In diffusion-based image-to-image editing, a single number between 0 and 1 decides how much of your source picture is allowed to live. That number, usually called strength or denoising strength, sets how much noise is added to the input before the model starts rebuilding it. At 0.2 the model barely touches the image. At 1.0 the original is drowned so thoroughly that the output is close to a fresh text-to-image render with a faint memory of the starting frame. Everything interesting happens in between, and that middle band is where creative tools, image metrics and copyright law now meet.
What the Noise Strength Dial Really Controls
The technique traces back to SDEdit, a 2021 method that showed you could push a rough sketch partway into noise and let a diffusion model denoise it back into something photographic. Open-source pipelines still work the same way. With 50 sampling steps and a strength of 0.75, the model skips the first quarter of the schedule and runs roughly 37 denoising steps starting from a partly noised version of your image. Lower strength means fewer steps and more of the original pixels steering the result.
What survives is not uniform, though. Broad composition and dominant color tend to hold on longest, because they live in the low-frequency structure that noise erases last. Fine texture goes early. Faces are fragile: a strength of 0.5 can keep the pose and lighting of a portrait while quietly swapping the person’s jawline, which is a problem for anyone editing a real subject. Text is the most brittle element of all. A shop sign that reads perfectly at 0.3 can turn into confident nonsense by 0.6.
Artlist and Playground Take Different Routes
Artlist approaches the problem by aggregating models instead of training one. Its image to image AI generator lists more than twenty engines, including Flux 3, GPT Image 2.5, Nano Banana 2.1, Seedream 5.0 and Krea 2, and splits fidelity into two modes: a strict one that keeps the layout and changes only what you ask, and a looser one that treats the source as a starting point. Several of those models accept multiple reference images, up to 16 in the case of Nano Banana 2.1, which matters for keeping a product or a character consistent across a set. Artlist says commercial use of the results is covered, that it makes no ownership claim over them, and that neither it nor its model partners train on uploads. For a creator delivering client work, that licensing position is arguably worth more than any single model.
Playground followed a different path. The company trained its own model, Playground v3, which a September 2024 technical report described as conditioning on a Llama 3 language model, with self-reported text-rendering scores ahead of several rivals. It then rebuilt the product around graphic design, with templates for logos, posters, T-shirts and other formats. The company has since been acquired, and its new owners say Playground is not slowing down. Playground suits designers who want layout and typography in one place. Artlist suits people who want to test the same source photo against many engines and keep the rights question simple.
Measuring What Survived
Saying an edit “kept the composition” is subjective, so researchers reach for numbers. The oldest workhorse is pixel error, which is a poor guide: two edits with nearly identical mean squared error can look completely different to a human. The structural similarity index from Wang, Bovik, Sheikh and Simoncelli measures structure instead of pixels, rewarding outputs that keep local patterns intact rather than matching raw pixel values. Their lab’s own demonstration shows five distortions of one photo with nearly identical mean squared error but SSIM scores ranging from 0.988 down to 0.662.
SSIM still misses a lot. A 2018 CVPR paper by Richard Zhang and colleagues introduced LPIPS, which compares images inside the feature space of a trained neural network, and found that these deep features matched human similarity judgments far better than SSIM or PSNR did. In practice, it is common for teams evaluating img2img pipelines to report both: SSIM for structural retention, LPIPS for how different the result feels.
Neither metric understands identity or meaning. A face can score well on both and still belong to someone else.
Guidance Beats Guesswork
The bigger shift is that strength is no longer the only lever. ControlNet, presented at ICCV 2023 by Lvmin Zhang, Anyi Rao and Maneesh Agrawala, attaches a trainable copy of a diffusion model’s encoder that reads an extra condition such as an edge map, a depth map or a pose skeleton. That lets a user throw away most of the pixels while locking the geometry. The result is a strange new kind of survival: the outline of the original persists exactly while every surface, color and texture is replaced. Commercial tools now hide this behind friendlier controls, such as region edits and “keep the layout” modes, but the underlying idea is the same.
Who Owns a Transformed Picture
Once an edit keeps the pose but not the face, or the layout but not a single original pixel, the question of what counts as the source becomes genuinely hard. Quantumrun has tracked how generative tools made everyone a potential creator while unsettling ideas of originality and intellectual property. US copyright authorities and courts have mostly been cautious, tending to protect the human contribution, such as the selection and arrangement of elements or the original photograph, rather than the machine’s rendering.
Europe is taking a disclosure route rather than settling authorship. Under Article 50 of the AI Act, which has applied since August 2, 2026, providers must ensure synthetic images are flagged in a machine-readable format, and deployers must disclose deepfakes that resemble real people, places or events. A voluntary code of practice was finalized in June 2026, and about 190 organizations had signed it by the end of July.
Where Image Fidelity Heads by 2030
A few trends look durable. Fidelity controls will become semantic: instead of a strength slider, users will pin specific attributes, such as this face or that exact headline, and let everything else change. Similarity metrics will move into products, with tools reporting how far an output drifted from its reference before export, which matters for brand compliance and for EU labeling. Provenance metadata will also travel with each edit, so a transformed image carries a record of what it came from. That last point may reshape licensing more than any model upgrade, because a buyer will be able to see whether the source was a stock photo, a client asset or someone else’s work.
The honest answer to how much of an original survives is that you now choose, attribute by attribute, and the law increasingly expects you to say so. Playground remains a capable option for design-led work built around its own model. Artlist makes more sense for creators who want many engines, explicit fidelity settings and outputs cleared for commercial use, which is what counts once a transformed image lands in paid client work.
