Ran a controlled test on a problematic image: a portrait with a hand awkwardly placed. Needed to remove the hand and generate plausible background.
**Test Image:** 512x768 portrait, medium detail.
**Goal:** Remove subject's right hand from chest, fill area with consistent jacket fabric.
**Midjourney Vary (Region) Method:**
* Prompt used: `[selected region] plain black jacket, no hand, seamless fabric`
* Processing: Via Discord. Slow. 4 variations per job.
* Result: Output 1: Seamless, but jacket texture too smooth vs original. Output 2: Generated a button instead of hand. Output 3 & 4: Inconsistent lighting.
* Consistency: Low. Each variation a total re-imagine of the region.
**Photoshop Generative Fill:**
* Prompt used: `jacket fabric`
* Processing: In-app. Faster.
* Result: Generated 3 options. All matched existing jacket texture and lighting closely. One option near-perfect.
* Consistency: High. Options felt like extensions of the existing image.
**Summary:**
* **Midjourney:** Unpredictable. Treats region as a new prompt canvas. Good for creative changes, not precise fixes.
* **Photoshop:** Predictable. Analyzes surrounding pixels for context. Superior for realistic, seamless edits.
For corrections, Generative Fill is a better tool. For ideation within an image, Midjourney's Vary has its uses.
- bench_beast
Benchmarks don't lie.
I run a 20-person dev shop. We do real estate photography retouching at scale - 2000+ images a week, mostly architectural. We ditched Midjourney for fixes after two weeks of testing. Here's what we actually measure.
**Core comparison**
- **Consistency with source** - Photoshop wins. It's a pixel-aware diffusion model that respects the original texture gradient. Midjourney treats the region as a fresh latent space generation - you get a new button, a different lighting angle, or a fabric crease that doesn't match the existing wrinkles. We saw >90% acceptance rate on Photoshop's third option vs ~15% on Midjourney's best of four.
- **Throughput** - Photoshop is about 8-10x faster per edit in my env. In-app generation takes 3-5 seconds on a 512x768 region with an M1 Mac. Midjourney via Discord: 30-45 seconds per job, plus 4 variations you have to manually review. At scale, that's 20 minutes saved per 100 images.
- **Cost** - Adobe Photography plan $9.99/mo includes Photoshop and Lightroom, and Generative Fill credits are bundled (no extra per-use cost last I checked). Midjourney is $10-30/mo depending on tier, but you burn GPU time fast - each Vary (Region) call consumes a chunk of your fast queue. We hit the $30 plan's limit in 3 days of moderate testing.
- **Control** - Photoshop gives you a prompt *and* a mask that constrains the generation to the exact area. Midjourney's region selection is a brush stroke that often bleeds into adjacent pixels, and the prompt is just a hint. For a hand swap, Photoshop's "jacket fabric" prompt produced three options that all looked like the same jacket. Midjourney's "no hand" prompt still hallucinated a hand in two of four outputs.
- **Where it breaks** - Photoshop's generative fill struggles with large areas of missing detail (e.g., a whole face replacement) or repeating patterns. Midjourney is better for total reimagination if you're ok with the style shift. But the OP's use case - precise fabric fix - is exactly where Photoshop dominates.
**My pick** - Photoshop Generative Fill, for any fix where the surrounding context is what you want to keep. If you're doing creative exploration (replace a jacket with a dinosaur), use Midjourney. For production retouching, use Photoshop.
Simplicity is the ultimate sophistication
> "treats the region as a fresh latent space generation"
This is exactly the core difference. Photoshop's inpainting model (likely Firefly-based) uses a masked diffusion that conditions on the surrounding pixel values directly, not just the latent representation. Midjourney's Vary Region is effectively a text-guided re-generation of the masked area from scratch, which is why you see lighting mismatches and hallucinated objects like buttons.
One thing I'd add: the prompt phrasing matters a lot more for Midjourney here. Your prompt `plain black jacket, no hand, seamless fabric` is actually telling the model to generate a new jacket, not to fix the existing one. A negative prompt approach might help, but Discord doesn't support negative prompts easily. I've seen better results with `same jacket texture, match existing lighting` but it's still hit or miss.
For production retouching, Photoshop is the obvious choice. But if you're doing creative work that requires a total reimagination of a region (say, replacing a hand with a different object), Midjourney's unpredictability can be useful. Different tools for different jobs.
Show me the bill.
You're spot on about prompt phrasing making a huge difference with Midjourney. I tried a similar fix on a landscape photo where I needed to remove a signpost. Using "continue the grass and dirt path" gave me weird new plants. "Match surrounding terrain" worked a bit better, but the color grading was still off.
It really does feel like you're fighting the model's desire to create something new instead of just fixing what's there. For that kind of precise, context-aware work, I'll stick with Photoshop's tool every time. Midjourney's great when you *want* that surprise element, though.
Infrastructure as code is the only way