Sign In

Dutch Angle: Camera Tilt for Krea 2

1

Not recently loaded

Download

1 variant available

fp16 SafeTensor

DutchAngle_Krea2_v1.safetensors

Half precision, best balance • 218 MB

Verified:

Type
LoRA
Stats

122

1

Reviews
Published

Oct 26, 2026

Base Model
Training
Steps: 2,000Epochs: 50
Usage Tips
Strength: 0.8
Hash
AutoV22AC4CB8B63
Trigger Words
dutchtilt
a1-s2.png

Krea 2 does not roll the camera from a prompt. In my Posing and Camera Angles benchmark, dutch angle scored 0 of 6: it slants a wall junction, but the frame stays level. So the advice was to render upright and rotate the image afterwards, and lose the corners to the crop.

This LoRA makes the model draw the tilt itself. No rotation, no cropped corners. The trigger is dutchtilt, and it works together with a tilt sentence, below. A note from testing: without the LoRA, the same sentence already gives a mild roll on some prompts, mostly the wrong way. With it, the roll lands in the direction you asked for.

HOW TO USE

  1. Load it in ComfyUI with LoraLoaderModelOnly at strength 0.8.

  2. The trigger alone does nothing. Every training caption carried the tilt sentence, so the sentence is what works. A test prompt with only dutchtilt stayed level on every checkpoint. Paste one of these:

  3. Clockwise: dutchtilt, the picture is tilted clockwise, about twenty degrees, so the horizon runs from the upper left down to the lower right and the verticals lean to the right,

  4. Counter-clockwise: dutchtilt, the picture is tilted counter-clockwise, about twenty degrees, so the horizon runs from the lower left up to the upper right and the verticals lean to the left,

  5. To set how far it leans, write slightly or strongly before the direction and change the degrees. The showcase uses fifteen to thirty, for example the picture is tilted strongly clockwise, about twenty-eight degrees.

  6. Put the sentence first. If you open with a style sentence, it goes right after it. The rooftop sprint in the showcase opens with my house sentence, A painterly anime illustration in a soft cel-shaded semi-realistic register., and then the tilt sentence.

  7. a photograph of prompts stay photographic. an illustration of prompts drift toward a flat gouache painted look. That comes from the dataset, where most of the scenes are illustrations in my own style.

Example prompt at 0.8 (896x1152, seed 7301). It is one of the showcase images, the detective. My first version looked straight down the alley and stayed upright, so this one turns the view side-on to a long wall:

dutchtilt, the picture is tilted strongly clockwise, about twenty-eight degrees, so the horizon runs from the upper left down to the lower right and the verticals lean to the right, an illustration of a rain-soaked back street at night seen side-on, one long brick wall running across the whole frame behind her with a row of lit square windows, a steel door and an iron fire escape climbing the wall, one street lamp over the door, puddles on the cobbles, seen at eye level. A Korean private detective, a woman of 38 in a belted grey trench coat and a fedora, stands under the lamp, her right hand aiming a small torch down the alley and her left hand on the iron rail of a fire escape, hard amber light and deep inky shadow.

IF IT DOES NOT ROLL

The frame stays upright. Look at the view. A symmetric one-point view, straight down an alley or along a street between two rows of houses, kept the frame upright twice in the showcase. A side-on view with one long straight line across the frame rolls cleanly: a wall, a parapet, a row of houses, the sea horizon.

It rolls the wrong way. The direction is a strong hint, not a guarantee. 2 of the 12 showcase images rolled the wrong way. Two fixes worked. Make the main straight line run with the asked direction: the burning village row is written rising toward the right for a counter-clockwise tilt. Or ask for the direction the scene already takes: the hospital corridor was asked clockwise, rolled counter-clockwise, and I asked counter-clockwise on the next try.

RECOMMENDED SETTINGS

  • Checkpoint: RedCraft 3.0 (Krea 2). The only one I tested.

  • 8 steps, CFG 1, sampler er_sde, scheduler simple.

  • About 1 MP: 896x1152 portrait, 1152x896 landscape or 1024x1024 square. All three are in the showcase.

  • Strength 0.8, the only strength I tested.

  • Other checkpoints and other UIs are untested. If you try it on plain Krea 2 Turbo or somewhere else, tell me how it went.

WHAT I TESTED

Before release I ran 6 test prompts it was not trained on, on steps 1000, 1500 and 2000, next to the same prompts with no LoRA:

  • a rainy city crossing at night

  • a woman running down a concrete stairwell

  • a castle gate with a knight riding through

  • a pub, in my house anime style

  • a kitchen, with the trigger and no tilt sentence

  • the city crossing again with no trigger, as a level control

The roll rose steadily from step 1000 to 2000. On the castle gate it leaned the wrong way at 1000 and the right way from 1500 on. On step 2000 all four prompts with the sentence rolled in the asked direction. The control with no trigger stayed level on every checkpoint, so the LoRA stays out of the way when you do not ask for a tilt.

Then the showcase at 0.8: 8 scenes, all 8 kept, from 12 images. 5 passed on the first try.

  • a detective in a back street at night

  • a skateboarder mid-ollie, a photograph

  • a rooftop sprint, in my house anime style

  • a dragon over a burning village

  • a hospital corridor at night, a photograph

  • a starfighter cockpit

  • a trawler on a storm wave, a photograph

  • a saloon standoff with two gunfighters

Of the 4 misses, two stayed upright and two rolled the wrong way. The tips above come from them.

LIMITS

  • Tested only at strength 0.8, and only on RedCraft 3.0.

  • The degree number is a rough guide. I did not measure the angles.

  • Wrong-way rolls happen: 2 of 12 showcase images.

  • Symmetric one-point views resist the roll.

  • Illustration prompts drift toward flat gouache shapes.

  • Text and signage are untested. The training images carry no readable text.

MAKER'S NOTE

Trained on Krea 2 Raw with ai-toolkit (arch krea2), rank 32, batch 1, lr 1e-4, 1024 px, 2000 steps, about 50 epochs on 40 images. The file is the final step, 2000.

The dataset is synthetic. I rendered 20 level scenes on Krea 2 with no LoRA, 12 in my house illustration style and 8 as photographs, each with a straight horizon, upright verticals and no text. From those I built 28 tilted images, 15 to 30 degrees, 14 each way, all cut with one centred crop rule, so it learns the tilt and not a zoom. The other 12 are level copies, captioned without the trigger.

The captions describe where the horizon and the verticals sit in the picture, not how the camera turned. Krea 2 follows where things sit in the frame, not camera words.

Free to use. Credit is appreciated. The licence is on this page.