Krea 2's Two-Pass Pipeline: Making AI Images That Don't Look AI

Most photorealistic AI images fail the same way: they look almost real, but something is off. A phone logo is slightly wrong. A reflection does not match its source. Skin has that uniform, plastic smoothness. Viewers cannot always name the problem, but they feel it instantly. This is the "AI look" — and it kills credibility in commercial work.

The fix is not adding more detail words to your prompt. The fix is a second sampling pass that lets the model correct its own mistakes.

What makes Krea 2 different

Krea 2 is a 12-billion-parameter diffusion transformer — not a U-Net. It pairs a dense SingleStreamDiT architecture with a Qwen3-VL text encoder, which gives it unusually strong prompt adherence. The training data was curated specifically to strip out the visual signatures that make images read as AI-generated: the over-smoothed skin, the too-perfect lighting, the anatomy that is 95% right.

But the architecture alone is not what removes the AI look. That comes from the two-pass sampling pipeline.

The two-pass pipeline explained

Pass one generates a complete image using Krea 2 Turbo's fast 8-step sampler. The result is good — composition, color, and subject are all correct. But if you look closely, fine details betray the generation. A brand logo on a phone back is misshapen. A reflection is physically impossible. Text on a label is garbled.

Pass two takes that first image and runs it through a second sampling round. Instead of starting from noise, the model starts from its own first-pass output and is prompted to refine toward realism. It effectively asks itself: "How would this object actually look in a real photograph?" — and corrects the small failures.

The difference is striking. Logos snap into correct shapes. Reflections align with their light sources. Skin gains pores and asymmetry instead of airbrushed perfection. The image stops looking generated.

RAW vs Turbo: two models, two jobs

Krea 2 ships as two variants with distinct purposes:

• Krea 2 RAW — the training base model. Highly plastic and ideal for fine-tuning. Train your LoRA here to teach the model a custom style, product, or character. Officially not recommended for inference.
• Krea 2 Turbo — the distilled generation model. Produces high-quality output in just 8 steps (4 steps if you provide a reference image). This is what you use for actual image generation.

The intended workflow: train on RAW, deploy on Turbo. You get a custom-styled model that still generates at turbo speed.

Running on modest hardware

A 12B model sounds intimidating, but the quantization options make it accessible:

• GGUF and FP8 quantization brings requirements down to roughly 8GB VRAM — a standard consumer GPU can handle it
• INT8-ConvRot leverages hardware acceleration on NVIDIA Ampere cards (RTX 3090, 3080, and similar), nearly doubling generation speed compared to FP8 on the same hardware — roughly 7.7 seconds per image instead of 14.8

This means you can run a state-of-the-art photoreal model locally without a data-center GPU or ongoing API costs.

Licensing for commercial use

Krea 2 permits commercial use for any business with under $1 million in annual revenue. For independent creators, small studios, and e-commerce brands, this removes a major friction point — you can generate product photography, marketing hero images, and design explorations without per-image licensing fees or restrictive SaaS lock-in.

Automating prompts with VLMs

The source workflow uses a vision-language model to automate prompt writing. Instead of hand-crafting long descriptions, you feed a short idea in plain language and let a VLM expand it into a detailed, structured prompt. Krea 2 rewards this — its Qwen3-VL encoder handles long, specific prompts exceptionally well, and Japanese-language prompts work natively without translation.

A typical automation loop:

1. Write a short concept: "product shot of a ceramic mug on a wooden table, morning light"
2. Let the VLM expand it into a full lighting, material, and camera description
3. Feed the expanded prompt to Krea 2 Turbo for pass one
4. Run the two-pass refine for final output

Bringing it into Phosphene

Phosphene's tag system already handles the structured prompt problem — instead of relying on a VLM to expand a sentence, you compose from typed tags for subject, lighting, lens, and style. That gives you the prompt precision Krea 2 rewards, without leaving the interface. When you combine Phosphene's composable prompts with a two-pass refine pipeline, you get images that hold up under scrutiny — the kind that stop looking AI-generated and start looking intentional.