Will the new GPT-Image 2 generate an electronic circuit diagram? Comparison with Nano Banana 2

p.kaczmarek2 24 Apr 2026 11:32 6 435 Cool? (+1)

📢 Listen (AI):

TL;DR

GPT-Image 2 (medium) is compared with gemini-3.1-flash-image-preview nano-banana-2 on electronics schematics, component layouts, product analysis, and photo-editing tasks.
The prompts test relay control, LED strip dimmers, PIC16F1459 USB pinouts, ESP32/Tasmota flashing, teardown annotations, and object placement on boards and photos.
The evaluation uses LMArena and includes concrete cases like a 12V LED strip dimmer, a 12V RGB strip dimmer, and a PIC16F1459 DIP schematic.
GPT-Image 2 often produces richer text, tables, and infographics, but its electronics logic and component accuracy remain unreliable, sometimes worse than Nano Banana 2.
Both models sometimes succeed on object relationships and simple edits, but the verdict stays cautious because only the medium GPT-Image 2 version was tested.

Generated by the language model.

I invite you to a practical test of the new image generator from OpenAI. In this topic I will test the performance of GPT-Image 2 on various electronics related tasks, there will be drawing schematics, describing devices and also editing photos and graphics. In addition, I will compare the whole thing with the Nano Banana 2. Is the new model really better? Let's find out!

A large part of the prompts and tasks in this topic have already been tested with the Nano Banana models, feel free to visit the related topic:
Will the Nano Banana generate an electronic diagram or infographic? Comparison of the two models

Here I have re-made them - I used the free LMArena and the two models available there:
- gpt-image-2 (medium)
- gemini-3.1-flash-image-preview (nano-banana-2) [web-search]
Please note that depending on the models available to you, the results may vary.

Presentation format
First the prompt will be given, then (if available) the photo attached for AI. Then the model name will be given, the images generated by it, then another model name and again the images. Finally, my loose comments.

Model comparison
a schematic showing how to control relay from arduino GPIO with a transistor and protection diode
gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

It is immediately apparent that GPT-Image 2 has a distinctive style, especially these annotations and explanations, worse that they are not correct. The relay connection to the manifold is lost....

draw detailed schematic showing how to make a 12V LED STRIP DIMMER with Arduino

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

GPT-Image 2 generates more complex images, puts a lot of explanations, diagrams, tables on them, but with the logic is still a bit worse. Still that shorting of the 12 V line with the drain of the MOSFET transistor.... interesting that the PARTS LIST however agrees with the schematic. Why does D9 (pin?) have a resistor schematic? What is the Schottky diode there for? The code is even there... and came out quite good? Information about the common ground also correct.

draw detailed schematic showing how to make a 12V RGB STRIP DIMMER using encoder with Arduino

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

As before - GPT-Image 2 likes to do complex graphics, but with logic it's worse. Still, those MOSFETs cascade....

show minimal connections of PIC16F1459 in DIP package schematic with USB connector

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

Pinout gets a bit mixed up with the actual one. Pin 10 is actually RB7 and 11 is RB6. USB is on 19 and 18. still those VBUS to MCLR.... but the Nano Banana did worse anyway, although .... D+ and D- tracked well.

Draw four elements on table: a 2200uF 30V capacitor, a Zener diode, a WS2812B single module, and a DIP8 socket

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

GPT-Image 2 vestigially handles text better, the rest the same, WS2812 pinout logic came out average.

draw a proto board with 3300uF capacitor 50V on it, on capacitor put a gold coin, on coin put an Arduino UNO, and on Arduino UNO put needle

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

These needle alignments are unnatural, but GPT-Image 2 did a slightly better job with the relationships between objects, they came out more natural.

flyback power supply schematic

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

GPT-Image 2 likes to apply a block of text to the images, but the sense in the images is not there. Still that shorted EMI FILTER... fB pin on the ground... aND the whole bit with the reference voltage taken from ground.

Arduino UNO with 28BYJ-48 and motor driver connect by wires

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

Similarly, but it is apparent that GPT-Image 2 likes diagrams and tables and text. Interestingly, both models know where the power and ground are on the Arduino.

Arduino UNO on a plate, next to old Unitra Silesian radio on beach during sunny day

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

Very similar, although the option with search mapped the radio better. I think it searched for them.

