--- title: "How to Train a Krea 2 LoRA (Character & Style)" description: "A practical guide to training a LoRA for Krea 2 Turbo — dataset size, the required training adapter, epochs, prose captions, and why LoRAs are architecture-specific." slug: how-to-train-krea-2-lora published: 2026-08-24 updated: 2026-08-24 tags: ["Krea 2", "LoRA", "training", "how-to"] tldr: - "Krea 2 LoRAs must be trained on the Krea 2 architecture — a Krea 2 LoRA is silently ignored on any other model, and vice versa." - "Aim for ~100 images and 20–30 epochs; more epochs tends to overfit." - "Training needs the Krea 2 Turbo training adapter and works well with ai-toolkit or OneTrainer's krea2 preset." - "Write captions in natural prose (Krea 2 reads an LLM), lead each with your trigger word, and never describe the face for a character LoRA." - "AIFLUX can train a Krea 2 LoRA for you from a photo set — no local GPU or config needed." reading_time: 7 --- A **LoRA** is a small add-on that teaches a base model a specific character, product, or style without retraining the whole thing. Krea 2 supports LoRAs well — but training one has a few Krea-2-specific rules that, if you miss them, produce a file that *looks* fine and does *nothing*. Here's how to get it right. ## Rule #0: Krea 2 LoRAs are architecture-specific This trips up almost everyone. **A LoRA trained on Krea 2 only works on Krea 2.** Load a Krea 2 LoRA on a non-Krea-2 model (or the reverse) and it is ignored **silently** — the output looks identical to no LoRA at all, and the logs quietly say the LoRA keys weren't loaded. If a LoRA "does nothing," the cause is almost always the wrong base model or a truncated download, **not** a broken file. Check the base architecture first, and verify the file's checksum/size. ## Dataset: quality and size - **Target ~100 images.** Smaller sets (say 40–50) can work but tend to produce a weaker, less consistent identity. Around 100 varied images is the sweet spot for a character. - **Vary everything except the subject** — different outfits, poses, lighting, backgrounds, distances. You want the model to learn *the subject*, not a single photo's context. - **Consistent subject, inconsistent surroundings** is the goal. ## Captions: prose, not tags (this matters on Krea 2) Because Krea 2's text encoder is a language model, **caption your dataset in natural sentences**, not comma tags. For a **character** LoRA: - **Start every caption with your trigger word** (e.g. `mychar, ...`). - Describe **outfit, pose, lighting, background** in plain prose. - **Do not describe the face or fixed features.** You want those baked into the trigger word, not tied to caption words — describing them teaches the model to only produce the face when those words appear. For a **style** LoRA, describe the content normally and let the consistent aesthetic across the set carry the style. ## Training settings that work A reliable Krea 2 Turbo LoRA recipe: - **Architecture:** `krea2:turbo`, with the **Krea 2 Turbo training adapter** loaded (training won't produce a usable Krea 2 LoRA without it). - **Rank / alpha:** 16 / 16, training the transformer only. - **Optimizer / LR:** AdamW 8-bit, learning rate ~1e-4, batch size 1. - **Precision:** bf16 with qfloat8 quantisation. (Note: Krea 2 has no low-bit training path — don't reach for 3/4-bit training quantisation here.) - **Resolution:** multi-resolution buckets, e.g. 512 / 768 / 1024. - **Regularisation:** ~5% caption dropout; cache the text embeddings to speed up training. - **Checkpoints:** save **every epoch** with fixed-seed sample images so you can compare. **Epochs:** aim for **20–30**. Krea 2 identity LoRAs overfit if you push much past that — the model starts reproducing dataset artifacts instead of generalising. The best practice is to **checkpoint every epoch and pick the winner by eye**, comparing samples generated at the *same prompt and seed* across epochs. Tooling-wise, both **ai-toolkit** (headless) and **OneTrainer** (which ships a `krea2` preset) handle this well. ## Using the LoRA Once trained, put the **trigger word at the very start of the prompt**, then describe the scene in prose. If you're also using body or style sliders, keep them near neutral when a strong character LoRA is active — the LoRA already knows the subject, and stacking too many competing LoRAs muddies the identity. ## The shortcut: train a Krea 2 LoRA on AIFLUX If you'd rather skip the adapter, the CUDA setup, and the epoch-picking, **AIFLUX trains Krea 2 LoRAs for you**. You upload a photo set, it auto-captions and trains on a cloud GPU with the recipe above, then hands you an **epoch-selection grid** — same-seed previews for each checkpoint — so you pick the best version. The finished model drops straight into the generator. 👉 **[Train a Krea 2 model on AIFLUX](/create)** ## Frequently asked questions **Why does my Krea 2 LoRA do nothing?** Almost always the wrong base model (Krea 2 LoRAs only work on Krea 2) or a truncated download. Verify the architecture and the file checksum. **How many images do I need to train a Krea 2 LoRA?** Around 100 for a solid character. Fewer can work but yield a weaker identity. **How many epochs should I train?** 20–30. Checkpoint every epoch and pick the best by comparing same-seed samples; more epochs usually overfit. **Should captions be tags or sentences?** Sentences. Krea 2's encoder is an LLM, so prose captions (starting with the trigger word) train better than tag lists. **Can I train a Krea 2 LoRA without a GPU?** Yes — AIFLUX runs the training on cloud GPUs and returns an epoch picker so you choose the best checkpoint. --- *Train your own Krea 2 model:* **[Start on AIFLUX →](/create)**