Krea Releases Krea 3, Beating GPT Image 2.5 at 11x Lower Latency

Krea has published a technical report for Krea 3, its image-to-image foundation model. The report details the architecture, training data, captioning pipeline, reinforcement learning system, production latency, and serving cost behind the release.

Krea reports that Krea 3 Ultra debuted fourth on the Arena, Artificial Analysis, and Design Arena editing leaderboards. When results are grouped by developer, Krea ranks second behind the GPT-Image-2.5 family. Its production benchmark gives Krea 3 Fast a mean latency of 8 seconds, compared with 22.3 seconds for Nano Banana Pro and 90.2 seconds for GPT Image 2.5 Sunburst at its maximum setting.

References change the editing task

Text-to-image systems generate from instructions alone, which limits their ability to preserve a specific person, product, layout, or visual identity. Image-to-image models accept reference images alongside the instruction. Krea 3 can combine as many as eight references, preserve characters and products across edits, and restrict changes to requested regions.

Creative workflows often require dozens of iterations before a designer chooses a result. Lower latency allows those iterations to remain inside an interactive workflow, while per-edit cost determines whether the same model can support high-volume design tools, commerce pipelines, and asset production.

A familiar architecture, tuned hard

Krea retained a conventional model architecture and concentrated its work on pretraining, data, and post-training. The team initialized Krea 3 from the Krea 2 checkpoint, expanded the training dataset, and added native 2K samples. Its image-to-image extension has two main components:

• Reference images pass through a variational autoencoder, or VAE, which converts pixels into compressed latent tokens. Those tokens are appended to the diffusion transformer sequence with the timestep fixed at t=0 , so the model always treats them as clean references. This supports key-value caching during causal distillation, allowing reference attention calculations to be reused.
• A vision-language model serves as the text encoder. It receives the edit instruction and reference images resized to 384 × 384 pixels, allowing instruction tokens to attend to visual context. The diffusion transformer receives the resulting instruction tokens, while higher-fidelity VAE latents carry the reference imagery.

2K resolution sends memory soaring

Native 2K training mainly improves fine text. At 1K resolution, small lettering can disappear during the VAE’s encode-decode cycle, leaving the model without a reliable target to learn. At 2K, those characters remain represented. A 2048 × 2048 target paired with seven references at the same resolution produces 131,072 tokens, 32 times the sequence length of a 1024 × 1024 text-to-image sample.

Krea controls that memory load with context parallelism, which distributes long sequences across devices; fully sharded data parallelism, which splits model and optimizer state; gradient checkpointing, which recomputes selected activations; and bf16 AdamW moments, which reduce optimizer memory. The settings vary by sequence length so shorter training examples avoid unnecessary coordination and recomputation. Reference and target resolutions can also differ, allowing a high-resolution product-placement output to use cheaper, lower-resolution references.

How billions of images become edit pairs

Krea built a cascaded pipeline to find related but nonidentical image pairs within a pretraining pool containing billions of images. Cheap similarity models remove most candidates before larger vision-language models inspect the survivors:

1. SigLIP 2 similarity retrieves candidate pairs.
2. A DINOv3 CLS-token filter retains 45% of those candidates.
3. Qwen3.5-9B checks whether each pair can form a useful reference and target, retaining about 4.7% of the pairs it reviews.
4. Qwen3.6-27B audits the remaining pairs and retains roughly 45%.
5. Target-level deduplication retains 18% of the audited set.

Multiplying those retention rates leaves 0.17% of the initial SigLIP 2 candidates. Because inexpensive embedding models handle the widest stages, Krea can run the pipeline as low-priority work on otherwise idle cluster capacity.

Naturally grouped sources, including Wikipedia galleries and albums, recover relationships that embedding similarity can miss, such as a floor plan paired with a rendering of the completed building. Synthetic datasets add rendered HTML and CSS edits for localized text, novel 3D viewpoints, and ControlNet-style depth, pose, and edge annotations.

