Has DALL·E 3 lost expressive variance for painterly art?

[Bronko_Kulitschka]

Title: Loss of expressive variance and painterly instability in DALL·E 3 compared to earlier versions

Category: Image generation / DALL·E

---

I’d like to raise a concrete, technical concern regarding DALL·E 3 that affects professional artistic use, particularly in painterly and expressionist contexts.

Summary

Compared to earlier DALL·E versions (notably 2023-era outputs), DALL·E 3 shows a consistent reduction in expressive variance, painterly instability, and non-systematic visual behavior. While instruction-following, clarity, and coherence have clearly improved, this appears to come at the cost of artistic freedom and malerische “roughness.”

Observed behavior

Across many prompts explicitly requesting painterly abstraction, impasto, visible brushstrokes, and imperfection, DALL·E 3 tends to converge toward:

• High detail clarity and object coherence

• Smooth, illustrative or airbrush-like aesthetics

• Systematic, evenly distributed texture patterns

• Conventionally “acceptable” outputs

Even when pushing strongly toward chaos, abstraction, or expressive error, the model stabilizes the image into a readable, polished form.

Style modes are insufficient

The current “vivid” and “natural” modes do not meaningfully address this:

• Vivid increases saturation and detail, but also reinforces smoothness and illustrative logic.

• Natural reduces contrast, yet remains within the same coherent, photoreal-adjacent aesthetic.

Both modes operate within the same optimization envelope and do not allow:

• non-uniform, irregular brush behavior

• intentional instability or visual “mistakes”

• expressive abstraction over detail fidelity

Practical impact

For artists working with expressionist, painterly, or deliberately imperfect aesthetics, this represents a functional regression. DALL·E 3 currently behaves more like an illustration/visualization system than a generative art tool.

DALL·E 2 (via API) sometimes produces more “raw” results, which further suggests a shift in optimization goals rather than a general limitation of generative image models.

Questions for the team

1. Is DALL·E 3 explicitly optimized toward “acceptable” or conventionally legible outputs compared to earlier versions?

2. Are there internal constraints (training, safety, post-processing) that suppress non-systematic or chaotic visual structures?

3. Are there plans to introduce explicit controls for abstraction, irregularity, or painterly expressiveness?

This is not a request for photorealism or prettier images.
It’s a request for creative degrees of freedom that previously existed and are now largely inaccessible.

I’m happy to provide before/after examples if that’s helpful.

Thanks for considering this feedback.

[jeffvpace]

@Bronko_Kulitschka

I’m happy to provide before/after examples if that’s helpful.

Yes, that would be helpful. Include the prompt as well if you can.

[Bronko_Kulitschka]

Thanks, happy to provide examples.

Prompt used (identical or near-identical across versions):

Expressionist oil painting, heavy impasto, thick paint, visible and irregular brushstrokes, unstable forms, painterly abstraction, dynamic color fields, human imperfection, no photorealism, no smooth surfaces, no illustrative or airbrush style

Comparison:

• The first images were generated with earlier DALL·E versions (2025).

• The later images were generated with DALL·E 3 (2026) using the same prompt.

Observed difference:
Earlier outputs show irregular, unstable brush behavior, visible painterly imperfection, and expressive abstraction.
DALL·E 3 outputs consistently converge toward smoother, more illustrative, and systematized textures, despite explicit instructions for instability, imperfection, and abstraction.

This behavior appears consistent across multiple runs and subjects, suggesting a shift in optimization rather than prompt sensitivity.

Let me know if additional examples or variations would be useful.

Earlier gen., 2025, very creative, human art:

[Bronko_Kulitschka]

Later gen., 2026, not much creativity, industrial vectors, no art, only system-print:

[jeffvpace]

@Bronko_Kulitschka

Note: The following was generated with the DALL-E-3 API and not ChatGPT - I’m not sure that ChatGPT can even create the DALL-E-3 images.

It appears that both your images were not created by DALL-E-3. More than likely, they were created by gpt-image-1.

