Shot Sizes for Krea 2

作者：Azimondious

Plain Krea 2 handles cowboy and medium shots and ignores most of the other sizes. Ask for an extreme close-up of a violinist on a dark stage and you get the man, the violin and the spotlight. Ask for a close-up of a potter at his wheel and the wheel comes too. On my test set the base model made 0 of 6 extreme close-ups and about 0 of 6 close-ups.

This LoRA makes those two sizes move the camera. There is no trigger: you write the size at the start of the prompt.

• Extreme close-up: 0 of 6 without it, 6 of 6 with it. The eyes and the bridge of the nose, no mouth, no outline of the head. It holds in a long paragraph too.

• Close-up: about 0 of 6 without it, about 2 of 6 with it. Short prompts only. 3 of 6 framed the size, and one of those turned an illustration into a photograph.

extreme long shot does something smaller, and it is still useful. It pulls the framing much wider than plain Krea 2: where the base model gives a medium or full figure, you get a long shot, the person about 17 to 35% of the frame height. A true extreme long shot, a tiny figure in a big landscape, landed on 1 of 6. Full shot is still weak. All of it is under LIMITS.

HOW TO USE

• Load it in ComfyUI with LoraLoaderModelOnly at strength 1.0.

• There is no trigger word. The six size phrases are the triggers. Put one at the very start of the prompt.

• Then the medium, then the subject: <phrase>, a photograph of <one sentence> or <phrase>, an illustration of <one sentence>. Every training caption opens this way.

• Extreme close-up holds in a long paragraph, and extreme long shot still pulls wider there. Close-up and full shot did not hold. I did not try cowboy or medium in a long paragraph, so for those four keep the prompt short.

• No size phrase in the prompt? Unload it, or lower it. Loaded at 1.0 with no phrase, it still frames tighter and turns the subject to face the lens. I have not measured how much a lower strength helps.

Example prompt at 1.0 (1152x896, seed 9101). It is the first side-by-side in the showcase, next to the same prompt and seed with no LoRA:

extreme close-up, an illustration of an Irish violinist of 30 playing on a dark stage, stage light from the left Example long prompt at 1.0 (1152x896, seed 9101). It is the first showcase image:

extreme close-up, a painterly anime illustration in a soft cel-shaded semi-realistic register of an Irish violinist of 30 with freckles and auburn hair pinned up, playing a worn violin on a dark wooden stage, a single warm spotlight from the left, dust drifting in the beam, the dark hall behind her with a few lit exit signs, her brow drawn in concentration. The violin, the stage and the exit signs are all still in the prompt. The LoRA cuts them out of the picture anyway.

THE SIX PHRASES

With the LoRA at 1.0, then plain Krea 2 on the same prompts and seeds:

• extreme long shot: meant for a tiny figure in a wide landscape. 1 of 6 got that far. All 6 came out wider than plain Krea 2, as long shots. Plain: 0 of 6.

• full shot: head to feet. About 2 of 8. Plain: about 1 of 8. This is the weak one.

• cowboy shot: the frame cuts across the thighs. 2 of 4. Plain: 4 of 4. Krea 2 already knows this one, and the LoRA makes it worse.

• medium shot: the frame cuts at the waist. 4 of 4. Plain: about 3 of 4.

• close-up: the face fills the frame. About 2 of 6 (3 of 6 framed the size, one of those turned an illustration into a photograph), and 0 of 2 in a long paragraph. Plain: about 0 of 6.

• extreme close-up: the eyes and the bridge of the nose only. 6 of 6, short and long. Plain: 0 of 6.

I scored these by eye, 2 seeds per test prompt. For the extreme long shot I used a strict rule: the figure under a tenth of the frame height. That is a small sample, so treat the counts as a ranking.

IF THE SIZE DOES NOT LAND

Extreme long shot comes out as a long shot. Expect this most of the time. The figure ends up about 17 to 35% of the frame height. The one tiny-figure result came from the long ridge paragraph.

