Last week I pushed out a blog article called “Can AI Separate Sim Process? I Put It To The Test.” If you missed that article, you might want to click and read it.
This article picks up after that one.
The idea that I’ve been working on lately is that shops can use AI tools in their art department to handle repetitive tasks. Whether it is building an art approval system, creating images, or processing art files, AI can add tremendous value and time savings to any busy shop.
Converting an image to spot color for screen printing is both an art and a science that takes years to master. I’ve been doing it since the late 1990s, and I’m still learning. Last week’s article was about my attempt to experiment with whether AI could handle this complex work.
But I want to admit something. The article proved that AI could build a sim process file.
It clearly had results. What it didn’t prove completely was that the steps I showed handled everything correctly.
This article is about my next steps on that journey.
What I’ve Discovered
Since last week, I’ve kept pushing.
Several shops reached out, as they were intrigued. A bunch of folks tried my Markdown file I provided and had good results. Better, a few people explained a few things they knew I was missing in my Markdown instruction file. Mark Coudray gave me some excellent tips on ink opacity, color measurement, and adding a confidence level rating.
All last week, I ran more tests through the process. Different shirt colors, art file types, and scenarios. I tested only ChatGPT and Claude as the AI platforms. The work I put in provided the discoveries that changed how I think about the process. The article and results I published last week weren’t wrong. I just didn’t know enough yet.
Here’s what I learned.
What Did The First Round Miss?
My biggest miss was simply judging the results by eye. Frankly, I was thrilled that it actually worked. But there were some noticeable challenges. Which is why I kept pushing.
- In some instances, Channel names were missing their last letter (“BLACK SHIR” instead of “BLACK SHIRT”)
- Shirt Color Channels were void of color and were white
- Directions given were not followed
- Spot Color Opacity settings all had the same setting, as if a bright orange, brown, or yellow print the same way
- There wasn’t a good way to determine whether the separation action was the best it could be.
The job of the Markdown file is to provide instructions and guardrails to separate the art correctly and ensure a repeatable outcome. In these new tests, it was apparent that I needed more rules.
How Much Testing Did I Do?
Below, I’m showing one result so you can see where we are now. Since last week, I’ve tested numerous art files, often 3 or 4 times to root out and solve issues. Some of these tests included some real-world scenarios, such as a single set of screens for both white and black shirts in a single order. I tested tonal designs, or designs with intricate areas. I didn’t want the results from the instructions to hinge on 1 lucky art file.
Every time something broke, the instructions got a little longer and more detailed. I’m not going to walk you through every sep run. That would be a long, boring article. Later in the article, I’ll show some examples from one set of separations for a typical simulated process sep, “Go Tigers,” because I think this shows what matters.
What Is ΔE, and Why Should You Care?
ΔE (pronounced “Delta E”) is a single-number measure of the difference between 2 colors for color matching. Like a golf score, the lower the number, the better. The bigger the number, the bigger the miss.
Think of it as a tape measure for color. Instead of arguing over whether an orange is “close enough,” use a scientifically validated method to measure color. Here’s a way to understand it:
| ΔE | What you’d see |
|---|---|
| Under 1 | No visible difference |
| 1 to 2 | A trained eye might catch it side by side |
| 2 to 5 | Noticeable side by side, acceptable for most garment work |
| 5 to 10 | Clearly a different shade |
| Over 10 | A different color |
What is ΔE 2000?
ΔE 2000 is the 2000 revision of the formula, and it’s the version the color industry uses today.
Older color formulas treated every color the same. Your eye catches a small shift in gray or blue much faster than the same shift in bright orange. ΔE 2000 weighs those differences the way your eye does.
Scoring Simulated Process Separations
At the suggestion of Mark Coudray, I added a Fidelity Score for color matching, and it turns ΔE into a rating from 0 to 100. The Fidelity Score measures how closely the print, predicted from actual Photoshop Channels in the file, matches the original art for color.
This was not in the first Markdown instruction file from my blog last week. I think it makes all the difference.
How Does That Turn Into A Score?
The formula is simple.
Fidelity = 100 − (average ΔE × 4) − (% of the art that misses by more than ΔE 10)
For example, if you look at Claude’s version of the “Go Tigers” separation below on a white shirt, the measurable results for the separation averaged a ΔE of 4.46, with 0.8% of the art missing by more than 10. That works out to a Fidelity of 81.3 out of 100.
More on that later.
What’s New in the Separation Markdown Instructions?
