Krea 3 vs. Event Horizon 3 — The Prompt Adherence Challenge

Six difficult prompts. Two generations of models. One question: how much better is Krea 3 at following instructions?

With the arrival of Krea 3, I decided to run a little experiment: put it head-to-head against Event Horizon 3 (EH3), my heavily optimized Krea 2 checkpoint.

The goal was simple: find out how much Krea 3 has improved at following complex instructions.

No beauty contest. No cherry-picked portraits. No cinematic lighting competitions.

Just prompt adherence.

First, an Important Disclaimer

PLEASE IGNORE IMAGE QUALITY IN THIS COMPARISON.

Seriously. That's not what we're testing.

The Krea 3 images were generated through the API, while EH3 is a heavily optimized checkpoint built on Krea 2.

It would be misleading to compare their visual quality under these conditions.

I fully expect Krea 3 to deliver significantly better visual results as workflows mature and the ecosystem develops. Judging its visual potential from these early API generations against a highly refined Krea 2 checkpoint wouldn't be particularly fair. :)

This comparison is EXCLUSIVELY about PROMPT ADHERENCE.

Can the models follow exact instructions? Respect object counts? Understand complicated spatial relationships? Handle impossible gravity? Maintain coherent reflections? Put things precisely where they're supposed to be?

That's what we're testing.

The Six Challenges

These aren't ordinary prompts. Each scenario contains multiple simultaneous constraints designed to push instruction-following capabilities beyond the usual subject-and-style descriptions.

1. The Impossible Elevator

Winner: Krea 3

Three people inside a glass elevator where gravity pulls sideways, while the surrounding hotel lobby obeys normal gravity.

Add an impossible mirror reflection, several architectural constraints, and precise object placement.

Krea 3 handles the overall spatial logic more convincingly, although it doesn't execute every instruction perfectly.

2. The Three Mirrors

Winner: Krea 3

A scene involving three mirrors, each presenting a different version of the same person.

The challenge is not simply generating reflections, but maintaining the specific relationships between each subject and their reflected counterpart.

Krea 3 demonstrates a substantial advantage here. EH3 captures the general idea but struggles to keep the individual instructions separated.

3. The Flooded Ballroom

Winner: Krea 3

A grand interior divided between submerged and dry areas, with specific characters and objects occupying precisely defined positions.

This tests whether the model can maintain an unusual environmental boundary without losing track of the scene's other constraints.

Krea 3 produces a considerably more faithful interpretation.

4. The Three Gravities Hotel

Winner: EH3!

Three different gravitational orientations coexist within the same architectural environment.

And here's the surprise.

EH3 handles a key gravity constraint better than Krea 3, which produces a plausible-looking composition but misses an important spatial instruction.

A reminder that newer models don't necessarily outperform their predecessors on every individual task.

Not bad for the old generation. :)

5. The Spatial Still Life

Winner: Krea 3

A deliberately constrained arrangement of objects, with precise quantities, positions, and spatial relationships.

No elaborate visual spectacle to hide behind. Either the objects are where they should be, or they aren't.

Krea 3 performs better, although neither model achieves perfect compliance.

6. Breakfast for Two

Winner: Krea 3

Probably the most ordinary-looking test, and one of the most revealing.

A breakfast table with two plates, carefully specified food quantities, differently colored cups, particular handle orientations, and multiple object-position constraints.

Generating a beautiful breakfast photograph is easy.

Generating the exact breakfast photograph described in the prompt is another matter entirely.

Krea 3 demonstrates stronger instruction following.

Final Results

Krea 3 wins 5 out of 6 challenges.

And frankly, that's what I expected.

We're comparing a new model generation against a checkpoint built on the previous architecture. Krea 3 should have the advantage.

The interesting question was never whether it would win, but how convincingly.

Across these six examples, Krea 3 shows noticeably better adherence to complicated instructions, especially when several spatial relationships and object constraints must be satisfied simultaneously.

However, it's not flawless.

It still makes mistakes, sometimes substitutes visually plausible arrangements for the requested geometry, and occasionally fails instructions that EH3 actually manages to follow.

Final Thoughts

Krea 3 is a genuine step forward in prompt adherence.

Not an overnight revolution that eliminates every instruction-following problem, but a meaningful generational improvement.

What surprised me was how competitive EH3 remained in certain scenarios, despite being built on the older Krea 2 architecture.

And it makes me genuinely curious about what comes next.

The real excitement will begin when Krea 3's model weights become available, workflows mature, and the community starts exploring what can be achieved beyond the base model.

For now, the result is straightforward:

Krea 3 wins the prompt adherence challenge, 5–1.

EH3 puts up a respectable fight.

And one last time: please don't judge Krea 3's visual quality from these images. That's an entirely different comparison, and one we're not making here.

— ElectricDreams

Want to try EH3? Get it here on Ko-fi — $11.90 + 6 FREE LoRAs!

Krea 3 open weights? Coming soon... hopefully!! 🤞