Pro Realism Edit Studio 🎨

Experimental Standard-Qwen FP8 Edition

This separate public test Space uses the pinned standard Qwen/Qwen-Image-Edit-2511 transformer — not Rapid-AIO v18. Its standard transformer and text encoder are dynamically quantized to FP8 only because the 57.7 GB BF16 stack could not bind to ZeroGPU. Normal mode defaults to 6 inference steps, has a hard maximum of 12, true CFG 4.0, and guidance 1.0.

FP8 is a compatibility requirement here, not a cheaper model-family substitution. This Space still exists to test whether the early standard-Qwen execution path produced the photorealistic behavior users remember. The main Pro Realism Edit Studio is separate and remains unchanged.

Upload one image for the first comparison and enter your prompt. Normal mode remains the recommended default. Advanced Settings contains an explicit experimental high-cost mode for separately testing aspect-preserving/custom 2K output or up to 50 steps; these settings may consume substantially more ZeroGPU quota, queue longer, time out, or exceed GPU memory.