---
license: creativeml-openrail-m
base_model: runwayml/stable-diffusion-v1-5
tags:
  - stable-diffusion
  - lora
  - cinematic
  - interstellar
  - nolan
  - christopher-nolan
  - text-to-image
---

# Absolute-Cinema-AI-Tiny

A cinematic LoRA fine-tuned on *Interstellar (2014)* stills,  
trained on 910 curated frames to capture Hoyte van Hoytema's  
iconic cinematography style — heavily inspired by the visual storytelling of Christopher Nolan.

## Disclaimer

This is a **tiny experimental LoRA**, not a production-grade model.
Results can be inconsistent, and in many cases, quality may fall short compared to larger or better-trained models. Think of this more as a **proof-of-concept / early iteration** rather than a polished release.
That said — it can still produce some surprisingly cinematic shots if prompted well

## Usage
**Trigger word:** `ISTELLAR`

## Example Prompt
ISTELLAR, cinematic masterpiece, hoyte van hoytema,  
anamorphic lens flare, 70mm film grain, IMAX format,  
dramatic natural lighting, shallow depth of field,  
teal and orange color grading, [your scene here]

---

## Training
- **Base model:** SD 1.5  
- **Images:** 910  
- **Epochs:** 8  
- **Network:** LoCon (dim 32, alpha 16)  

---

## Performance

- **Size: 1.8GB**
- Strong cinematic composition out-of-the-box  
- Handles lighting, contrast, and framing exceptionally well  
- Works best with detailed prompts (camera terms = boost)  
- Minimal overfitting despite stylization  

---

## Generation Time (Estimated on T4)

*(No upscaling unless stated otherwise)*

| Resolution        | Aspect Ratio | Label          | Time (Approx) |
|------------------|-------------|---------------|--------------|
| 512×512          | 1:1         | Square        | ~8–12 sec    |
| 768×768          | 1:1         | Square HD     | ~20–36 sec   |
| 1024×1024        | 1:1         | Square HQ     | ~32–45 sec   |
| 768×432          | 16:9        | Cinematic     | ~10–16 sec   |
| 1024×576         | 16:9        | Cinematic HD  | ~18–23 sec   |
| 1280×720         | 16:9        | HD (720p)     | ~25–35 sec   |
| 1920×1080        | 16:9        | Full HD       | ~45–60 sec   |
| 2048×858         | ~2.39:1     | Ultrawide     | ~65–110 sec   |
| 3840×2160*       | 16:9        | IMAX / 4K     | ~120–240 sec  |

> *4K/IMAX typically achieved via 4× upscaling pipeline*  
> Times vary based on steps, guidance scale, and VRAM optimization settings

> **NOTE: Upscaler is buit in!**
---

## Notes

- Designed for cinematic realism, not anime/cartoon styles  
- Best paired with camera/film keywords (IMAX, 70mm, anamorphic, etc.)  
- Works insanely well for space, landscapes, and emotional close-ups  
- Exeptions are humans and anything human body related unless specified very much

---

## Future Work

| Version | Base Model | Dataset | Status |
|---|---|---|---|
| **Absolute-Cinema-AI-Medium** | SDXL 1.0 | Interstellar + Oppenheimer | Next |
| **Absolute-Cinema-AI-Large** | SDXL 1.0 | All 4 movies | Planned |
| **Absolute-Cinema-AI-XL** | Flux.1-dev | All 4 + extra Nolan | Someday |

## Credits

- Cinematography inspiration: Hoyte van Hoytema  
- Style influence: Christopher Nolan  
- Creator: Andy-ML-And-AI