shiccup.

Updated Nov 11, 2026

Local AI models, compared

The open models we use to make our games, next to the newest ones on Hugging Face. Same prompts, same seeds, one home PC (one NVIDIA RTX 3090 (24 GB)). Tap any picture to see it full size.

Pictures

Game art from a text prompt: textures, sprites, icons, backgrounds, item art. Same prompts and seeds for every model, 1024 x 1024.

We use FLUX 3 Image Klein (10.7 s each).

Pictures: speed, memory, release date and licence of each model, on one NVIDIA RTX 3090 (24 GB)
ModelReleasedSpeed (seconds per picture)Graphics memoryLicence
FLUX 3 Image KleinBlack Forest Labs · bf16, 15 stepsWhat we useNewestOct 13, 202610.7 sfirst load 33.1 s15.2 GBFLUX Kommunity LicenseFree for businesses under $5M a year; outputs reviewed before use; AI use disclosed where the law asks.
Krea 2 TurboKrea · fp8Used beforeJun 16, 2026about (first public copy)15.9 sfirst load 38.2 s18.3 GBKrea 2 Community LicenseFree for businesses under $1M a year.
MOSS-Image BaseOpenMOSS · int8, 25 stepsTestedOct 11, 202617.1 sfirst load 39.9 s11.6 GBApache-2.0Free for any use.

Long captions

app icon

Prompt: Glossy stylized digital illustration of a mobile game app icon: a rounded square icon shape that fills the whole frame, with softly rounded corners and a smooth deep-blue to purple gradient background. In the center, a bright orange basketball with black seams drops through an orange metal hoop with a white net; the ball is just passing the rim, tilted slightly toward frame-left. Bold flames in yellow, orange and red sweep behind the ball in a curved swoosh from the upper left down toward the hoop, as if the ball blazed through the air. The ball and rim carry a strong glossy highlight on their upper surfaces, and a soft inner glow lights the area around the flames. Bold, saturated colors with clean shapes and crisp outlines, readable at a small size. The icon holds only the ball, flames, hoop and net on the gradient background, with plain unmarked surfaces.

FLUX 3 Image Klein (ours)upsampled caption (FLUX 3 system prompt, same for every model); 1024x1024, seed 11; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboupsampled caption (FLUX 3 system prompt, same for every model); 1024x1024, seed 11; fp8 (our default)
MOSS-Image Baseupsampled caption (FLUX 3 system prompt, same for every model); 1024x1024, seed 11; int8 base (ComfyUI template: 25 steps, cfg 2.5)

card art

Prompt: Rich painterly fantasy illustration for a card game unit portrait, vertical composition with the figure filling most of the frame and the painting running to all four edges. A stout armored blacksmith knight stands in the center foreground in a low three-quarter view, a broad short man in dented steel plate armor over a scorched leather apron, wearing a full helmet with a narrow visor. His right arm raises a heavy hammer high above his head toward the upper right, its head glowing bright molten orange; sparks burst from it and fly across the upper half of the frame. His left gloved hand rests on an anvil at frame-left. Behind him, a dark stone forge fills the background, its furnace mouth glowing deep red in the lower right and dim chains hanging from the shadowy ceiling. Dramatic lighting: the hammer's orange glow lights the top of his helmet and pauldrons, while cool blue shadow wraps his back and the room. Visible brushwork, deep shadows, warm-cool contrast. The image is the bare painting, with no frame, no border, no lettering and no card layout elements.

FLUX 3 Image Klein (ours)upsampled caption (FLUX 3 system prompt, same for every model); 768x1024, seed 11; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboupsampled caption (FLUX 3 system prompt, same for every model); 768x1024, seed 11; fp8 (our default)
MOSS-Image Baseupsampled caption (FLUX 3 system prompt, same for every model); 768x1024, seed 11; int8 base (ComfyUI template: 25 steps, cfg 2.5)

golfer sprite

Prompt: Stylized 3D cartoon render of a full-body game character sprite, side view facing frame-left, centered in the frame with empty space around the figure. A friendly older man in his sixties with fair skin, rosy cheeks, a short white mustache and a round belly stands mid backswing: both hands grip a silver golf club raised over his right shoulder, the club head pointing up and back toward the upper right, his weight shifted onto his back foot and his front heel slightly lifted. He wears an orange bucket hat, a white short-sleeved polo shirt stretched over his belly, khaki trousers with a brown belt, and brown-and-white golf shoes. His eyes look toward the lower left where the ball would be, with a small content smile. Soft even studio lighting from the upper left gives smooth shading and a gentle rim of light on his back, and a faint soft contact shadow sits under his shoes. The figure has a clean, readable silhouette with simple rounded forms and slightly oversized hands and head. Plain flat light-grey background, uniform from edge to edge, with nothing else in the frame.

