A painted-shard look for Krea 2. Every form breaks into hard-edged shards with broken brush strokes inside them, like shattered glass or torn coloured paper, in a palette that you name: a deep field colour, and a hot colour on the lit side of every form. It runs darker and moodier than the plain model, and a few stray iridescent specks come with it.
If you want clean low-poly triangles, you do not need this LoRA. Plain Krea 2 already draws them. Start the prompt with <code>a faceted geometric poster illustration of</code> and it cuts the picture into crisp triangles, and it even faceted a coffee cup. I tested both side by side before release (below). This LoRA gives a different picture, painted shards instead of clean geometry, and that is the only reason to load it.
The trigger is <code>facetposter</code>, the working name from training, but the palette does more of the work. Every training caption ended with one. My other Krea 2 utilities are <a rel="ugc" href="https://civarchive.com/models/3022257">Gritty Ink</a>, for brush-ink storybook pictures, <a rel="ugc" href="https://civarchive.com/models/3022090">Dutch Angle</a>, for a tilted camera, and <a rel="ugc" href="https://civarchive.com/models/3022787">Camera Angles</a>, for low, worm's eye and from-behind shots.
HOW TO PROMPT
<ol>
<li>Load it in ComfyUI with <code>LoraLoaderModelOnly</code> at strength 1.0.</li>
<li>Put <code>facetposter</code> at the start of the prompt.</li>
<li>Then write a plain prompt: subject, scene, what each hand holds, where the view is from. No style words are needed.</li>
<li>End the prompt with a named palette. Use one of the eight below. Without one, the colour follows the scene, and people and close-up objects stay close to a photograph.</li>
<li>Strength 0.8 to 1.0 works. Close-ups of people and objects need 1.0. At 1.2 it overcooks into confetti specks.</li>
</ol>
Example prompt at 1.0 (896x1152, seed 7701). It is one of the showcase images, the mountaineer. Every piece of her clothing has a palette colour:
<pre><code>facetposter a Chilean mountaineer, a woman of 29 in an ultramarine down jacket, rose climbing trousers, cobalt boots and a peach wool hat, climbing the last steps of a sharp snow ridge toward the summit at sunrise, her right hand driving an ice axe into the snow, her left hand gripping a coil of rope over her shoulder, a sea of clouds and far peaks below, seen from low on the ridge with the sky filling the top of the frame, cobalt and ultramarine with peach and rose light</code></pre>
NAME A PALETTE
The eight palettes from the training captions. Paste one as the last words of the prompt. The colours before <code>with</code> become the field and the shadows, and the colours after it land on the lit sides.
<ul>
<li><code>teal and petrol blue with tangerine and gold light</code></li>
<li><code>navy with lemon yellow and orange light</code></li>
<li><code>cobalt and ultramarine with peach and rose light</code></li>
<li><code>crimson and wine red with gold and saffron light</code></li>
<li><code>charcoal and slate grey with scarlet light</code></li>
<li><code>plum and burgundy with turquoise and aqua light</code></li>
<li><code>violet and indigo with magenta and coral light</code></li>
<li><code>forest green and emerald with flame orange light</code></li>
</ul>
Give every garment a palette colour. White, khaki or uncoloured clothes resist the palette and stay photo-like. One training image failed twice on a white shirt and khaki trousers, and passed once they were recoloured into its palette.
Keep animals and still lifes concrete. They wander from the prompt more than people do. Name each object, how many, and where it sits.
Widen a close-up that comes out as a photograph. Step back until the whole table, stall or counter and a window are in frame. That fixed a wok cook, a bread board and a bowl of noodle soup. Wide scenes with one strong subject, such as a landscape, an animal or a ship, never needed it.
Name each half of a hybrid creature. A griffin came out as a bald eagle until the prompt said <code>the tawny furred body, back legs, paws and tufted tail of a lion</code>.
RECOMMENDED SETTINGS
<ul>
<li>Checkpoint: <a rel="ugc" href="https://civarchive.com/models/958009?modelVersionId=3139241">RedCraft 3.0 (Krea 2)</a>. The only one I tested.</li>
<li>8 steps, CFG 1, sampler er_sde, scheduler simple.</li>
<li>About 1 MP: 896x1152 portrait, 1152x896 landscape or 1024x1024 square. All three are in the showcase.</li>
<li>Strength 1.0. 0.8 is fine for landscapes and animals, weaker on people and close-ups.</li>
<li>Other checkpoints and other UIs are untested. If you try it on plain Krea 2 Turbo or somewhere else, tell me how it went.</li>
</ul>
WHAT I TESTED
