Grain, bloom, and 23 other sliders

Krea 2 grew an unusually good LoRA community, and most of what it produced isn’t style packs — it’s sliders. Nineteen of the twenty-five on our shelf are bipolar: a negative strength runs the effect backwards instead of switching it off. You aren’t choosing a look so much as trimming one.

A LoRA is not a filter

Look again at the four frames at the top of this page, or at any comparison row below. They are not one picture with four treatments applied. They are four different pictures.

A LoRA edits the weights of the model before anything is drawn. Krea 2 Turbo renders in eight steps, so that change is in play from the very first one — the model walks a different path and lands somewhere else. Same seed, same prompt, same settings, and you still get a different composition, a different crop, sometimes a different face. Two things follow, and both matter more than the grade itself:

• You can’t A/B a look in isolation. Every measurement below is reading two things at once: what the LoRA does to the grade, and what it does to the picture the model chose to draw.
• Stacking is not layering. Two LoRAs on one render is not the first one’s output with the second applied over it. It is a third model. Sometimes that lands between the two, sometimes it compounds them, and sometimes it goes somewhere neither went alone.

So the useful question isn’t what does this LoRA do to my image? It’s what kind of image does this model now want to make?

Opposites, on purpose

BloomGirls UltraRealism by coolstrad is a full fine-tune rather than a slider: soft bloom, richer colour, smoother skin, close phone-camera framing. It carries a strong subject bias, so it restyles faces and hair as well as the grade — treat it as a look to build on, not a finishing pass. It holds together to about 1.0; by 1.5 flat surfaces start to patch, and at 2.0 fine texture smears. The slider runs 0 to 2 and defaults to 0.8.

Purple Grainy by purplelady bakes in the post-processing pass its author would otherwise apply by hand: film grain, softened edges, and a muted grade that pulls saturation down and cools the midtones. It is bipolar, so negative strengths run the other way into a cleaner, punchier, grain-free digital look. Detail survives all the way to 2.0, but past about 1.5 it starts drawing an instant-film print border around the frame. The slider runs −2 to +2 and defaults to 1.

Being opposites is also why the community tends to run them together rather than picking one. The pairing has become a staple of AI influencer and lifestyle content, where the picture has to read as real and aspirational at the same time: BloomGirls supplies the aspirational half — the polish, the bloom, the close phone-camera intimacy — and Purple Grainy puts back the grain and the slightly-off colour that keep it from looking rendered. Frames 04 and 06 below are the clearest cases of that trade landing.

Six scenes, four renders each

Every row is one seed, one size, one step count, with only the LoRAs changing. All twenty-four renders came off GPUs in the Sogni network — the same hardware that serves everyone else’s jobs. Frames 01–03 are from the validation sweep that set the published ranges; 04–06 were written specifically for grain, skin and hard light and went through as ordinary jobs, landing on whichever nodes happened to be free. Purple Grainy runs at 1.0 everywhere except frame 01, where the prompt named 1.2.

Worker nodes aren’t identical — they’re tuned around different attention backends, so the same prompt at the same seed won’t necessarily land on the same pixels twice. Read any single row as indicative rather than definitive; the evidence for what these two LoRAs do is the pattern repeating across all six scenes, on hardware nobody hand-picked.

Under each row is a signal readout. It measures four things about the frame as rendered, each one relative to that scene’s own no-LoRA baseline — so the notch at the middle of every track is 100%, and the dots show how far each version moved.

Frame 01

Long exposure on red

This prompt already asks for film grain, so Purple Grainy is pushing on an open door: grain and edge energy both more than double, and the red goes muddier and warmer. BloomGirls goes the other way and cleans the frame up — fewer high frequencies, glossier leather, a face pulled toward the influencer register. Stacked lands between the two, which is the most useful thing this pair does.

Prompt

film grain, faded color tones, low contrast, long shadows, almost cinematic feel. A young woman with dark makeup and elegant styling against a solid deep red background, black leather jacket, neutral contemplative expression. A long-exposure technique blurs her figure while her face and shoulders stay relatively sharp. Minimalist, moody, introspective.

Measured against this scene’s own no-LoRA render. The notch on each track is 100%.

Signal	No LoRA	BloomGirls	Purple Grainy	Both
Grainhigh-frequency residual against a 3×3 blur	100%	89%	230%	132%
Saturationmean channel spread	100%	110%	115%	124%
Edge energymean luma gradient — falling means softening	100%	89%	234%	136%
Luminancemean brightness	100%	93%	137%	118%

Frame 02

Black cat in the bubbles

The cleanest single read of the six, because the subject is one texture. The baseline is a cool, almost-white field with the cat barely emerging. BloomGirls warms the foam to cream, glosses the eyes and blooms the whole frame. Purple Grainy takes the whites down to a grey-green and lays visible speckle over the bubbles — the only one of the four that looks printed rather than rendered. Stacked, you get BloomGirls’ warmth and gloss with the speckle back on top.

Prompt

film grain, faded color tones, artistic framing, almost cinematic feel, moody photography, dramatic composition. A close-up, mysterious photo of a black cat taking a bubble bath. The thick white bath bubbles cover the entire frame, with only the cat’s nose and a little fur around the mouth poking out. Soft lighting, minimal background, dreamy and surreal atmosphere.

Measured against this scene’s own no-LoRA render. The notch on each track is 100%.

Signal	No LoRA	BloomGirls	Purple Grainy	Both
Grainhigh-frequency residual against a 3×3 blur	100%	97%	170%	120%
Saturationmean channel spread	100%	150%	89%	149%
Edge energymean luma gradient — falling means softening	100%	95%	168%	119%
Luminancemean brightness	100%	96%	90%	88%

Frame 03

Palladium selfie

The baseline is the most convincingly lo-fi frame in the set — crushed shadows, real sensor noise, a marquee blowing out. BloomGirls undoes exactly that: edge energy drops to 77%, skin brightens and smooths, and the picture becomes a posted selfie rather than a found one. Purple Grainy barely moves the needle here, because the prompt had already asked for everything it adds.

