img2img denoising strength: a working scale from 0.2 to 1.0

In img2img the input picture is partly noised, then denoised toward your prompt. Denoising strength is how much noise is added, so how much of the original survives. A practical scale - 0.15-0.3: retouch. Same composition and colours, cleaner surfaces. Good after an upscale. - 0.35-0.5: restyle. Shapes and layout hold, rendering and texture change. - 0.55-0.7: reinterpret. The pose and broad layout survive, details are new. - 0.75-0.9: loose suggestion. Colour masses and rough placement only. - 1.0: the input is effectively gone; you are doing text-to-image with extra steps. Steps interact with strength In A1111, a strength of 0.5 with 30 steps runs about 15 actual steps. Low strength with low step counts can therefore barely change anything. ComfyUI expresses the same thing as "denoise" on the KSampler. When to use ControlNet instead If you want a high strength (lots of change) but the pose or outline must hold, img2img cannot do both. Lower the strength, or keep strength high and add a ControlNet (canny, depth or openpose) to lock the structure. The Studio's Remake works on that second principle: it keeps the outlines of the source and re-renders everything else. Reproduce it Take one input, one seed, and render strengths 0.3, 0.5, 0.7 side by side. Post the strip with the input. That one image answers most img2img questions on this board.