# Loras AI

> Train custom LoRA models for consistent characters, products and styles

## Product and reference

- [Loras AI](https://loras.im/): Custom LoRA training, image generation and dataset preparation.
- [Detailed product reference](https://loras.im/llms-full.txt): Configuration-derived costs, limits, compatibility and availability boundaries.
- [Training workspace](https://loras.im/train): Prepare image or video datasets; cloud LoRA training requires sign-in, credits and configured services.
- [Model directory](https://loras.im/models): Compare model families; a directory listing does not mean cloud training is available.
- [Pricing](https://loras.im/pricing): Subscriptions and one-time credit packs; proposed tier benefits are marked as previews.
- [Training guides](https://loras.im/guides): Public guides covering preparation, captions, steps and evaluation.
- [Blog](https://loras.im/blog): Published articles.

## Model training references

- [Krea 2 LoRA](https://loras.im/krea-2-lora-training): Fal training and paired inference integration; use the detailed reference for requirements and costs.
- [FLUX.1 Dev LoRA](https://loras.im/flux-lora-training): WaveSpeed training integration with compatible inference; requires configured API credentials and R2 storage.
- [FLUX.2 Dev LoRA](https://loras.im/flux-2-lora-training): Fal training and paired inference integration; use the detailed reference for requirements and costs.
- [Qwen Image 2512 LoRA](https://loras.im/qwen-image-lora-training): WaveSpeed training integration with compatible inference; requires configured API credentials and R2 storage.
- [Z-Image Turbo LoRA](https://loras.im/z-image-lora-training): WaveSpeed training integration with compatible inference; requires configured API credentials and R2 storage.
- [Wan 2.2 LoRA](https://loras.im/wan-2-2-lora-training): WaveSpeed training integration with compatible inference; requires configured API credentials and R2 storage.
- [LTX-2 19B LoRA](https://loras.im/ltx-2-lora-training): WaveSpeed training integration with compatible inference; requires configured API credentials and R2 storage.

## Training objectives

- [character LoRA training](https://loras.im/character-lora-training): Dataset and workflow guidance for character concepts; compatibility depends on the selected model.
- [style LoRA training](https://loras.im/style-lora-training): Dataset and workflow guidance for style concepts; compatibility depends on the selected model.
- [product LoRA training](https://loras.im/product-lora-training): Dataset and workflow guidance for product concepts; compatibility depends on the selected model.
- [video LoRA training](https://loras.im/video-lora-training): Wan 2.2 I2V and LTX-2 19B video LoRA training and compatible inference; configuration required.

## Guides

- [How to train a LoRA](https://loras.im/guides/how-to-train-a-lora): How to train a LoRA: concept, dataset, captions and evaluation.
- [How many images to use and how to choose useful variation.](https://loras.im/guides/how-many-images-for-lora-training): How many images to use and how to choose useful variation.
- [Captioning examples for characters, styles and products.](https://loras.im/guides/lora-captioning-guide): Captioning examples for characters, styles and products.
- [Training steps, learning rate and overfitting explained.](https://loras.im/guides/lora-training-steps-explained): Training steps, learning rate and overfitting explained.
- [Evaluate adapters with new prompts, baselines and failure cases.](https://loras.im/guides/how-to-evaluate-a-lora): Evaluate adapters with new prompts, baselines and failure cases.

## Languages and policies

Public training pages and guides are available in English and Chinese. Japanese and Korean are currently homepage-only.
- [中文首页](https://loras.im/zh): Chinese homepage; public training and guide paths also have a /zh prefix.
- [日本語ホーム](https://loras.im/ja): Japanese homepage.
- [한국어 홈페이지](https://loras.im/ko): Korean homepage.
- [Privacy policy](https://loras.im/privacy-policy): Read the published privacy terms before uploading material.
- [Terms of service](https://loras.im/terms-of-service): Published terms and usage conditions.
- [Sitemap](https://loras.im/sitemap.xml): Public page URLs and supported language alternates.

## Product capabilities and availability

Loras AI helps creators prepare image datasets, train custom Krea 2, FLUX.1, Qwen Image 2512 and Z-Image Turbo LoRAs and generate images with compatible adapters.
Integrated means application code exists. It does not verify deployed credentials, provider uptime or a successful end-to-end run. Paid operations require sign-in, sufficient credits and the corresponding configured provider.
Model directory statuses are catalog labels, not live service health or purchase entitlements.

