004
Four ways to double an image
Which upscaler can I trust to double an image without redrawing it?
Part 01
One image, four upscalers
I upscaled one Krea 2 image to 2× with four methods to decide which one I use for each kind of work. I measured each result and then looked at the crops, because the numbers alone reward the method that changes the least.
Krea 2 · txt2img · seed 460701 · 1024×12805 attempts
The face and the hair, at 2×
The same 320×400 area of the test image, taken at each method’s own scale and brought to the same size. The numbers come from scaling each result back down to 1024×1280 and comparing it with the original.
- OriginalReference
The starting point: a 1024×1280 image. This crop is stretched from the original, so it is the softest.
Original · 1024×1280 - LanczosTest
The neutral baseline: it doubles the size and adds no detail. Good for a plain file copy.
PSNR 39.17 dB · mean error 1.136 · edge energy 3.806 - AnimeSharpBetter
The most faithful: it keeps the drawing, the identity and the line, and makes the edges clearly sharper.
PSNR 44.20 dB · mean error 0.752 · edge energy 4.674 - SeedVR2 3BDiscarded
It smooths a lot and changes the face and the small shapes more than the others. Not the right one for this style.
PSNR 28.08 dB · mean error 5.042 · edge energy 2.816 - SeedVR2 7BMy pick
It recovers definition with less drift than the 3B, but it is still generative: it reinterprets slightly. The best finish for a semi-real anime portrait.
PSNR 32.45 dB · mean error 3.251 · edge energy 3.925
5 attempts
The helmet and the hand
Hands, accessories and text are where a generative upscaler can redraw what was there. I look at them separately.
- OriginalReference
Helmet and hand in the original: small details that a generative upscaler can redraw.
Original · 1024×1280 - LanczosTest
Same details, just bigger.
PSNR 39.17 dB · mean error 1.136 · edge energy 3.806 - AnimeSharpBetter
The helmet, the hand and the trousers keep their shapes; the edges get crisper.
PSNR 44.20 dB · mean error 0.752 · edge energy 4.674 - SeedVR2 3BDiscarded
The softest of the four in this area too.
PSNR 28.08 dB · mean error 5.042 · edge energy 2.816 - SeedVR2 7BMy pick
More definition in the helmet and the fabric. This is where I check hands and accessories, because it is the method that can invent.
PSNR 32.45 dB · mean error 3.251 · edge energy 3.925
What I learned
AnimeSharp is the safe choice when I need to keep the drawing exactly. It had the highest fidelity (44.20 dB) and the sharpest edges. I use it for flat anime, chibi, posters and line art.
Attempts: AnimeSharp, AnimeSharp
For semi-real anime, SeedVR2 7B Sharp gives the best finish if I accept a slight reinterpretation. With noise at zero and LAB color correction it drifted less than the 3B (32.45 dB against 28.08 dB).
Attempts: SeedVR2 7B, SeedVR2 3B
The metrics measure similarity, not beauty. Scaling each result back down and comparing it with the original rewards the least change, so I also look at the crops, and at the hands and accessories first.
Attempts: Lanczos, SeedVR2 7B
A saved workflow can break quietly after a node update. My old SeedVR2 workflow had its values shifted (resolution, max_resolution, batch_size, uniform_batch_size and color_correction), which may explain earlier soft results and unexpected sizes. I made two corrected copies and kept the original.
If the image has text or readable symbols, I use AnimeSharp or Lanczos: a strong generative upscaler can redraw the glyphs.
The next experiment
Giant robots
Can the pilot stay anime while the robot stays cinematic?
OpenFollow the lab
Every experiment goes up as it happens: the picked images, the clips in motion and the batches that didn't work.
Write to me on Instagram