Part 01
Ten concepts, six corrections
I took ten reference concepts and rebuilt each one with text alone, aiming for anime with semi-realism only in the light, the volume and the materials. The real question turned out to be where the line is.
Krea2 · txt2img · seed 92401 · 1024×1280 (4:5)6 attempts
Too realistic, then fixed
I rebuilt each concept from a reference using only text (no img2img). In the first review, six of the ten concepts needed a correction, four of them for being too realistic or photographic. Top row: the first batch. Bottom row: the corrected one.
- Dragon · 1stDiscarded
Ivory Dragon Oracle, first batch: it reads as a realistic fantasy painting, not as anime.
Seed 92301 · 62 % anime / 38 % semi-real - Cat · 1stDiscarded
Golden Cat Goddess, first batch: too realistic.
Seed 92601 · 58 % anime / 42 % semi-real - Fae · 1stDiscarded
Botanical Fae Empress, first batch: a realistic portrait.
Seed 92701 · 55 % anime / 45 % semi-real - Dragon · fixedBetter
Eyes, nose, mouth, hair and face planes described as unmistakably anime. The dragon, the violet and the opal ornament stay.
Seed 92401 · corrected batch - Cat · fixedBetter
The halos, the gold and the flowers stay; the face and the hair are clearly anime now.
Seed 92701 · corrected batch - Fae · fixedBetter
The botanical density, the wings, the jewelry and the palette stay; the face is anime with a soft painterly volume.
Seed 92801 · corrected batch
5 attempts
The dial, from 84 to 55
Every concept had its own balance written into the prompt. Placed side by side, the numbers show how far I pushed the semi-realism before the image stopped reading as anime.
- 84 / 16Option
Scarlet Occult Sorceress: high-contrast crimson and black with cel shading. The most anime of the set.
84 % anime / 16 % semi-real in the prompt - 70 / 30Option
Atompunk Hotrods, batch 3: the balance I recovered from Atompunk Girl.
70 % anime / 30 % semi-real in the prompt - 62 / 38Option
Ivory Dragon Oracle: ornate dark fantasy with opal scales.
62 % anime / 38 % semi-real in the prompt - 58 / 42Option
Golden Cat Goddess: molten gold, filigree and backlight.
58 % anime / 42 % semi-real in the prompt - 55 / 45Option
Botanical Fae Empress: painterly, with art-nouveau curves. The most realistic balance I could still call anime.
55 % anime / 45 % semi-real in the prompt
6 attempts
The other way round: when the face goes flat
In another series, Atompunk Hotrods, I had the opposite problem: the faces came out flat, like a cartoon. I blamed the negative and removed it. They came out just as flat. The cause was in my own text. I was asking for winks and big smiles, and I never described the face. So I added one sentence that describes the face and swapped the gestures for calm ones. Top row: before. Bottom row: after, same scene and same seed.
- 1 · ADiscarded
A flat face: big eyes and almost no volume.
Hotrods · scene 02 · face not described - 2 · ADiscarded
A cartoon smile. The nose is barely a line.
Hotrods · scene 04 · face not described - 3 · ADiscarded
Huge eyes and a smooth, flat face.
Hotrods · scene 09 · face not described - 1 · BBetter
Same scene. Now there is volume in the cheekbones and the lips.
Hotrods · scene 02 · with one sentence describing the face - 2 · BBetter
A calm look, and a nose with shape. It also changed her eye color, which I did not ask for.
Hotrods · scene 04 · with one sentence describing the face - 3 · BBetter
It no longer reads as a cartoon.
Hotrods · scene 09 · with one sentence describing the face
What I learned
The percentages are words, not settings. 84/16, 70/30 or 62/38 are art-direction descriptors in the prompt; the model doesn’t measure them. What changes the result is what I describe.
Realism gets in through the face. Every time a first batch came out too realistic, the fix was the same: describe the eyes, nose, mouth, hair and face planes as anime, and keep the realism for light, volume and materials. It works the other way too: when a face came out flat, describing it in the text brought the volume back. Cartoon gestures, like a wink or a big smile, flatten it.
Attempts: Dragon · 1st, Dragon · fixed, Cat · 1st, Cat · fixed, Fae · 1st, Fae · fixed, 1 · A, 1 · B
A correction round is the normal path, not a failure: six of the ten concepts needed one in the first review.
Attempts: Dragon · 1st, Cat · 1st, Fae · 1st
Rebuilding a reference from text alone keeps the idea, not the pixels. I describe subject, composition, palette, line and shading, and generate from zero, without img2img.
Different concepts want different balances: a graphic sorceress works at 84/16, a painterly botanical portrait at 55/45.
The next experiment
Four ways to double an image
Which upscaler can I trust to double an image without redrawing it?
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