Prompt

An authentic, high-angle smartphone selfie at a concert venue, brightly lit marquee behind her, dimly lit with mixed uneven lighting, harsh cool backlight and blown-out highlights, warm tungsten cast, deep crushed shadows. Grainy with significant digital sensor noise, soft focus, distinct dutch angle, nightlife aesthetic, candid IG post.

Measured against this scene’s own no-LoRA render. The notch on each track is 100%.

Signal	No LoRA	BloomGirls	Purple Grainy	Both
Grainhigh-frequency residual against a 3×3 blur	100%	91%	108%	97%
Saturationmean channel spread	100%	157%	113%	148%
Edge energymean luma gradient — falling means softening	100%	77%	102%	83%
Luminancemean brightness	100%	113%	100%	105%

Frame 04

Butter-yellow studio glam

The scene that proves Purple Grainy is a grade and not a contrast boost: against a flat butter-yellow seamless it pulls saturation down to 89% and cools the wall toward chalk. BloomGirls does the opposite, warming the backdrop and opening the shadow inside the coat. Both together keep BloomGirls’ colour at Purple Grainy’s lower key — the most magazine-like frame of the four.

Prompt

A glam studio portrait against a solid smooth butter-yellow seamless backdrop filling the frame edge to edge. Low angle, full body, preppy collegiate styling with an oversized black overcoat over the shoulders. A nostalgic instant-film snapshot with the soft contrast and creamy color of a Polaroid, faint vignetting, gentle overexposure, warm chemical color cast. Soft even frontal light with a bright catchlight in each eye.

Measured against this scene’s own no-LoRA render. The notch on each track is 100%.

Signal	No LoRA	BloomGirls	Purple Grainy	Both
Grainhigh-frequency residual against a 3×3 blur	100%	95%	96%	98%
Saturationmean channel spread	100%	124%	89%	126%
Edge energymean luma gradient — falling means softening	100%	94%	101%	98%
Luminancemean brightness	100%	100%	93%	91%

Frame 05

Minimalist living room, 1990s

Same room, four film stocks. BloomGirls brightens and warms it into furniture-catalogue light. Purple Grainy cools it and drops the key. Stacked is the darkest and lowest-key of the four at 79% luminance — here the pair makes the picture quieter, not louder.

Prompt

Film grain, faded color tones, emotional candid photography, long shadows, almost cinematic feel, 1990s vibe. A serene, intimate scene in a modern minimalist living room: cream shag rug, small round wooden coffee table, a yellow book titled “The Art of Grainy,” a light blue sectional sofa, a large window onto an urban cityscape. Natural light streaming in, cosy and contemplative.

Measured against this scene’s own no-LoRA render. The notch on each track is 100%.

Signal	No LoRA	BloomGirls	Purple Grainy	Both
Grainhigh-frequency residual against a 3×3 blur	100%	116%	102%	98%
Saturationmean channel spread	100%	162%	110%	139%
Edge energymean luma gradient — falling means softening	100%	110%	111%	97%
Luminancemean brightness	100%	103%	88%	79%

Frame 06

Extreme close-up selfie

The skin test, and the clearest argument for reading these two as opposites. At 1:1 the baseline keeps freckles, pores and fine noise. BloomGirls removes all three, enlarges the eyes and lips and drops grain to 84%. Purple Grainy keeps the texture and adds its own on top at 141%. Stacked gives you BloomGirls’ face with grain re-applied over it — which reads as film, not as skin.

Prompt

An indoor close-up selfie, held very close so a smartphone wide-angle slightly distorts the face. Warm yellowish indoor incandescent cast, moderate saturation, low contrast, soft and grainy. A tender candid 35mm film portrait on Kodak Portra 400 — soft fine grain, gentle pastel color, mild halation around the highlights, delicate soft-focus falloff. Background out of focus but visible.

Measured against this scene’s own no-LoRA render. The notch on each track is 100%.

Signal	No LoRA	BloomGirls	Purple Grainy	Both
Grainhigh-frequency residual against a 3×3 blur	100%	84%	141%	117%
Saturationmean channel spread	100%	118%	102%	114%
Edge energymean luma gradient — falling means softening	100%	84%	143%	112%
Luminancemean brightness	100%	101%	109%	100%

420 px — 1:1 where the screen allows

Faces at 1:1, matched to the subject rather than to a fixed pixel box. This is the only honest way to judge what happens to skin.

The rest of the shelf

Two creators account for most of it: Loraholic with ten and alcaitiff with nine. Each card below shows the range the slider actually offers in the app, with the recommended band in mint, the default marked, and zero marked where the slider crosses it. Every name links to the creator’s original model page.

Working with them

• Up to eight at once, on any Krea 2 model — Turbo, Dark Beast and the identity-edit variants all draw from the same shelf.
• Order matters. These models run fp8-quantized and the patches do not commute, so the same set in a different order is a measurably different image.
• Negative is not off. Zero is off. Negative is the inverse, and on the bipolar sliders it is often the more interesting direction — Purple Grainy at −1 is a clean, saturated digital look you can’t reliably prompt for.
• Watch what the prompt already asked for. Frame 03 is the lesson: a LoRA that adds what the prompt already had barely registers, and a look LoRA pointed at a prompt built around restraint will simply overrule it.
• Lock the seed while you tune. One variable at a time is the only way to tell a slider apart from a re-roll.