API setup required
Krea 2
Integrated with fal Krea 2 Trainer. Use the completed weights with Krea 2 Turbo LoRA in Generate. Requires a configured Fal API Key.
ImageExplore
Fine-tuned models / Base models
Pick a base model to fine-tune.
Compare the base models you can fine-tune with LoRAs on Loras AI: FLUX, Krea 2, Qwen Image, Z-Image, Wan 2.2 and LTX-2. See what each does well and which you can train today.
Quick answer: Train supported image and video LoRAs with configured services, sufficient credits and a compatible dataset. Each LoRA must use its paired generator; catalog listings do not establish compatibility.
Cloud training is integrated for Krea 2, FLUX.1, FLUX.2 Dev, Qwen Image, Z-Image, Wan 2.2 I2V and LTX-2 19B. Configure the required services, sign in and check the quoted credit cost before submitting.
One workspace. Multiple model families.
Choose a model for your concept. Check availability before preparing a training run.
API setup required
FLUX.1 Dev
FLUX.1 Dev image adapters. Training and compatible inference still need end-to-end validation.
ImageExplore
API setup required
FLUX.2 Dev
Fal FLUX.2 Dev V2 training and paired LoRA inference, with R2 weight archival. FLUX.1 weights are not interchangeable.
ImageExplore
API setup required
Qwen Image 2512
Qwen Image 2512 adapters. Keep the exact base-model version with your training configuration.
ImageExplore
API setup required
Z-Image Turbo
Z-Image training candidate. Base / Turbo compatibility must be tested before release.
ImageExplore
API setup required
Wan 2.2
Train from MP4 clips; compatible I2V generation preserves separate high- and low-noise LoRA weights.
VideoExplore
API setup required
LTX-2 19B
Video LoRA training and text-to-video inference for LTX-2 19B. LTX-2.3 22B is a different model and cannot reuse this pairing.
VideoExplore
Planned
FLUX.2 Klein 9B
A planned FLUX.2 Klein 9B training workflow.
ImageNot open
Planned
FLUX.1 Kontext
Paired before / after editing data requires a dedicated workflow.
ImageNot open
Planned
Wan 2.1
A separate video adapter family; availability still needs verification.
VideoNot open
Planned
SDXL
Trainer availability, licensing and output compatibility are under review.
ImageNot open
Planned
Stable Diffusion 1.5
A legacy model family. Training and generation compatibility need review.
ImageNot open
Under review
Krea 3
Research candidate. No public LoRA training integration is claimed.
ImageNot open
Under review
Nano Banana 2
Generation popularity does not establish LoRA training support.
ImageNot open
Under review
GPT Image 2
LoRA training availability has not been established for this catalog entry.
ImageNot open
Under review
Kling
Video generation and LoRA training are distinct capabilities.
VideoNot open
Under review
Veo
Training availability and permitted adapter workflows remain under review.
VideoNot open
Under review
Krea 2 Medium
Seen in the reference workflow. This version is listed for comparison; its training and inference integration has not been validated.
ImageNot open
Under review
Krea 2 Large
Seen in the reference workflow. This version is listed for comparison; its training and inference integration has not been validated.
ImageNot open
Under review
Flux Realtime
Seen in the reference workflow. This version is listed for comparison; its training and inference integration has not been validated.
ImageNot open
Under review
LTX-2.3 22B
Seen in the reference workflow. This version is listed for comparison; its training and inference integration has not been validated.
VideoNot open
Catalog entries are not live training integrations. No model is marked ready until training, weight retrieval and compatible generation have been tested together.
