On bare Krea 2 a short camera phrase barely moves the camera. Write low angle shot, worm's eye view or low angle shot from behind and the picture comes out at or near eye level. high angle shot from behind and high angle shot work a little without help. Until now the fix was a long framing paragraph for every image: where the ground meets the bottom of the picture, which part of the body is big and which is small, what fills the top.

This LoRA makes the short phrase do that job. The low phrases drop the camera to the ground, worm's eye puts the boot sole in front of you, and the two high phrases come out steeper.

NEW IN V2 (OUT 29 OCTOBER): LONG PROMPTS

v1 needs short prompts. In a long paragraph its phrase gets weaker. When I put v1 next to v2, a low shot from behind on a stone bridge went high with v1 on all 3 seeds. v2 is trained on the same 39 images twice, once with the short caption and once with a long paragraph caption of 100 to 160 words. The same bridge paragraph on v2 gives a true low angle from behind on all 3 seeds: the lens at the cobbles, the sky filling the top of the frame.

Long worm's eye, high from behind and plain low shots came out as strong as on v1 or stronger. Short prompts did not get worse. The triggers and the strength are the same, so everything below holds for both versions unless it names one.

HOW TO USE

1. Load it in ComfyUI with LoraLoaderModelOnly at strength 1.0.

2. There is no trigger word. The six angle phrases are the triggers:

   • low angle shot

   • worm's eye view

   • low angle shot from behind

   • high angle shot from behind

   • eye level shot

   • high angle shot

3. Write the prompt in this order: the phrase, then the medium, then one sentence about the subject. <phrase>, a photograph of <one sentence> or <phrase>, an illustration of <one sentence>. This is the form of every training caption. On v2 the sentence can grow into a long paragraph, because half of its captions are one.

4. v1 needs short prompts. One sentence, about 30 to 60 words: who or what, what they wear, where they stand, what is in each hand, the light. v2 holds the angle in long paragraphs as well, so a full scene can follow the phrase. On both, leave every other camera word out.

5. For worm's eye, end the subject with looking down at the viewer. On v1, keep it short too.

6. If you stay on v1 and need a long paragraph, the phrase gets weaker. Add a closing sentence that repeats the angle and names the ground, for example Low angle shot from behind, the lens just above the paving stones. It is a partial fix. It helped the low angle from behind and did not rescue worm's eye. v2 held the long bridge prompt without it.

7. a photograph of prompts stay photographic. an illustration of prompts come out in flat cel colour.

Example prompt at 1.0 (896x1152, seed 8001). It is the cover prompt:

worm's eye view, a photograph of a huge old matriarch elephant with long curved tusks and deeply wrinkled grey skin standing on dark cracked mud at a dry waterhole on the savannah, its trunk curled up and its ears spread wide, looking down at the viewer, dark storm clouds and low gold evening sun

Example long prompt on v2 at 1.0 (896x1152, seed 7915). It is the bridge in the v2 showcase, and the prompt v1 failed on:

low angle shot from behind, a painterly anime illustration in a soft cel-shaded semi-realistic register, drawn rather than photographed, with clean line art and flat colour areas. A wide stone bridge over a dark river at dusk, the parapet and old iron lamp posts running ahead, a cathedral with two spires on the far bank against an orange and violet sky. A tall Romanian man of 34 with iron-grey temples, in an ankle-length black leather coat and black boots, walks away along the bridge, one hand on the parapet, the other holding a paper bag of bread, looking back over his shoulder at the viewer. The last warm light of sunset on his shoulders, the lamps just lit.

IF THE CAMERA DOES NOT MOVE

It stays near eye level. Look at the length first, and at the version. On v1, a long paragraph about a man walking along a stone bridge stayed at eye level or went high, on all 3 seeds I tried. The same scene as one short sentence passed on the one seed I tried it on: a strong low angle from behind, the boots big in front. The length was the problem, not the scene. v2 fixes it: the same long paragraph comes out low from behind on all 3 seeds.

Worm's eye comes out as a plain low angle. Shorten the prompt and add looking down at the viewer. On v1, a long worm's eye prompt gave a strong low angle from the cobbles both times I tried, with and without the closing sentence. On v2 the same long prompt came out as low as on v1 or lower.

From behind comes out side-on. Give the subject somewhere to face, away from you: facing away up the slope toward the clouds. A look back over the shoulder can turn the whole body sideways. It did on one showcase seed, and it worked on others.

Raising the strength does not help. On v1 at 1.3, three of four long prompts barely changed, and the fourth only moved from high to about eye level. Nothing looked overcooked. It just did not fix them.

