Concept
The concept for the final project is not to create a character, but to teach the neural network a style: taking the aesthetic of retro-futuristic space posters from the 1950s–1960s (raygun gothic) and transferring it into Stable Diffusion v1.5 as a training trigger.
Instead of using pre-existing, borrowed imagery, the dataset is assembled by the author: first, a «style anchor”—a contrasting poster featuring a rocket and halftone grain—is generated. From this anchor, 18 img2img variations and 12 txt2img variations are derived with different subject matter, ensuring the style is not tied to a single motif. The fine-tuning utilizes DreamBooth via an LoRA adapter with rank 16 on the UNet attention projections; the text encoder and VAE are frozen, and the latents and embeddings are cached, allowing 800 steps to take approximately 6 minutes on an RTX 3080.
The rare trigger token sks ties the style to a word, after which the check is strictly enforced: the series «alternative history of astronautics, ” composed of 12 storylines, does not repeat the training sample even once. A passenger liner to Mars, a cross-section of a planetary city, a space zoo—the style remains consistent across all of them. The typography on the posters is intentionally suppressed by a negative prompt: letter generation in SD 1.5 is unstable, and a clean graphical presentation is more important for the series than the captions.
Generations
The project was created using Stable Diffusion 1.5 with DreamBooth and LoRA applied. The model was fine-tuned on a proprietary dataset of 30 images, generated locally using the same model. The goal of the project was to master the technology of fine-tuning generative networks and to create a custom style trigger capable of reproducing the aesthetic of retro-futuristic space posters from the 1950s–60s in new narratives. The fine-tuned model was then used to generate a series of images for the project.
