Key Takeaways

  • AI image models like Image 2.0 are exceptional at following precise visual directions from a sketch, but often render designs two-dimensional, losing realistic fabric flow.
  • Balancing strict adherence to a visual input with realistic textures and movement remains a core AI challenge in fashion design, demanding constant prompt refinement from the designer.
  • Even with advanced AI tools like Codex, generating accurate production patterns directly from designs is an "unsolved" problem, forcing Yana Welander to run human and AI pattern makers in parallel.
  • To achieve truly novel, yet realistic designs, Welander discovered she had to iterate aggressively, pushing AI models beyond generic runway variations while maintaining a sense of realism.

The Method: Parallel Play When AI Stalls

Yana Welander's approach to building an AI-native fashion startup, Yana Banana, isn't about letting AI take the wheel entirely. Instead, it's a shrewd, iterative dance between human and machine, especially in areas where AI's current capabilities hit a wall. She tackles the AI's precision-versus-realism dilemma head-on.

First, Welander leverages AI for visual direction, not always for a perfect final product. She starts with tools like Image 2.0 (and previously, Image 1.5 and Nano Banana) for initial design concepts. These models are great at taking a basic visual input—her sketches—and turning it into an image. “Image 2.0 does even better is it really well follows the sketch. In fact, it follows the sketch sometimes a little bit too precisely,” Welander explains. This precision often leads to images that look flat, "two-dimensional," or lack the natural flow of fabric. Her solution? Relentless prompt engineering, nudging the AI to understand and render the material properties: "You got to like it's got to flow." She pushes the AI to move beyond a literal interpretation of lines to a more physical representation of textiles.

Second, for critical production steps like pattern generation, she runs a parallel race. The biggest "unsolved" problem, as Welander describes it, is getting AI to create accurate patterns directly from her designs. This isn't a future challenge for her; it's a current, daily bottleneck. Her tactical response is to engage both human pattern makers and AI (like Codex) simultaneously. "And there I'm kind of running humans and AI in parallel," she says, “So, I am working with human pattern makers and Codex and just saying who's going to get to making my things fastest like most accurate that like to my vision, you know?” This approach ensures progress, leveraging the strengths of each while highlighting AI's specific gaps.

Finally, Welander invests heavily in iterative refinement for both novelty and realism. Other AI models, she found, might make a design “look beautiful and realistic,” but often stray too far from her original vision. They tend to produce variations of existing designs rather than truly new concepts. The hard part, and where she feels she's "cracked that nut," is pushing AI to create "really unique things per your vision while still making it realistic." This means constant experimentation, understanding the models' biases, and course-correcting when they default to generic or flat outputs.

Where This Breaks Down

This "parallel play" method is effective because Welander has a deep, specific understanding of AI's current strengths and weaknesses in fashion. It falters if you approach AI with an expectation of full autonomy or ignore its inherent limitations. Specifically:

  • Blind Trust in AI Precision: Assuming AI can handle highly precise, technical elements like garment patterns without human oversight or verification is a mistake. Current AI models often lack the geometric exactness and physical understanding required for manufacturing-ready patterns. Welander's need for human pattern makers confirms this.
  • Ignoring AI's Generative Bias: Believing AI will automatically produce truly novel designs without extensive, detailed prompting and refinement is naive. Most models are trained on existing data, making them prone to generating variations of what's already out there. Achieving "really unique things" requires a human designer's strong vision and persistence.
  • Underestimating the "Flow" Problem: Overlooking that realistic material properties—like how fabric drapes or flows—are not intuitively understood by image generation AI, especially when given flat, precise sketches. Without explicit prompting or human intervention, the AI might deliver visually accurate but physically unrealistic renderings.

What to Do With This

Look at your own creative or technical processes and identify areas where you need both precision and a nuanced, realistic output. Where does AI excel at one, but struggle with the other? Don't wait for perfect AI; instead, implement a "Welander-style" parallel workflow this week. Task an AI with its strengths (e.g., rapid ideation, generating diverse visual variations) while simultaneously engaging human experts for the areas where AI currently lacks the necessary precision or nuanced understanding (e.g., technical specifications, complex pattern generation, or final quality assurance). By running these processes side-by-side, you'll accelerate development and uncover the specific gaps you need to bridge with human intelligence.