The Markdown File you can download for free below has a big rewrite. Here’s what changed from last week, in plain shop terms:
- More fail conditions up front. The file lists exactly what makes a separation a failure. For example, if the white underbase prints even slightly past the edge of the art, you get a white halo on a black shirt. That’s an automatic fail.
- Ink opacity classes. Inks are grouped from A (opaque, like whites, black, and browns) to D (very translucent, like deep blues and purples). Each class gets its own opacity setting, because a bright orange doesn’t cover like a brown. My original Markdown file treated every color the same, and printing inks don’t behave that way.
- Your stock black. The black plate is the black already sitting in your ink bucket, not some Pantone the AI picked. I don’t know any shop that is printing Pantone Black 4 C.
- Colored underbase as an option. A tan, light gray, or light blue base can keep burgundy and royal from shifting. I’ve used this on press for years, and now the instructions offer it.
- The layered solve. Every pixel gets its own plate values, calculated based on how ink actually layers on a shirt. No posterizing the art into a handful of colors.
- A swatch strip on every file. Solid patches of each ink, on the open shirt and over the underbase, placed below the art. You print it on the test shirt to see how each ink really behaves on that garment, then block it off for production. This is used to measure the actual color, which we can report back to the AI to learn how the seps work. If this bothers you, just tell the AI not to run it.
- 1 set of screens for multiple shirt colors. Each shirt color gets its own preview Channel you can switch on and off in Photoshop. I tested a scenario with four shirt colors, and the AI chose to separate four completely different separation files. This resolves to building Channels for shirt colors and keeping everything in one file.
- The build script and build log are now deliverables. You get the actual code that built your file. More on why that matters in a minute.
How Do You Rate Separation Work Without Guessing?
Another thing I added to the instructions was a Confidence Score, which I also built as a rating from 0 to 100. This is different than the Fidelity Score. It’s good to know if the AI is confident that the separation files it built are as accurate as can be.
The Confidence Score uses the Fidelity Score (color-matching rating), subtracts points for known risks, and is capped by the amount of real data in your process that your shop has given the AI.
Risks include an art file being enlarged by the AI more than 2x, or files that have not yet been opened in Photoshop. Each one costs Confidence Points.
I’ve capped different score levels based on how the shop operates and what is known. A preview based on estimated ink values that haven’t actually been measured can’t earn a high score, no matter how good it looks.
Want a higher score for your separation work? Start measuring the results and giving the AI the information to use the next time you build a separation.
| What your shop has done | Highest possible score |
|---|---|
| No measurement | 70 |
| Photographed the swatch strip from a test print | 80 |
| Measured the swatch strip with a spectrophotometer | 90 |
| Measured, plus at least 1 press check compared to the prediction | 100 |
You don’t need a spectrophotometer to use this. But until you do, the confidence that the seps are accurate will be less.
The Go Tigers Test: Where the Rubber Meets the Road
As I mentioned, I gave simultaneous tests to both ChatGPT and Claude constantly. Every time, they received the same files and the exact same instructions.
For the “Go Tigers” art, here’s what I handed both platforms:
- The “Go Tigers!” art file. This was a ChatGPT file, 1431 px wide, just like a client might send you. The AI had to enlarge it 2.73x to reach 13” at 300 ppi.
- Requested to separate spot colors for screen printing, using one set of screens to print on both a white shirt and a black shirt.
- Use as few colors as possible. I did not specify colors or the number of colors as I wanted to see how the AI tools could work the problem.
Same art. Same instructions. Same request.
ChatGPT finished first. It was horrible. In earlier tests I thought I was making progress, but this time the orange came out muddy brown, and the entire file lacked any zip.
Then I had an epiphany. Maybe there was a higher level of ChatGPT, and that could do a better job. Turns out, there was. ChatGPT recommended that I use GPT-6 on High or Extra-High Thinking Mode. I reran the seps in GPT-6 High, and it worked much better.
Here are the previews for the three seps shown above. The first is the ChatGPT V1 sep, the second is the ChatGPT GPT-6 High sep, and the third is the Claude sep.
| ChatGPT V1 | ChatGPT V2 (GPT-6, High) | Claude | |
|---|---|---|---|
| Screens | 6 | 6 | 5 |
| Fidelity, white shirt | −23.7 | 75.8 | 81.3 |
| Fidelity, black shirt | 13.5 | 78.1 | 81.2 |
| Art missing by more than ΔE 10, white/black | 71.9% / 42.4% | 5.2% / 3.4% | 0.8% / 1.5% |
| Confidence, white/black | 0 / 0 | 65.8 / 68.1 | 70 / 70 |
ChatGPT V1 built a file that opened. The structure was fine. But the colors were wrong. The orange printed like burnt rust, and almost 3/4 of the art on the white shirt missed by more than ΔE 10. A negative score is possible when it’s that far off. Yuck.