1960 old room with TV showing modern Arduino UNO in workshop
gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

I find it difficult to give a verdict on which is better.

draw graphic guide - ESP32 flashing for firmware change to Tasmota
gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

On this board you don't need to manually short GPIO0 to ground.... interesting, the rest came out pretty good. Nicer infographics are made by this GPT-Image 2.

So how to connect USB TO TTL to TYWE3S module for flashing?
gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

A bit faint, wrong, pinout is also wrong. This is how you can burn a module. A disastrous tool.

please plug a white phone charger with empty USB A slot into my power strip

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

Banana got it wrong once, and yes both average.

change image to remove Flash Memory Chip so only empty pads are visible

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

GPT-Image 2 total fail, doesn't seem to understand what Flash memory is.

Change voltage to 9.03V , current to 2.50A and capacity to 46444mAh

gpt-image-2 (medium)

gpt-image-2 (medium)
gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

Both managed.

change hdd to 2.5" SSD Samsung EVO 4TB

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

Both generators have a problem.

convert schematic to cadsoft eagle style and format

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

GPT-Image 2 transferred the style better, but also messed up the schema.

make advertisement for this product

gpt-image-2 (medium)

gpt-image-2 (medium)
gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

Ad-style graphic of a VL-34U4 time relay on a wooden workbench with tools and circuit sketches

I find it difficult to judge which version is better.

make analysis of this product (what is inside, what is on back)

gpt-image-2 (medium)

Please annotate sections of this device:
Be detailed and technical

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

Looks nice, but the errors are fundamental.

please put a common mode choke in the correct place

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

Rather more realistic results from Nano Banana

make a teardown of this device

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

A rather meaningless result, but you can see that GPT-Image 2 prefers infographics and diagrams.

remove usb sockets from the photo

gpt-image-2 (medium)

And here's what happened? GPT-Image-2 removed the ground along with the USB...
gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

Create a realistic photo showing how would this device look in real life

gpt-image-2 (medium)

gemini-3.1-flash-image-preview (nano-banana-2) [web-search]

a13b6bb9e
Rather comparatively - once better, once worse.

In summary , it seems that GPT-Image 2 is not at all significantly better than Nano Banana 2 in this context. My impression is that GPT-Image 2 is very fond of generating text, tables, diagrams and infographics, but nevertheless with the field of electronics is still poor. In addition, on the other hand, I saw that GPT-Image 2 does slightly better with object relationships (the needle test on the Arduino) and slightly worse with the electronic components themselves (it does not know what Flash memory is). In summary, based on these tests I would say that GPT-Image 2 has other stronger points than Nano Banana 2, it handles text better, but is not a great step forward overall.
For a full verdict I would still need to test the high version, if one is available. So far I have tested everything on medum.
And what is your opinion? Have you tested GPT-Image 2?

About Author

p.kaczmarek2 wrote 14368 posts with rating 12284 , helped 649 times. Been with us since 2014 year.

Comments

Add a comment

gulson 24 Apr 2026 12:50

Interesting. I expected better results from image 2 One test to try at your leisure - where he gets confused in a diagram or similar, give him a netlist of connections between elements. So that he knows... [Read more]

p.kaczmarek2 24 Apr 2026 13:13

I was also just thinking of trying to write in the prompt what to connect and how to connect, as I will then generate. I'll try it at my leisure. I'll try to redo this example "draw detailed schematic... [Read more]

gregor124 24 Apr 2026 13:29

All in all, if someone asked me about flash memory, I wouldn't know which specific model or enclosure they were referring too. And yet in professional design you have to be accurate. This is what AI... [Read more]

p.kaczmarek2 24 Apr 2026 13:33

Just in a situation like there on the PCB, the choice is quite obvious. What's more, the Nano Banana 2 copes with it, and recognises well the distinctive casing and markings of a chip that is common even... [Read more]

gregor124 24 Apr 2026 13:54

I don't know, for me such a choice is not obvious and it is difficult to judge on this basis whether one is better than the other. I'll give an example like this: I send the first messenger to get rice... [Read more]

gulson 24 Apr 2026 15:34

But it would be interesting to see how he would handle the netlist. I found one case in point. https://github.com/nturley/netlistsvg Schema original created with JSON: https://obrazki.elektroda.pl/5230585500_1777037602_bigthumb.jpg... [Read more]

FAQ

TL;DR: Across 20+ practical tests, GPT-Image 2 looked richer but often failed basic circuit logic; as one expert summary put it, "a disastrous tool" for risky wiring prompts. This FAQ helps electronics users judge when GPT-Image 2 is useful, when Nano Banana 2 is safer, and where both still fail on schematics, PCB edits, and teardown graphics. [#21889868]

Why it matters: If you use AI images for electronics, visual polish can hide wrong pinouts, unsafe wiring, and destructive flashing advice.

