Generate or edit with your own LoRAs

LoRA Image to Image, one of eight generation types in the image workspace.

Open the image workspace and choose LoRA Image to Image from the Select Generation Type dropdown, alongside the other seven image generation types listed there.

Everything below follows from that choice: pick a base model, supply a source image where the model needs one, add LoRA adapters, then set sizing, strength, output format and a seed to match.

Four models are available: Z-Image Turbo at 10 credits (the default), FLUX.2 Klein at 21, Krea 2 Turbo at 18 for 1k or 30 for 2k resolution, and Qwen Image at 25. Switching models resets the image inputs, sizing fields, strength and output format to that model's own defaults.

The model choice matters beyond price. It decides whether a source image is required, optional, or accepted more than once, whether a prompt is mandatory, and which of the sizing, strength and output-format controls further down actually appear.

Four models, four different credit costs and four different sets of controls below.

Five ways to supply a source image, though not every model needs one.

Five sources are available: upload a file, pick one of your previous completed tasks, select an AI Actor or AI Character, or paste a direct image URL. Which ones you need depends entirely on the model chosen above.

Z-Image Turbo and Qwen Image each require exactly one source image. FLUX.2 Klein accepts up to three optionally, letting the result draw on more than one reference. Krea 2 Turbo accepts zero or one; leave it empty and the model generates from the prompt and LoRA stack alone.

The prompt runs up to 2,500 characters on every model. Krea 2 Turbo and Qwen Image require one; Z-Image Turbo and FLUX.2 Klein do not, since a strong source image and LoRA stack can carry the result without written instructions.

Write it as an instruction for what should change, not a description of the whole picture, the same principle that applies on Image to Image: naming only the change gets a source image respected instead of overwritten.

The prompt. Required on two of the four models, optional on the other two.

Up to three LoRA adapters, each with its own source and scale.

Up to three adapters can be stacked, each independently sourced from your own "My LoRAs" list or an external link to a .safetensors, .bin, .ckpt or .pt file. An adapter picked from My LoRAs shows its trigger word if it has one, as a reminder to include it in the prompt.

Each adapter has its own scale from 0.1 to 1.0, defaulting to 1.0 when added. Lower it if one LoRA is overpowering the others, and keep the stack small when tight control matters, since three styles competing at full strength rarely blend cleanly.

Sizing follows the model. Z-Image Turbo uses explicit width and height from seven presets or a custom 256 to 1536px pair. FLUX.2 Klein and Qwen Image default to matching the source image's size, with a toggle for a custom WIDTHxHEIGHT string instead, capped at 1536x1536 on Qwen. Krea 2 Turbo uses an aspect ratio picker, which disappears once a source image is added since the result then follows the image instead.

Krea 2 Turbo also has a 1k or 2k resolution choice sitting alongside its aspect ratio picker, and unlike the other sizing controls it changes the credit cost directly: 18 credits at 1k, 30 at 2k. The other three models charge a flat price regardless of the size chosen.

Sizing, shaped entirely by the model chosen above.

Transformation strength. On two of the four models, controlling how much survives from the source.

A transformation strength slider from 0 to 1 appears on Z-Image Turbo, defaulting to 0.6, and on Krea 2 Turbo once a source image is added, defaulting to 0.8. It controls how much of the source image survives the edit: low values keep the original mostly intact, high values treat it as loose inspiration only. FLUX.2 Klein and Qwen Image do not expose this control.

Keep the seed fixed while adjusting strength to compare results fairly, the same principle that applies to every other setting on this form. A small move, from 0.6 to 0.7 for instance, can shift the balance more than it looks like it should, so nudge it gradually rather than jumping to the extremes.

Z-Image Turbo and Qwen Image offer a JPEG, PNG or WebP choice for the finished image; FLUX.2 Klein and Krea 2 Turbo do not expose this control at all.

PNG is the safer choice when the result is heading into another editing pass, since it avoids the compression artefacts that build up across repeated JPEG saves. JPEG and WebP are both fine for review and sharing, where a smaller file size matters more than preserving every pixel for a later edit.

Output format. Only on two of the four models.

The seed. Random by default, or fixed to reproduce a result.

The seed defaults to -1 for a fresh result every run. Entering a specific number reproduces the same result later, as long as the prompt, LoRA stack and other settings stay identical, and a dice button next to the field fills in a random one.

This is most useful for comparing changes fairly: fix the seed, then adjust one setting at a time, such as a LoRA's scale, the strength slider or the prompt, to see what actually moved.