A bridge, a hallway or an aisle running away into the distance makes it harder. A view like that resisted the low camera while I built the dataset, and the bridge is the scene v1 failed in the long prompt. In short form the bridge passed on v1, and on v2 it passed in the long prompt too. On v1 an open place is the safer bet: a hill, a beach, a field, a cliff.

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.

• Strength 1.0, range 0.8 to 1.2. On v1, 0.8 and 1.0 were close, and 1.0 was a little stronger on worm's eye. I tested v2 at 1.0.

• Other checkpoints and other UIs are untested. If you try it somewhere else, tell me how it went.

WHAT I TESTED

Before the v1 release I ran 11 short prompts on subjects it was not trained on, one phrase each, on steps 1000, 1500 and 2000, next to the same prompts with no LoRA. Two more had no phrase at all.

• Low angle: without the LoRA it was mild or near eye level. With it, the camera dropped and sky or ceiling filled the top. Step 2000 was the strongest.

• Worm's eye: without the LoRA, a plain figure at eye level. With it, the boot sole at the bottom of the picture. On step 2000 it held even when the subject looked away.

• Low angle from behind: without the LoRA, from behind at about hip height. With it, the camera on the ground and the heels big.

• High from behind and high: these already passed without the LoRA here. With it they came out steeper and wider.

• Eye level: it held on every step. Step 2000 framed a little wider.

• No phrase: the framing went a little wider and lower, and in one prompt the subject turned to look at the viewer. Mild.

Then long prompts, a full paragraph in my house illustration style after the phrase. The high angle from behind passed and came out steeper than without the LoRA. The worm's eye came out as a low angle. The low angle from behind on the bridge failed, as above. Two new long scenes with nothing running into the distance: a hill meadow, low from behind, passed. A rooftop observatory, low angle, came out only mildly low.

Then the v1 showcase at 1.0: 8 scenes, all 8 kept, from 10 images. 6 passed on the first try. The archer's bow broke below the grip, so she carries a spear now. The highland herder came out side-on while he looked back over his shoulder. Written as facing away up the slope, he came out from behind.

For v2 I ran the long and short test prompts again with the same seeds, v1 next to v2, all at 1.0.

• Long low angle from behind, the bridge: v1 went high on 3 of 3 seeds. v2 came out low from behind on 3 of 3, the lens at the cobbles and the sky filling the top of the frame.

• Long worm's eye, long high from behind, long low and the long meadow from behind: v2 as strong as v1 or stronger.

• Short prompts: no regression.

• No phrase: v2 still framed slightly low, and on that seed the subject turned to look at the lens.

The v2 showcase is from that test set, v2 at 1.0: the bridge at dusk, the rainy street worm's eye, the storm meadow from behind, the forest worm's eye, the night market from behind and above, the rooftop telescope, the dock from behind and the plaza worm's eye.

LIMITS

• On v1, long paragraphs weaken it. The closing sentence is a partial fix there, and worm's eye needs the short form. v2 holds the angle in long paragraphs.

• an illustration of comes out in flat cel colour, sometimes with blotchy flecks in skies and snow. v2 still adds the small paint flecks. The photographs keep their look.

• Loaded with no phrase in the prompt, it frames slightly lower (v1 also a little wider), and the subject can turn to look at the lens. Unload it when you do not want an angle.

• Eye level and high mostly work without it. They are in the training set as controls.

• The training captions carry no other camera words, so mixing it with other framing language is untested.

• Text is untested. The training images carry no readable text, but a test locomotive grew a number plate. Things that carry writing in real life may grow it.

• Tested only on RedCraft 3.0.

MAKER'S NOTE

v1 was trained on Krea 2 Raw with ai-toolkit (arch krea2), rank 32, batch 1, lr 2e-4, 1024 px, 2000 steps, about 51 epochs on 39 images. It runs on RedCraft. The file is the final step, 2000.

v2 uses the same settings and the same 39 images, each in the dataset twice: once with the short caption and once with a long paragraph caption of 100 to 160 words. 3000 steps, about 38 epochs over the 78 pairs. The file is the final step, 3000.

The dataset is synthetic. I rendered 39 images on Krea 2 with no LoRA, each with the long framing block for its angle from my Posing and Camera Angles benchmark, fitted to the scene: the ground at the bottom edge, the size chain (feet big and head small from below, shoulders big and feet small from above), and what fills the top. 31 people, a different nationality each, and 8 animals, buildings and objects. About half are illustrations in my house style and half are photographs.

Each image was then captioned with only the short phrase, the medium and plain content. No framing words. So the LoRA learns to make the short phrase do the long block's job. The v2 long captions follow the same rule: the phrase and the medium, then a full paragraph of plain content.

Its sibling is Dutch Angle: Camera Tilt for Krea 2, which rolls the camera. I have not tested the two together.

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