Training capabilities and requirements
Costs and limits below follow the application configuration. Integration describes the code, not verified provider uptime or a successful live training run.
| Model | Integration | Cloud input limits | Credits and settings | Output and compatibility |
|---|---|---|---|---|
| Krea 2 | Integrated; configuration required | 1–30 JPEG, PNG or WebP images with captions and a trigger phrase. Up to 15 MiB per image; 25 MiB total. | Training: 100 credits / task. Generation: 10 credits / image. Fixed training steps: 100. | LoRA weights after success; use the paired Krea 2 generator. Compatibility with other base models is not implied. |
| FLUX.1 Dev | Integrated; configuration required | 4–30 JPEG, PNG or WebP images with captions and a trigger phrase. Up to 15 MiB per image; 25 MiB total. | Training: 167 credits / task. Generation: 10 credits / image. Fixed training steps: 500. | LoRA weights after success; use the paired FLUX.1 Dev generator. Compatibility with other base models is not implied. |
| FLUX.2 Dev | Integrated; configuration required | 10–30 JPEG, PNG or WebP images with captions and a trigger phrase. Up to 15 MiB per image; 25 MiB total. | Training: 2134 credits / task. Generation: 14 credits / image. Fixed training steps: 1000. | LoRA weights after success; use the paired FLUX.2 Dev generator. Compatibility with other base models is not implied. |
| Qwen Image 2512 | Integrated; configuration required | 4–30 JPEG, PNG or WebP images with captions and a trigger phrase. Up to 15 MiB per image; 25 MiB total. | Training: 167 credits / task. Generation: 10 credits / image. Fixed training steps: 500. | LoRA weights after success; use the paired Qwen Image 2512 generator. Compatibility with other base models is not implied. |
| Z-Image Turbo | Integrated; configuration required | 4–30 JPEG, PNG or WebP images with captions and a trigger phrase. Up to 15 MiB per image; 25 MiB total. | Training: 209 credits / task. Generation: 10 credits / image. Fixed training steps: 500. | LoRA weights after success; use the paired Z-Image Turbo generator. Compatibility with other base models is not implied. |
| Wan 2.2 | Integrated; configuration required | 3–10 MP4 clips with a trigger phrase. Up to 8 MiB per clip; 24 MiB total. | Training: 1667 credits / task. Generation: 67 credits / 5-second video. Fixed training steps: 100. | LoRA weights after success; use the paired Wan 2.2 generator. Compatibility with other base models is not implied. |
| LTX-2 19B | Integrated; configuration required | 3–10 MP4 clips with a trigger phrase. Up to 8 MiB per clip; 24 MiB total. | Training: 584 credits / task. Generation: 34 credits / 5-second video. Fixed training steps: 500. | LoRA weights after success; use the paired LTX-2 19B generator. Compatibility with other base models is not implied. |
Cloud tasks require sign-in and sufficient credits. Krea 2 needs Fal; WaveSpeed image and video trainers need WaveSpeed access and R2 storage.
Cloud training sends your selected media and captions to the provider. Image and video dataset limits differ. Confirm account readiness before submitting.
A focused workflow
How to read the model list
Every fine-tuned LoRA is tied to one base model. Compare them side by side so you fine-tune on the one that fits your idea.
Family and version
A LoRA trained on one version may not load on another. Note both before you choose.
Image or video
Wan 2.2 and LTX-2 are video families; the rest produce still images.
Training status
Integrated families can be trained here; others are listed for reference.
Cost before you start
Credit costs for training and generation are shown before you submit anything.
Ideas worth teaching
Examples by model family
Explore three ways to define a training concept. Each reference links to a dedicated workflow, with dataset advice and evaluation prompts.
A style that travels.
Use this coastal illustration as a visual-language reference. A style dataset should vary the subject while keeping its materials, colors and mark-making coherent.
Caption exampleA coastal village in ohwx_style, gouache texture, warm flat colors and fine ink contours.
Explore style LoRA training
One character. New settings.
An original ceramic explorer gives a dataset a clear identity. Vary the angle, pose and surroundings instead of repeating one composition.
Caption exampleohwx_robot, a ceramic explorer with blue glass eyes, standing beside an alpine lake in daylight.
Explore character LoRA training
Keep the product recognizable.
A blank amber bottle makes shape and materials easy to discuss. Include alternate views of its glass body and cap rather than teaching a single background.
Caption exampleohwx_bottle, an amber glass bottle with a round walnut cap, on a travertine surface in window light.
