Sea of Clarity

Between 1969 and 1972, six Apollo crews landed on the Moon. Each brought back film from the expedition: 70 mm, a Hasselblad camera, and a Zeiss 60 mm lens. Black crosses are visible on every frame. This is not a defect, but a marking: a glass plate with a grid of reference points was placed in front of the film, and distances on the shots were measured against these points later.

Everyone recognizes these images. The black sky without stars. Gray soil resembling ash. The sun strikes from the side, casting long, sharp shadows. There are few objects, and each stands independently: a module, a flag, a rover, a person in a spacesuit. Nothing superfluous, because nothing extra could be transported there.

More missions were planned. Apollo-20 was the first to be canceled on January 4, 1970, followed by Apollo-18 and Apollo-19 in September of the same year. Their film remained undeveloped.

Why a country house? It is also a place where everything is brought with you and where every object matters. It’s an expedition for the weekend—with your own gear, rituals, and specimen collection. Moreover, a country plot is easily recognizable by a single object: a greenhouse, a well, a banya, laundry on a line—each serves as a flag.

For training the model, 146 frames were selected from the NASA Image and Video Library archive, captured by astronauts from the Apollo 11 to Apollo 17 missions.

Initially, 197 frames were found. After review, images featuring the lunar surface, individual objects, and characteristic details of the lunar landscape were chosen.

Images from the cockpit and command module, crew portraits, frames viewed through the viewport, blurred and overexposed photographs, as well as near-duplicate images, were excluded from the selection.

All 146 images have been converted to a square format of 1024×1024 pixels. Each frame has been manually captioned in English, describing the main objects and their placement:

«srnt photo, an astronaut walking towards the camera, reflection in the gold visor» «srnt photo, the lunar rover far away below a mountain, tracks in the foreground» «srnt photo, a huge split boulder»

The captions contain no film grain, noise, crosses, or harsh sunlight. This is consistent across all 146 frames, and the model should associate this with the «srnt photo» trigger.

Model: Stable Diffusion XL 1.0, native resolution 1024 pixels. Method: LoRA, fine-tuning with captions, official diffusers script «train_text_to_image_lora_sdxl.py». Hardware: one Tesla T4 GPU, Kaggle.

Reflection is intentionally disabled. The frames contain flags, text on the module, and shadows that fall in one direction. A mirror image of the Moon would look fake.

Noise offset is necessary because of the sky. Standard SD pulls every frame toward a mean gray, causing the sky to look dark gray. The noise shift allows the model to render true black.

After 250 steps, the images acquired the characteristic features of lunar photography: a cold gray palette, graininess, and pronounced shadows. With further training, the details of the lunar module became more precise, but the images of the greenhouse and the car still resembled Earth landscapes.

Results: Without LoRA, the astronaut looks like a staged shot: bluish sky, plastic spacesuit, smooth terrain. After 250 steps, it’s already like an Apollo film. Cold gray color, grain, shadows beneath the feet, the module is cropped by the edge of the frame, like a real photo taken by an astronaut. After 500, 750, and 1000 steps, the frame changes almost not at all. The module becomes more precise: foil, supports, ladder. The greenhouse and the «Zhiguli» remained on Earth throughout all stages. The film became duller and flatter, the color closer to «Apollo,» but with grass, trees, and a pond surrounding it.

Key takeaway from the training: the trigger learned the film, but not the Moon. In the captions, the Moon is named with the words: «grey plain, ” „black sky, ” „lunar module.“ Therefore, for the model, the Moon exists within these words, not in the srnt photo. For the series shoot, this means the location must be named directly in the prompt.

For the series, the weights taken after 1000 steps were used: the lunar module is the most accurate on those weights, and there is no sign of overfitting.

Flight Assignment: Twelve frames. Each frame must feature one piece of country life equipment: a small house, a greenhouse, a «Zhiguli» car, a well, a scarecrow, a banya (sauna), a samovar, a booth, a cart, or a gate. In two frames, an astronaut: one digging potatoes and one hanging laundry.

Camera Settings: SDXL 1.0 with LoRA srnt photo, weights after 1000 steps; Adapter strength: 1.15; Sampler: DPM++ 2M Karras, 30 steps, CFG 7; Six takes per frame, 1024×1024.

CLIP ViT-L/14 yields the best take in prompt adherence. Final selection is manual.

The prompt looked like this in the first shot:srnt photo, a small wooden dacha house with carved window frames and a porch, standing alone, on the grey lunar surface, black sky

Three things have been changed.

1. The location has been moved to the beginning, immediately following the trigger, and stated explicitly: on the moon, lunar surface, pitch black sky, harsh sunlight. The model pays closer attention to the first words of the prompt.
2. The negative prompt has been expanded. Added gray and overcast skies, fog, snow, winter, forest, and buildings in the background.
3. The LoRA strength has been increased from 1.0 to 1.15, and CFG from 6 to 7.

Captions follow the NASA catalog format: number, subject, and description of the shot. Tape number 14 corresponds to the site number. The Apollo-20 expedition did not take place; all images were generated.