Captions must account for every change

Krea found that vision-language models often produced incomplete training instructions, invented changes, or ignored references. Its captioning pipeline requires four structured outputs before the final caption:

• A comprehensive OCR transcription.
• A global description of the target image.
• A dependency map linking target elements to their source references.
• An explicit list of every change between the references and target.

Ablation tests found that training on a mixture of long and short prompts performed worse than training on detailed prompts alone, including evaluations that used detailed instructions. Implicit requests such as “a sign about rocket science” provide no exact character sequence to learn and can reward plausible-looking gibberish.

RL targets repeatable edits

Krea’s reinforcement learning stage optimizes reliability on existing editing tasks. For each training input, the system samples 16 candidate edits, called rollouts, and a collection of vision-language agents selects task-specific rewards. Before optimization, planner agents can generate three types of evaluation plans:

• A rubric containing binary visual checks.
• A text plan listing every string that must appear or remain intact.
• A region plan containing bounding boxes for localized edits.

Pointwise rubric judges eventually produced too little variation. Near the end of an RL run, about 70% of rollouts received a perfect score, and more than half of the 16-sample groups had identical rewards. Krea replaced that signal with pairwise comparisons. A full comparison among 16 outputs requires 120 judgments; its pivot method selects one random rollout and compares it with the other 15, approximating the full ranking at linear cost.

When text rewards backfire

OCR systems struggled with stylized fonts, so Krea used a vision-language model as the text reader. Showing that reader the original instruction caused it to autocorrect generated errors, including interpreting “eleghaant” as “elephant,” so the reward model now reads the output without the instruction. The generator then learned to increase its score by producing larger, bolder, and plainer text. A size penalty drove text into tiny captions. Krea’s final reward combines a typography judge, which evaluates edited text against the image’s overall design, with a size constraint measured against a fixed anchor.

Diffusion outputs also showed limited diversity across random seeds. On a 97-task stress set, one sample per task solved none of the tasks. Sixteen samples with different initial noise produced at least one correct result for seven tasks. Constrained instruction paraphrases solved almost twice as many hard cases as seed variation alone, so an augmentation agent rewrites the instruction for half of each rollout group.

Measured gains from RL

On held-out public text-editing benchmarks, Krea reports that reinforcement learning reduced text errors by roughly 25%. On CPI-Bench, its judge recorded a 17% reduction in flawed generations overall and a 27% reduction in the benchmark’s “intelligent” category. Krea says the CPI-Bench judge was absent from reward training.

Qualitative evaluations showed narrower, more accurate edits: the model translated the requested poster lines, preserved a man while removing the referenced woman, and changed leaf ornaments on a pan handle while preserving the rest of the handle.

Latency and cost, by model

Krea’s production benchmark reports mean latency and estimated cost for 100,000 edits across five models:

At the reported rates, 100,000 Krea 3 Fast edits cost $7,000 less than Nano Banana Pro and $35,660 less than GPT Image 2.5 Sunburst. Mean latency is 14.3 seconds lower than Nano Banana Pro and 82.2 seconds lower than Sunburst.

Quality requires a separate comparison because the leaderboard placement belongs to Krea 3 Ultra, while the latency and cost table names Krea 3 Fast. Teams evaluating the API should benchmark the exact configuration they plan to deploy.

Known limits

• Dense small text: Extremely small or tightly packed lettering remains unreliable.
• Identity consistency: Facial identity can drift across edits. Krea plans to support character- and product-specific LoRAs, which are compact fine-tuning adapters.
• Interactive latency: The fastest configuration completes an edit in about 4 to 5 seconds, while Krea 3 Fast averages 8 seconds in the reported benchmark. Continuous, frame-by-frame editing remains out of reach.

What to benchmark before shipping

Product teams building design tools, commerce systems, or game-asset workflows should test instruction adherence, text accuracy, identity preservation, tail latency, and cost on their own requests. Krea’s reported results show how training data, structured captions, reward design, and serving optimizations can improve those deployment metrics without changing the model’s basic architecture.

The model is available through the Krea 3 app and the Krea API .