Full shot comes out as a medium. In my tests a counter, a wheel or a table in front of the figure pulled the camera in, on every checkpoint I tried. The chef at the pass and the ceramicist at his workbench both came out as medium shots. Leave the furniture out of the sentence, or put it behind the figure. I have not tested that fix.

Close-up in a long paragraph stays medium. My long close-up test named canvas trousers, red sneakers and a brush in one hand, a paint tin in the other. It stayed a medium shot on both seeds. Use the short form for close-ups.

Close-up comes out too tight. In short prompts it can cut through the forehead and the chin, nearly an extreme close-up. It did that on the potter.

Cowboy comes out as a medium. On both seeds a safari guide lifted his binoculars to his eyes, and the frame went medium. On one rancher seed the frame cut near the belt. Plain Krea 2 did cowboy 4 of 4, so for cowboy, unload this LoRA.

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 1.0. I only rendered 1.0. If you try 0.8, tell me what it did.

• Other checkpoints and other UIs are untested.

WHAT I TESTED

18 test prompts with 2 seeds each, with people of nationalities that are not in the dataset. Twelve short ones, one illustration and one photograph per size, the photographs in the other orientation. Four long paragraphs: an extreme long shot, an extreme close-up, a full shot and a close-up. One with no size phrase at all, and one with long shot, a phrase it was not trained on. All of it at 1.0 on steps 1250, 1750, 2000, 2250 and 2500, next to the same prompts and seeds with no LoRA.

• Step 1250 already pulled the extreme long shot wider and fixed the short close-up.

• Step 2000 started the extreme close-up.

• Step 2250 was the only step with a true extreme close-up on all three of its test prompts. That is the file.

• Step 2500 relaxed the extreme close-up back to a close-up and added nothing anywhere else.

The showcase is from that test set, step 2250 at 1.0, with two side-by-side pairs against plain Krea 2.

LIMITS

• A true extreme long shot, with a tiny figure, landed 1 of 6. Most of the time extreme long shot gives you a long shot.

• Full shot is weak, about 2 of 8, and furniture in front of the figure turns it into a medium.

• Close-up fails in a long paragraph, and in a short prompt it can come out too tight.

• On one close-up written as an illustration, it drew a photograph. Plain Krea 2 kept it an illustration.

• Cowboy is worse than plain Krea 2: 2 of 4 against 4 of 4.

• The bleed with no size phrase is noticeable. On my one test prompt, plain Krea 2 drew the librarian in three-quarter view from the knees up. The LoRA at 1.0 drew her from the waist up, facing the lens. That is one prompt and two seeds.

• Strength 0.8 is untested.

• Most test prompts were plain photographs and plain illustrations. Heavily styled prompts are barely tested.

• long shot and medium close-up are not trained.

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

• Tested only on RedCraft 3.0.

MAKER'S NOTE

Trained on Krea 2 Raw with ai-toolkit (arch krea2), rank 32, batch 1, lr 2e-4, 1024 px, 2500 steps. It runs on RedCraft. The file is step 2250, not the final step.

The dataset is synthetic. I rendered 48 images on plain Krea 2, each one forced to its size by a long framing block: where the edges of the picture cut the body, or for the wide sizes, how small the figure is against the land. The caption never says any of that. It carries only the size phrase, the medium and plain content. So the LoRA learns to make the short phrase do the long block's job.

36 of the images are people, a different nationality each, ages 26 to 70. The other 12 are animals, vehicles and objects: an elephant, a heron, a scooter, a snowman, a brass lantern, a horse's eye, a drop of water on a leaf. Photographs, illustrations in my house style, watercolour and comic ink, each captioned by its medium. The cowboy images have no hats, guns or ranches in them. The phrase is a crop.

Each image is in the dataset twice, once with a short caption and once with a long paragraph caption, 96 pairs. The long captions are there so the size holds in a paragraph. I tested four sizes in a paragraph: the extreme close-up held, the extreme long shot still pulled wider, and full shot and close-up did not hold.

Its siblings are Camera Angles, which moves the camera up and down, and Dutch Angle, which rolls it. I have not tested them together.

If this worked for you, a thumbs-up here helps the next person find it.

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