Explore product LoRA trainingThese original images were AI-generated as concept references. They are not Loras AI training outputs, before/after comparisons or evidence of model performance.
From intent to evaluation
Choosing a model family in four steps
Prepare the image or video dataset required by the selected model, review its trigger and captions, then submit training. Use the completed weights with their matching generator.
- 1
Choose one concept
Decide whether you want a recognizable character, a transferable style, a distinctive product or a specific motion.
Read the note - 2
Collect useful examples
Keep the intended concept consistent while varying details you do not want the model to memorize.
Read the note - 3
Review captions & export
Inspect file dimensions, edit each caption and export the local configuration before leaving the workspace.
Read the note - 4
Plan your test prompts
Test completed, compatible LoRA weights with new prompts. Record the model, settings, successes and failures, and compare with the base model to evaluate the concept.
Read the note
Dataset & compatibility
Before you train on a model family
Image, paired-edit and video LoRAs are different kinds of files. Status labels show what is integrated today; they are not a promise that every dataset will be accepted.
Read the dataset preparation note- Pick the family by the look you want, not by its name.
- Check the version a LoRA was trained on.
- Use a video family only when you need motion.
- Compare two families with the same prompt.
Prompt notebook
Prompts for comparing model families
Save these ideas for evaluating a compatible trained adapter later. They are prompt examples, not generated results or a promise of consistent output.
Architecture
A hillside library in ohwx_style, warm flat colors, textured paper and delicate ink contours.
An everyday object
A vintage desk lamp in ohwx_style, simple composition, warm ochre and indigo palette.
A natural scene
A quiet forest path after rain in ohwx_style, layered shapes, visible gouache texture.
A different subject
A small sailboat in ohwx_style, bold simplified forms, soft paper grain, no lettering.
A character study
An original fictional gardener in ohwx_style, full-body pose, simple pale background.
A new palette test
A night market in ohwx_style, restrained blue and amber palette, hand-painted details.
Related notes
A little preparation goes a long way.
Practical reading on image selection, captions, training settings and evaluating what a model actually learned.
How many images do you need?Choose useful variation rather than relying on one universal image count.A practical LoRA captioning guideSeparate the concept from changing details, with character, style and product examples.Training steps without the guessworkUnderstand steps, learning rate and overfitting before changing advanced settings.How to evaluate a trained LoRATest new prompts, compare a baseline and document useful failures.
FAQ
LoRA model families: questions
Model-specific guidance, with the current availability stated clearly.
Which base model should I fine-tune?
Start from the result you want. Pick an image model for characters, styles or products, and Wan 2.2 or LTX-2 when you need video. Compare a few examples before you commit.
Can one fine-tuned LoRA work across base models?
No. A LoRA only loads in the base model family and version it was fine-tuned on. To use the same subject elsewhere, fine-tune a separate LoRA for that model.
Can I submit a paid training job on this page?
Krea 2, FLUX.1, FLUX.2 Dev, Qwen Image, Z-Image, Wan 2.2 I2V and LTX-2 19B have training and paired generation integrations. Paid tasks require configured services, sign-in and sufficient credits. Other catalog models remain previews.
What happens to the images I add?
Local draft preparation processes images in the current browser tab; exported JSON excludes source files. Refreshing or leaving clears local images. Submitting cloud training sends the selected dataset and captions to the provider; WaveSpeed training also uses configured R2 storage for datasets and archived artifacts. Storage privacy depends on bucket access settings. Generation sends prompts and selected LoRA references to the relevant service. Only submit material you have permission to use.
Can I reuse one LoRA with every image or video model?
Adapters are not interchangeable. Match the model architecture, base version and loading workflow. Krea 2, MOSS and Anima need their own compatible weights; similar names do not imply compatibility. Video workflows may also require multiple adapter files.
Keep exploring
Go deeper into one base model
Compare related model families before choosing a training workflow. Adapter formats and base-model versions must remain compatible.
Picked a base model? Fine-tune your LoRA on it
Choose a compatible model, prepare images and captions, then submit an available training or generation workflow. You can also export a dataset configuration draft for free.