the cup body resting against the palm, the other hand resting lightly on the table edge," cutting the hands' workload in half. I regenerated 4 images, and two came out with completely normal hands.
Step two: upscale to a finalist. I picked the one with the best window light, kept the prompt unchanged, and upscaled it to the 2K tier. The finalist candidate's hand shape was correct, but the second knuckle of the pinky still had a faint trace of fusion.
Step three: inpaint the hand. I framed the whole hand plus the lower half of the cup, and wrote a prompt describing only "a hand naturally holding a white ceramic cup, five fingers clearly defined, distinct knuckles, warm window light." It worked on the first try — the skin tone and lighting transition at the seam showed no visible sign of retouching.
If step three hadn't worked on this job, my fallback plan was to switch to Nano Banana 2 and regenerate the full image using the finalist as a reference — that's exactly how I bailed out another job two weeks earlier, a couple's portrait with clasped fingers, where V7 failed three rounds in a row and switching models produced two usable images on the very first round. The whole process took about 40 minutes, and the client approved it on the first review, zooming in specifically to check the hands.
Check This Before Delivery: A Portrait Hand Checklist
- Count the fingers: count each finger on every hand individually — any hidden fingers should still make anatomical sense.
- Count the hands: the total number of hands in frame should match the number of people — watch for an extra hand appearing out of nowhere in a corner.
- Check the knuckles: no fusion, no joints bending backward, no fingers that thin out and then thicken again mid-length.
- Check the contact surfaces: a hand holding an object shouldn't clip through it — the contact point should show natural pressure and shadow.
- Check the proportions: the size relationship between hand and face should match the character — an adult's hand is roughly the length from chin to hairline.
- Check for retouching seams: at the edges of an inpainted region, skin tone, lighting, and brushstroke style should match the surrounding area.
- Zoom in for a final check: view the image at full delivery resolution, zoomed to its original size — thumbnails can be deceiving.
When Does an Aggregator Platform Not Make Sense?
A few honest notes. If your creative style always avoids showing hands — profile pictures, chibi style, abstract art — broken hands simply aren't a problem for you, and you don't need to add any tool for this. If you're already subscribed directly to Midjourney and your quota fits your needs, there's no reason to pay twice; direct Midjourney access requires an overseas network environment and an overseas account, which this article won't get into. For illustrators with strong hand-drawing skills, fixing a hand directly in your tablet software might be faster than regenerating — generation-side fixes are just one more option, not a replacement. The straightforward path is: access via Flux Art's aggregation, sign up on the web and start immediately, pay by credits, full-power access with no queueing. What's often called a "local access point for overseas models" really just means an aggregator platform connects original models like Midjourney V7 and Nano Banana 2 for use with stable access — the model capability itself belongs to the original developer, and 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, Xinhua News Agency report (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 data on total retail sales of consumer goods and online retail sales (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+ top 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 with no extra network setup, up to 4K with no watermark, and commercial use allowed, plus 20K+ prompt templates and 150+ specialized agents. It's operated by MORNING STAR INDUSTRY LIMITED. Official site: https://flux-art.ai and https://flux-art.cn. Note: Flux Art is an aggregator platform, not Black Forest Labs' FLUX.1 or any single model in itself; each model's capabilities belong to its original developer, made accessible via Flux Art. Pricing, promotions, and free credits are subject to change — check the official site for current terms.