Prompt: Image of a bearded man walking through the red light district. He is smoking a cigar. Expressionist oil painting, heavy impasto, thick paint, visible and irregular brushstrokes, unstable forms, painterly abstraction, dynamic color fields, human imperfection, no photorealism, no smooth surfaces, no illustrative or airbrush style.

Here is the image using the gpt-image-1.5 API:

Here it is for ChatGPT in the Playground - note that DALL-E-3 is not available

[Bronko_Kulitschka]

Thanks for the clarification — this helps narrow it down.

Based on the comparison, I think we’re actually observing two different behaviors:

• Older images (2023-era, ChatGPT image pipeline) show continuous, gestural stroke behavior: short directional strokes, overlap, irregular pressure, and visible interruption.
• Newer images generated via ChatGPT / Playground (gpt-image-1.x) consistently resolve into discretized, rectangular color patches.

Even when the prompt explicitly requests “visible irregular brushstrokes” and “unstable forms,” the texture logic appears tile-based rather than stroke-based. Impasto is simulated volumetrically, but the physical logic of brush movement (drag, direction, breakage) is no longer present.

The Red Light District example actually illustrates this well:
The semantic prompt is followed, but the surface structure converges toward a uniform mosaic-like pattern.

So the core issue seems to be a regression in painterly stroke continuity within the ChatGPT image model (gpt-image-1.x), not a misunderstanding of expressionist style labels.

If useful, I can label my examples explicitly as:

• older ChatGPT image output (unknown model ID, 2023)
• newer ChatGPT / Playground output (gpt-image-1.x)

Let me know if that helps with internal reproduction.

[jeffvpace]

Prompt: Create an image of a bearded man walking through a red light district in the style of Edvard Munch.

I think you can eventually get what you want using the right prompt with gpt-image-1.5. In many cases “over prompting“ leads to undesirable results. For example, your later prompt was applied to the entire image without nuance. Maybe the latest gpt-image-1.5 in not automatically appling nuance?

[Bronko_Kulitschka]

Thanks for the example — it helps clarify the distinction I’m trying to make.

I agree that using a Munch-style prompt reduces harshness and makes the image feel more painterly at first glance. However, the core issue I’m describing remains unchanged:

The surface texture is still composed of short, discretized segments rather than continuous, gestural brush strokes. Even here, there is no clear stroke trajectory (start–drag–release), no pressure variation across a long movement, and no overlapping directional pull that would reflect human brush behavior.

In other words, the style shifts, but the underlying stroke-generation logic does not. This isn’t about nuance in subject treatment or over-prompting, but about the absence of continuous stroke continuity in the model’s texture synthesis.

So while certain prompts can mask the issue aesthetically, they don’t restore the physical logic of painterly gesture that earlier outputs occasionally exhibited.

[jeffvpace]

With all due respect, I think you are being overly critical. Good luck.

[Bronko_Kulitschka]

Understood. My goal here wasn’t to debate whether the images are “good” or whether certain styles can be approximated.

I was specifically documenting a change in stroke continuity and texture generation behavior compared to earlier outputs. I’ve provided examples and technical observations for that purpose.

I’ll leave it there. Thanks for engaging. Maybe you try to paint Van Gogh Style, then you will see the limitations.

[Bronko_Kulitschka]

2025:

[Bronko_Kulitschka]

2026:

[Bronko_Kulitschka]

2025: van Gogh Style, Art

2026: Crap

[_j]

@Bronko_Kulitschka

2025: van Gogh Style, Art

2026: Crap

“2025” and before - 1536x1024 = gpt-image-1, not DALL-E
2026 - also 1536x1024
What you said was DALL-E - also 1536x1024 and not DALL-E

This is DALL-E 3, API, HD

I suspect that the latest image_gen was trained on a lot of DALL-E 3 distillation, BTW, as it seems to mimic the style on occasion.

[jeffvpace]

Here is DALL-E-3 without the fancy prompt.

Prompt: Create an image of a bearded man walking through a red light district in the style of Vincent van Gogh.

[jeffvpace]

@Bronko_Kulitschka

Maybe you try to paint Van Gogh Style, then you will see the limitations.

Yes, you are correct. Gpt-Image-1.5 does not do Van Gogh very well…