### Krea 2 training and image generation — integrated, conditional

Training: 100 credits per submitted training task. Image generation: 10 credits per image task.
Training input: 1–30 JPEG, PNG or WebP images with captions and a trigger phrase. Maximum 15 MiB per image and 25 MiB combined image bytes.
Training settings sent by the application: 100 steps, resolution 768. Advanced reference controls do not imply configurable provider settings.
Training output: LoRA weights and configuration links after successful completion. The application can generate one PNG using a successful Krea 2 training task owned by the signed-in user.
Generation sizes: square_hd, portrait_4_3, landscape_4_3, landscape_16_9, portrait_16_9.
Krea 2 training weights are used by the Krea 2 generator; compatibility with MOSS, Anima or other base models is not implied.
Source images and captions are sent to the cloud provider when cloud training is submitted. Local preparation is distinct from cloud upload; provider links are not a promise of private permanent storage.
When R2 is configured, newly completed Krea training weights and configuration are archived. Existing provider-hosted records remain accessible through their original links.

### FLUX.1, Qwen Image 2512 and Z-Image Turbo — integrated, conditional

- FLUX.1 Dev: 500 fixed training steps; 167 credits per training; 10 credits per generated image. Training endpoint wavespeed-ai/flux-dev-lora-trainer; paired inference wavespeed-ai/flux-dev-lora.
- Qwen Image 2512: 500 fixed training steps; 167 credits per training; 10 credits per generated image. Training endpoint wavespeed-ai/qwen-image-2512-lora-trainer; paired inference wavespeed-ai/qwen-image/text-to-image-2512-lora.
- Z-Image Turbo: 500 fixed training steps; 209 credits per training; 10 credits per generated image. Training endpoint wavespeed-ai/z-image-lora-trainer; paired inference wavespeed-ai/z-image/turbo-lora.
Requires the configured WaveSpeed account and R2 storage. Input: 4–30 JPEG, PNG or WebP images, at most 15 MiB each and 25 MiB total. Images and captions are sent to WaveSpeed on submission.
Trained weights and configuration are archived in R2 before success is reported. Owned successful records can be downloaded through authenticated routes or used with their paired inference endpoint. Provider access uses expiring signed URLs; storage privacy also depends on the bucket access configuration.
Public HTTPS deployments with an AUTH_SECRET receive signed per-job callbacks that trigger authoritative provider queries. Authenticated polling provides fallback synchronization. No automatic resubmission or refund occurs while submission is uncertain.

### MOSS Image — integrated, conditional

Input: text prompt up to 10000 characters, aspect ratio, resolution and 1–3 compatible trained LoRA adapters with positive strength. Output: one image after successful completion.
Credits per task by resolution: 1k: 10; 1.5k: 20; 2k: 40.
Aspect ratios: 1:1, 1:2, 2:1, 1:3, 3:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 9:21, 21:9.
Output format choices: jpeg, png, webp.
Required adapters need MOSS compatibility confirmation and a supported weight reference. A supported URL does not prove model compatibility.

### Anima — integrated, conditional

Input: text prompt up to 10000 characters, aspect ratio, resolution and 1–3 compatible trained LoRA adapters with positive strength. Output: one image after successful completion.
Credits per task by resolution: 1k: 10; 1.5k: 20.
Aspect ratios: 1:1, 1:2, 2:1, 1:3, 3:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 9:21, 21:9.
Required adapters need Anima compatibility confirmation and a supported HTTPS weights URL. MOSS or Krea adapter compatibility is not implied.