ChatGPT V2 was a real step up. It’s a legitimate separation. But it used 6 screens when 5 would have done the job just as well. Its underbase extends beyond the edge of the art, showing up as light specks and a tan edge around the splatter on the black shirt. And it went back and edited the plates after it had solved them.
[IMAGE: Edge close-up on the black shirt, Claude vs. ChatGPT V2]
Claude used 1 fewer screen and scored higher on both shirts: Underbase White, PANTONE 151 C, PANTONE 7416 C, Black, and Highlight White.
Here's a screenshot of the Claude-based seps in Photoshop. There are two Channels for the White or Black Shirt colors, plus the 5 correctly separated spot color Channels.
To be fair, Claude’s first build missed too. It scored 55 on the white shirt. It had capped the black plate at 95%, which let a little white shirt show through the linework and turned the blacks gray. The instructions say anything under 60 gets reworked. So it found the problem, let the solid linework print at 100%, and came back at 81.3. It did this automatically, without me asking.
That’s a solid win and instructions doing their job. The rules caught the mistake before I ever opened the file.
One more thing to keep in mind. These scores measure how well each file fits the print model. “Go Tigers” hasn’t been on press yet.
As always, the test print is the real judge.
Getting Information
During the Go Tigers work, a big lesson was dumped in my lap. Even the GPT-6 High version had problems I couldn’t see in the preview. That changed when I asked ChatGPT for the actual build code it used to make the separation.
The report ChatGPT wrote said it validated everything. The code showed what it actually did:
- Its halo check counted faint, nearly transparent edge pixels as part of the art, so the underbase printed under them.
- After solving the plates, it pushed the darkest areas toward solid red and black on its own, outside the rules.
- It only tested ink sets with 6 or more screens, even though I asked for the fewest colors.
The actual code showed how ChatGPT processed the work and made mistakes, so I could create instructions and rules to address incorrect assumptions or actions.
What I changed in the instructions because of it
The AI follows exactly what you wrote, not what you meant. So the new instructions now:
- Define exactly which pixels count as art for the halo check.
- Ban any changes to the plates after the solve, other than the listed cleanup steps.
- Require testing smaller ink sets from scratch when you ask for the fewest screens.
Claude is still the champ
One thing I can report is that across all the work, while testing several different files, through multiple sets of instructions, Claude consistently delivered what I requested each time.
I’m sure some of you out there love ChatGPT and want to keep working and using it for projects like this. To tell you the truth, I’m all for it. I like to keep working using the same tools too.
But ChatGPT was like that hyperactive kid that can’t sit still. It kept straying away from what I wanted.
Claude, on the other hand, was rock solid and delivered the goods every time. Even when it missed something, it caught it, fixed the cause, and told me what happened.
So, Where Is This Work Going?
Below is the latest version of my Markdown Instruction file. Is it finally perfect? Probably not. I need more feedback, testing, and experiments.
I do know that it is much better than last week. Take the Pepsi challenge and download the Markdown yourself and try it. Use the last file you separated, or maybe the most difficult one you slaved over to get right. See what happens.
What’s Left On My Radar
I want to build more structure around this, and I’ll create a Basic Separation Tool that I’ll provide here for free. It’s going to have 4 parts:
- Shop Profile: your inks, mesh counts, garments and press results
- Separation Engine: the instructions doing the work
- Job Ticket: what each order needs
- Feedback Log: what actually happened on press
Work like this needs data to get better. Press checks, color measurements, and data help provide a more accurate set of separation instructions. I’m just getting started.
I also want to be clear about one thing. I don’t think my work is about replacing anyone’s creativity or skilled judgment with separation work. To me, the more someone spends their time doing higher value work, such as designing, and less time doing task work, such as getting the art approved or separating the files for production, the better.
What Should You Do This Week?
- Download the updated Simulated Process Separation Instructions
- Run 1 past job you’ve already printed through it
- Compare the new Photoshop file to the one you had already created
- Burn some screens with the file
- Print the swatch strip on a test shirt
- Compare the printed shirt to the preview
- Take a pic with the image, test strips and a blank white piece of paper and email me
Lastly, I’m looking for a few shops that want to explore automating their art department workflows. I’m looking for explorers who want to go on a journey. I’m 100% going to be building out content and education around this. Want to be a part of that? I have a handful of shops already interested in this. Join us!
Shoot me an email (marshall@marshallatkinson.com) or book a call.