Before release I ran 6 plain prompts it was not trained on, on steps 1000, 1500 and 2000 at 0.8: a lighthouse cottage at dusk, a fisherman mending a net, a coffee and a croissant by a rainy window, a night market, a bear catching a salmon, a spaceship on a red planet.
With no palette, step 1000 broke only the lighthouse and the bear into shards. Step 2000 also took the night market and most of the spaceship, but the fisherman and the coffee stayed near a photograph. 1.0 added a little texture, and 1.2 put iridescent specks on the fisherman and the coffee.
With a training palette added at the end, at 0.8, the lighthouse, the bear and the spaceship matched the training look. At 1.0 the fisherman broke into shards from cap to boots, and the cup and the croissant began to. His face stayed the most realistic part of the picture.
A pass at 1.0 with the trigger left out: the lighthouse came out as strongly as with it. The style lives in the weights, so strength and the palette are the real controls. Keep the trigger anyway, since every caption opened with it.
Then the test against plain Krea 2. The same 6 palette prompts and seeds on plain Krea 2 with no LoRA, with <code>a faceted geometric poster illustration of</code> in place of the trigger. On all 6, plain Krea 2 drew crisp low-poly triangles. On 5 of them it faceted harder than the LoRA at 0.8, and it even faceted the coffee cup, which the LoRA left close to a photograph. The LoRA drew something else: painterly broken-brush shards with hard edges, darker and moodier, the palette's hot colour on the lit side, and a few stray iridescent specks. So it is not a stronger low-poly. It is a different look.
Then the showcase at 1.0, 8 scenes with a different palette on each. All 8 passed on the first render:
<ul>
<li>a mountaineer on a snow ridge at sunrise</li>
<li>a canyon river bend at sunset</li>
<li>a blacksmith at her anvil</li>
<li>a river city at night</li>
<li>a red fox trotting through snow</li>
<li>a shuttle launch under a ringed planet</li>
<li>a fisherman casting a round net from a boat</li>
<li>pomegranates in a copper dish</li>
</ul>
Then 24 more images for the promo posts, from 30 renders. 19 passed on the first render. The misses were three close-ups that came out as photographs, an elephant whose tail tuft read as a hand, and a griffin that came out as an eagle. Saying where the tail hangs fixed the elephant, and the two tips above fixed the rest.
LIMITS
<ul>
<li>For clean low-poly triangles it is the wrong tool. Plain Krea 2 with <code>a faceted geometric poster illustration of</code> does that better, with no LoRA.</li>
<li>Not tested with that phrase and the LoRA in the same prompt.</li>
<li>Tested only on RedCraft 3.0, and only from 0.8 to 1.2.</li>
<li>Faces break up last. At 0.8 they can stay semi-real, and at 1.0 a face can still read more real than the clothes.</li>
<li>At 1.2 and above it breaks into iridescent confetti specks. Even at 1.0 a few stray specks can show, as in the pomegranates.</li>
<li>Leaving out the trigger does not switch the style off. For plain pictures, unload the LoRA.</li>
<li>The base model's pseudo-lettering on lanterns, signs and boat hulls comes through.</li>
<li>Not tested together with character LoRAs, other style LoRAs, Gritty Ink, Dutch Angle or Camera Angles.</li>
</ul>
MAKER'S NOTE
Trained on Krea 2 Raw with ai-toolkit (arch krea2), rank 32, batch 1, lr 2e-4, 1024 px, 2000 steps, about 56 epochs on 36 images. The file is the final step, 2000.
The dataset is synthetic: 36 images rendered on Krea 2 with no LoRA from long style prompts, 12 at each of the three sizes. Landscapes, buildings, rooms, still lifes, animals, 8 people and a few creatures, each in one of the eight palettes. The captions carry only the trigger, the subject and the palette, with no style words.
It is the same recipe as Gritty Ink, where the trigger is the switch. Here the style comes through without it. I do not know why yet.
Free to use. Credit is appreciated. The licence is on this page.
If this worked for you, a thumbs-up here helps the next person find it.

Shard Brush Style for Krea 2

v1, step 2000. The final step of the run. Painted shards with hard edges, like shattered glass or torn coloured paper, in a palette you name.
Not a low-poly LoRA. For clean low-poly triangles, plain Krea 2 with <code>a faceted geometric poster illustration of</code> does better, with no LoRA. The side-by-side test is on the model page.
Start with <code>facetposter</code> and end every prompt with one of the eight palettes on the model page, for example <code>teal and petrol blue with tangerine and gold light</code>.
Strength 1.0. 0.8 works on landscapes and animals. 1.2 overcooks into specks.
Tested only on RedCraft (Krea 2), 8 steps, CFG 1, in ComfyUI.

Files with SHA256: 513e2fc70ad51f002189788a668ad7ac4907fb170c644bf2a39d65ed380356b9