Qwen-Image-Edit

Qwen-Image-Edit takes an existing photo and a text instruction and produces a modified version that keeps everything the instruction didn't ask to change — a face, a pose, a background — intact. It handles fine-grained object edits, style transfer, and semantic transformations like changing a scene's viewpoint.

On AutorunX it's positioned as the budget-conscious alternative in services where GPT Image 2 or Nano Banana are the polished default: it's selectable in the model picker for AI Influencer persona shots and Banner Ad edits when identity-preserving editing at a lower credit cost is the priority.

Qwen-Image-Edit pairs a 20-billion-parameter multimodal diffusion transformer (MMDiT) with a Qwen2.5-VL vision-language model for text conditioning. The VL component reads and reasons about the source image before the diffusion transformer edits it, which is what lets the model make fine-grained, semantically aware edits — changing one object, transferring a style, or preserving a face — instead of regenerating the whole scene.

AutorunX runs it on two tracks. The owned track, internally called AX-QWN, is the Apache 2.0-licensed open-weight model self-hosted on AutorunX's own RunPod GPU fleet — no per-call fee to a third party, so it's one of the cheapest models on the platform. A separate proprietary hosted tier, Qwen-Image-Edit-Plus, is accessed through a different provider's API rather than self-hosted, and requires a reference image to lock in exact-identity edits.

Both tracks share the same core strength: because the weights are Apache 2.0-licensed and openly published (unlike closed competitors), the underlying model can be inspected, fine-tuned, and deployed independently of any single vendor's API — which is exactly why AutorunX was able to bring it in-house as AX-QWN instead of only ever renting it.