Test area	GPT-Image 2	Nano Banana 2
Schematic appearance	More text, tables, infographic style	Simpler drawings
Circuit correctness	Often wrong connections	Usually less polished, sometimes more realistic
Object relationships	Slightly better in stacked-object test	Slightly weaker
PCB/photo editing	Mixed; sometimes removed wrong copper/ground	Mixed, but sometimes more faithful
Overall verdict in thread	Not a major step forward	Competitive, sometimes safer for electronics

Key insight: GPT-Image 2 can make electronics images look authoritative, but the thread shows that visual confidence did not translate into correct electrical logic. For schematics and flashing guides, wrong details mattered more than prettier output.

Quick Facts

The main comparison used 2 models in LMArena: gpt-image-2 (medium) and gemini-3.1-flash-image-preview (nano-banana-2). [#21889868]
Repeated test prompts included 12V LED-strip dimmers, 12V RGB strip control, a PIC16F1459 DIP USB schematic, relay drivers, flyback supplies, and PCB/photo edits. [#21889868]
In the PIC16F1459 check, the author states pin 10 = RB7, pin 11 = RB6, and USB should be on pins 19 and 18. [#21889868]
One display-edit task required changing readouts to 9.03V, 2.50A, and 46444mAh; both image generators completed that prompt. [#21889868]
A follow-up post added a netlist-style idea and showed a JSON-driven schematic example, but GPT-Image 2 still produced a result described as pretty yet nonsensical. [#21890013]

How well does GPT-Image 2 generate electronic circuit schematics compared with Nano Banana 2?

GPT-Image 2 generated more elaborate schematic-like images, but it was not clearly better at actual electronics. The thread’s overall verdict says GPT-Image 2 handled text better, yet still showed weak circuit logic and was “not at all significantly better” than Nano Banana 2 in this use case. Nano Banana 2 often looked simpler, but several outputs were judged more realistic or less dangerous for wiring tasks. [#21889868]

Why does GPT-Image 2 add lots of text, tables, and infographics to electronics images while still getting the circuit logic wrong?

GPT-Image 2 appears to favor presentation over electrical correctness. Across multiple prompts, it added annotations, parts lists, code blocks, tables, and infographic elements, yet still shorted lines, misplaced feedback pins, or miswired MOSFET stages. The author’s repeated observation was that the model has a distinctive explanatory style, but “with logic it’s worse,” which makes the images look authoritative even when the circuit is broken. [#21889868]

What mistakes did GPT-Image 2 make when drawing an Arduino relay driver with a transistor and flyback diode?

It lost a key relay connection and added incorrect explanatory notes. In the relay-driver test, the author says the “relay connection” was effectively lost, even though the prompt explicitly asked for Arduino GPIO control through a transistor and protection diode. That means the image looked like a teaching diagram but failed the core requirement: a complete transistor-switched relay path with the diode shown meaningfully. [#21889868]

How should a 12V RGB LED strip dimmer with Arduino, MOSFETs, and buttons actually be connected?

The thread’s corrected prompt says each RGB channel should go to the drain of its own MOSFET, and each MOSFET source should go to ground. The LED strip should use a common +12V rail, and the control side should use 3 buttons with pull-up resistors: one for mode and two for brightness up/down. That setup avoids the model’s repeated mistake of inventing cascade-like MOSFET wiring. [#21889912]

What happened when the prompt for the Arduino RGB strip dimmer was rewritten with explicit connection details?

The clearer prompt improved the instruction quality, but GPT-Image 2 still miswired the design. After the rewrite, the author says the model “absurdly connected the PWM to the gates of the transistors,” while Nano Banana 2 produced the better result in that round. The test shows that even explicit connection guidance did not reliably force GPT-Image 2 to draw a valid schematic. [#21889912]

In the PIC16F1459 DIP USB schematic test, which pins were mixed up and what are the correct USB connections?

