TLDR: Conditioning looks successfull!
[2023/03/04] Please note that yes, while AstroSleuthNEXT's ability to color match leaves a lot to be desired, turns out, humams have like 100x the number of lumance cones than for color! The bandaid solution was to simply upsample the YR/CB color channels with bilinear interpolation, and it's good as gold. See against blurx for (simple) color matching results. Otherwise, the other models at this stage have good enough color matching that they don't need any special treatment. Back to the original content
It certainly failed to keep the same color, however I believe this is because the color information may be contained within the condition embedding, and because I used a generalised one for inference, the output is inaccurate.
Will be very interesting to see what other possible conditioning prompts could do?
Ideas:
- Saturation
- Contrast
- Nebula detail
- Diffraction spikes
| Original | AstroNEXT (no cond) |
|---|---|
 |
 |
| Low detail, high stars | Low detail, low stars |
|---|---|
 |
 |
| high detail, high stars | high detail, low stars |
|---|---|
 |
 |