### Other models and workspaces

- FLUX.2 Klein 9B: catalog status planned; no cloud training integration in this application.
- FLUX.1 Kontext: catalog status planned; no cloud training integration in this application.
- Wan 2.1: catalog status planned; no cloud training integration in this application.
- SDXL: catalog status planned; no cloud training integration in this application.
- Stable Diffusion 1.5: catalog status planned; no cloud training integration in this application.
- Krea 3: catalog status research; no cloud training integration in this application.
- Nano Banana 2: catalog status research; no cloud training integration in this application.
- GPT Image 2: catalog status research; no cloud training integration in this application.
- Kling: catalog status research; no cloud training integration in this application.
- Veo: catalog status research; no cloud training integration in this application.
- Krea 2 Medium: catalog status research; no cloud training integration in this application.
- Krea 2 Large: catalog status research; no cloud training integration in this application.
- Flux Realtime: catalog status research; no cloud training integration in this application.
- LTX-2.3 22B: catalog status research; no cloud training integration in this application.
- /image (Loras AI Image): text-to-image and image editing with up to 5 reference images, ratios 1:1, 2:3, 3:2, 9:16, 16:9; 10 credits per generation. No trained LoRA required.
- /video (Loras AI Video): text-to-video and image-to-video with sound, 480p or 720p, 6 or 10 seconds; 480p 24 / 40 credits; 720p 45 / 75 credits.
- /image FLUX.2 Pro · 1K: text and image inputs, maximum 5 reference images. 9 credits per task.
- /image GPT Image 1.5 · Medium: text and image inputs, maximum 5 reference images. 10 credits per task.
- /image Imagen 4 Fast: text input only, maximum 0 reference images. 10 credits per task.
- /image Ideogram V3 · Balanced: text and image inputs, maximum 1 reference images. 12 credits per task.
- /video Seedance 2.5: text and image inputs, maximum 2 reference images. Durations 4/5/6/8/10/15/20/30 seconds; resolutions 480p/720p/1080p. Prices are quoted before submission.
- /video Veo 3.1 Fast · 720p: text and image inputs, maximum 2 reference images. Durations 4/6/8 seconds; resolutions 720p. Prices are quoted before submission.
- /video Kling 2.6 · Audio: text and image inputs, maximum 1 reference images. Durations 5/10 seconds; resolutions 1080p. Prices are quoted before submission.
- /video Runway Gen-4 Turbo · 720p: text and image inputs, maximum 1 reference images. Durations 5 seconds; resolutions 720p. Prices are quoted before submission.
- Wan 2.2 I2V video LoRA: 100 training steps, 1667 training credits; 67 credits per 5-second video. Requires WaveSpeed and R2. Inference requires a start image and preserves separate high/low noise adapters.
- LTX-2 19B video LoRA: 500 training steps, 584 training credits; 34 credits per 5-second video. Requires WaveSpeed and R2. Text-to-video inference with owned compatible weights.
- FLUX.2 Dev: 1000 training steps, 2134 credits per training; 14 credits per bounded image preset. Requires Fal and R2. Trainer fal-ai/flux-2-trainer-v2, paired inference fal-ai/flux-2/lora.
Video training input: 3–10 MP4 clips, 8 MiB each, 24 MiB total. Successful training weights are archived before becoming available for inference.
- /generate: use trained Krea 2, FLUX.1 Dev, Qwen Image 2512, Z-Image Turbo, Wan 2.2 I2V or LTX-2 19B LoRAs, or MOSS / Anima with compatible LoRAs.
Failed generation tasks refund their credits. Completed video training awaiting artifact review or archival is not automatically refunded or resubmitted.
Concept illustrations are disclosed as references, not measured results of this training pipeline. No success rate, completion time or quality guarantee is asserted.

## Credit plans and billing

The following products are advertised on the pricing page. Amounts below are full checkout totals in the listed currency, not monthly equivalents of annual totals. Credit quantities are the catalog grants for that purchase or billing period.
| Product | Billing | Checkout total | Credits |
| --- | --- | --- | --- |
| Starter | 1 month | USD 19.90 | 1000 |
| Pro | 1 month | USD 39.90 | 2200 |
| Studio | 1 month | USD 79.90 | 3200 |
| Starter | 1 year | USD 99.99 | 12000 |
| Pro | 1 year | USD 199.99 | 26400 |
| Studio | 1 year | USD 279.99 | 38400 |
| Studio Credit Pack | one-time credit pack | USD 39.99 | 6500 |
Annual plans are billed as annual totals. One-time credit packs do not create recurring subscriptions; legacy product identifiers containing lifetime do not promise lifetime model access.
Proposed tier-specific model access, training image limits, concurrency, queue and batch benefits are design previews, not enforced entitlements. Purchasing credits does not enable an unavailable model.
Payment methods depend on site configuration. Confirm the current price and applicable terms on the pricing and checkout pages. Credit expiration and account-specific grants must be checked in the account; no unlimited validity is promised here.