Pins 10 and 11 were mixed up, and the USB pins were misplaced. The author states that pin 10 is RB7, pin 11 is RB6, and USB should be on pins 19 and 18. He also flagged the model’s odd tendency to route VBUS to MCLR, which is another example of a plausible-looking but incorrect microcontroller schematic. [#21889868]

What is a netlist, and how could giving a netlist help an image model draw a correct schematic?

“Netlist” is a connection-description format that lists which pins and nodes join together, using explicit connectivity rather than visual guesswork. In the thread, one participant suggested giving the model a netlist so it would “know how to connect the elements,” almost like importing into KiCad. That could reduce guesswork because the model would receive the wiring rules directly instead of inferring them from plain-language prompts. [#21889908]

What is netlistsvg, and how is it used to turn a JSON netlist into a circuit diagram?

“netlistsvg” is a schematic-rendering tool that converts a JSON netlist into a visual circuit diagram, using declared modules, ports, cells, and connections. The thread links to the GitHub project and shows a JSON example whose original diagram was generated from that data. That example was then used as a comparison point, and GPT-Image 2 still produced a decorative but nonsensical redraw. [#21890013]

Which kinds of electronics tasks did GPT-Image 2 handle better than Nano Banana 2 in these tests?

GPT-Image 2 did better on text-heavy visuals, infographic styling, and some object-relationship tasks. The author says it handled text better overall, made nicer firmware-flashing infographics, and performed slightly better in the “needle on Arduino” spatial-relationship test. It also transferred the style of a CadSoft EAGLE-like schematic better, even though it still damaged the actual circuit logic. [#21889868]

Where did Nano Banana 2 produce more realistic or safer results than GPT-Image 2 for electronics-related images?

Nano Banana 2 often produced safer results in exact electronics tasks, especially where wrong wiring could damage hardware. The thread says Nano Banana 2 was better on the rewritten 12V RGB dimmer prompt, more realistic in common-mode-choke placement, and better at recognizing the flash-memory chip case and markings. For the TYWE3S flashing diagram, GPT-Image 2 was called “a disastrous tool,” which strongly favors the safer alternative. [#21889868]

How do you connect a USB-to-TTL adapter to a TYWE3S module for flashing without damaging the module?

The thread does not provide a correct pin-by-pin TYWE3S wiring recipe; it only warns that the generated diagrams were wrong enough to burn the module. The direct takeaway is procedural: 1. do not trust the AI image alone, 2. verify the TYWE3S pinout independently, 3. confirm every adapter-to-module connection before powering it. The author explicitly says the shown result was wrong and could damage hardware. [#21889868]

What are the correct steps for flashing an ESP32 board to Tasmota, and when is grounding GPIO0 unnecessary?

The thread gives one concrete rule: on the tested board, you do not need to manually short GPIO0 to ground. The author says GPT-Image 2 made a nice-looking flashing guide, but that specific step was unnecessary for that board. So the safe reading is: verify the board’s boot method first, because a generic ESP32 flashing diagram may add steps that do not apply to your exact hardware. [#21889868]

Why is removing or editing specific PCB parts like a flash memory chip or USB sockets so difficult for image generators?

These edits require the model to preserve board context while changing only one component. In the thread, GPT-Image 2 failed to remove a flash-memory chip cleanly and later removed USB sockets along with surrounding ground, which shows poor understanding of PCB structure. A commenter also notes that “flash memory” is too vague professionally unless package and exact device are clear, so ambiguity worsens image editing errors. [#21889933]

How accurate are AI-generated annotations and teardown graphics when identifying parts inside power supplies and other devices?

They were visually attractive but technically unreliable. The thread says GPT-Image 2 liked annotation-heavy graphics and teardown-style diagrams, yet made “fundamental” errors in section labeling and produced meaningless teardown results for some devices. The same pattern appeared in the flyback power-supply test, where text blocks looked informative while core details like EMI filter and feedback placement were still wrong. [#21889868]

What is a common mode choke, and where should it be placed in a mains power supply schematic?

“Common mode choke” is an EMI suppression component that attenuates common-mode noise on paired mains lines, using two coupled windings placed in the input filter path. In the thread, the test asked the models to place it correctly, and the author judged Nano Banana 2’s results more realistic. That implies the choke belongs in the mains-input EMI section, not as a random added symbol elsewhere in the supply. [#21889868]

Generated by the language model.