Seeds: why the same seed sometimes gives a different image

A seed sets the starting noise. Same model, same seed, same settings, same software: same image. When that breaks, it is one of these. The usual causes - Different resolution. The noise tensor has a different shape, so it is different noise. A seed only means something at one size. - Ancestral sampler with a different step count (see the sampler thread). - GPU vs CPU random numbers. A1111 has a setting for the RNG source (GPU, CPU, or NV); an image made with one will not match another. ComfyUI generates the starting noise on the CPU by default, which is one reason ComfyUI and A1111 give different pictures for the same seed. - Batch position. Image 3 of a batch of 4 is not "seed + 0"; tools number batch seeds differently. - Non-deterministic speed-ups. Some attention optimisations (xformers in certain versions) produce tiny run-to-run differences that grow over many steps. - A LoRA, embedding or VAE you forgot was loaded. What a seed is good for - Comparing settings fairly: change one variable, keep the seed. - Coming back to a composition you liked and refining it with img2img or inpainting. What a seed is not good for Character consistency across different prompts. Change the prompt and the same seed gives a different person. Use a reference image or a LoRA for that (see the Characters & Consistency board). Write it down A1111 and ComfyUI both embed the generation settings in the PNG. If you post the PNG somewhere that strips metadata, paste the settings into the post too.