The homepage toggle is easy to ignore and expensive to misunderstand. Text-to-image (text2img) invents a picture from words. Image-to-image (img2img) starts from a file you provide and treats the prompt as a list of edits. If you use the wrong mode, you either lose the identity you already had, or you fight a blank canvas that will never match a real product’s silhouette.
Use text-to-image when you do not have a source picture that is worth protecting. Concept art, festival scenes, invented characters, and “what if this interior were at night” all belong here. The model is free to choose pose, lens, and layout. Your job is to write a complete brief, as in the prompt guide.
Text-to-image is a poor tool for “make my exact storefront look like dusk” if you already photographed the storefront. The model will invent a different building that merely feels similar. That can be fine for mood boards. It is not fine for a listing that must remain the same shop.
Use image-to-image when the source already contains the composition you want to keep: a product on a table, a character sheet, a sketch, a logo lockup you have rights to, a photo of a room you may restyle. The prompt should describe the change, not the entire universe. “Keep the bottle and label, replace the kitchen with a marble counter and north-window light” is an img2img prompt. “A beautiful product photo of a bottle in a luxury kitchen, cinematic, 8k” is a text-to-image prompt pretending to be an edit; the label will drift.
On Nanomelon, img2img needs a source image URL or upload. If the source is missing, the job cannot do what the mode promises. The public gallery marks generation type on each card. When you click Make Same on an img2img work, the site tries to carry the source into the editor. When you click Make Same on a text2img work, you are cloning a prompt, not a photograph.
Different models interpret “how hard to push” differently. As a working rule: low change keeps geometry (the horse lantern still has four legs and a rider). High change keeps palette and vibe but will move limbs and architecture. If you need the same face, stay low and mention “preserve face and camera angle.” If the source is a rough sketch, you want a higher change so line art can become a painting.
Do not use img2img to launder someone else’s photo into a “new” artwork you claim as original. You need the right to the source. Our terms put that on you; the model will not check your license.
Product. Photograph the object on a clean background if you can. Use img2img to move it into a lifestyle set. Name materials in the prompt so plastic does not become glass. Keep type on the label short in the prompt; models still mangle small text—plan to composite real type in an editor if the label must be readable.
Character. A single reference sheet (front view, even light) is more useful than three random selfies. If you only have words, use text-to-image until you have a face you like, then switch to img2img for wardrobe and location so the person does not become a sibling every run.
Scene. Architecture and streets are where text-to-image shines, because you rarely need a specific real address. If you do need a real place, you are in photography or 3D, not this tool. Do not generate “evidence” of events that did not happen and present it as fact. See About for how we think about misleading images.
Identity drift: you asked img2img to “make it night” and the person aged ten years. You over-specified a new style and under-specified what to keep. Add “same person, same crop.”
Texture soup: the source was already noisy or upscaled. Img2img will double the artifacts. Start from a cleaner file.
Mode mismatch: the gallery says text2img but you expected it to match a photo you have on disk. That page never saw your photo. Generate in the correct mode from the homepage.
Failed and policy-blocked jobs do not deduct credits. Successful jobs do, including ones you dislike. That is why quantity 1 is the default for tests. Image-to-image is not “free because you already had a picture”; it still calls a model. See models and credits for how to budget a session.
A practical sequence we recommend to new accounts: two text-to-image tests to find a composition, one img2img pass to lock identity or product, then a last text-to-image only if you decide the source was holding you back. That sequence wastes fewer credits than five random text-to-image rolls with a paragraph that changes every time.
Toggle the mode on the homepage before you paste a prompt. Read the user guide if you have not used the form yet.