FLUX 3 Image Klein (ours)upsampled caption (FLUX 3 system prompt, same for every model); 1024x1024, seed 11; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboupsampled caption (FLUX 3 system prompt, same for every model); 1024x1024, seed 11; fp8 (our default)
MOSS-Image Baseupsampled caption (FLUX 3 system prompt, same for every model); 1024x1024, seed 11; int8 base (ComfyUI template: 25 steps, cfg 2.5)

item lamp

Prompt: Hand-painted stylized 3D render of a single game item: a vintage brass table lamp in three-quarter view, centered in the frame with generous empty space around it. A round stepped brass base with a slightly worn polished rim supports a short turned brass column with two decorative rings, topped by a dome-shaped shade of translucent emerald-green glass with a flared lower lip. A small brass pull chain hangs from under the shade on the right side. Soft key light from the upper left makes the brass glow warm gold with smooth painted highlights and darker bronze in the recesses, and the green glass shows a gentle light gradient from bright lime at the top edge to deep green at the rim. Soft, simple shading with slightly exaggerated rounded forms. A faint soft shadow sits directly under the base. Plain white background, uniform edge to edge, with the lamp as the only object in the frame.

FLUX 3 Image Klein (ours)upsampled caption (FLUX 3 system prompt, same for every model); 1024x1024, seed 21; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboupsampled caption (FLUX 3 system prompt, same for every model); 1024x1024, seed 21; fp8 (our default)
MOSS-Image Baseupsampled caption (FLUX 3 system prompt, same for every model); 1024x1024, seed 21; int8 base (ComfyUI template: 25 steps, cfg 2.5)

level bg

Prompt: Soft painted digital illustration of a side-scrolling platform game level background on a wide canvas, a sunny fantasy forest clearing built in layered parallax planes. In the foreground along the bottom edge, a flat strip of grass-topped brown earth runs the full width of the frame as a level walking surface. In the midground, giant mushrooms stand at frame-left and frame-right: tall pale-cream stems with broad red caps dotted with white spots, the left one larger and closer, between them a few clusters of small orange mushrooms and round green bushes. Tall trees with dense green canopies frame the upper corners, their trunks partly overlapping the mushrooms. In the background, rolling green hills fade to soft blue-green with distance under a light-blue sky with a few white rounded clouds. Warm late-morning sunlight from the upper left casts soft shadows to the right and gentle light beams through the canopy. The clearing is empty of people and animals. Smooth brushwork, slightly desaturated distant layers, saturated foreground colors.

FLUX 3 Image Klein (ours)upsampled caption (FLUX 3 system prompt, same for every model); 1344x768, seed 11; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboupsampled caption (FLUX 3 system prompt, same for every model); 1344x768, seed 11; fp8 (our default)
MOSS-Image Baseupsampled caption (FLUX 3 system prompt, same for every model); 1344x768, seed 11; int8 base (ComfyUI template: 25 steps, cfg 2.5)

pixel knight

Prompt: 16-bit pixel art game sprite of a small knight in an idle pose, centered in the frame and drawn large so the sprite fills about half of the frame height, with crisp hard-edged square pixels and no anti-aliasing or blur. The knight faces frame-right in a three-quarter view: a silver-grey helmet with a dark eye slit and a small blue plume, silver armor with darker grey pixel shading, a flowing blue cape behind him that falls to his ankles, brown boots, and a short silver sword held upright in his right hand with a brown grip. A limited palette of about twelve colors: three greys, three blues, two browns, black outline and a white highlight pixel on the helmet and blade. Each shape is outlined in a one-pixel dark outline. Plain flat pale-cream background with no gradient, nothing else in the frame.

FLUX 3 Image Klein (ours)upsampled caption (FLUX 3 system prompt, same for every model); 1024x1024, seed 11; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboupsampled caption (FLUX 3 system prompt, same for every model); 1024x1024, seed 11; fp8 (our default)
MOSS-Image Baseupsampled caption (FLUX 3 system prompt, same for every model); 1024x1024, seed 11; int8 base (ComfyUI template: 25 steps, cfg 2.5)

text sign

Prompt: Glossy cartoon game title screen logo on a wide canvas, centered on a plain flat sky-blue background. Two lines of big chunky rounded cartoon letters read "PEBBLE" on the top line and "PATROL" on the bottom line, the text spelled exactly "PEBBLE PATROL". The letters are bright orange with a thick dark-brown outline, a lighter orange highlight along their top edges and a soft dark drop shadow offset down and to the right. The words are slightly bouncy, each letter tilted a few degrees in alternating directions, and the two lines are the same width and stacked tightly. A small smiling round grey pebble character with two black dot eyes and a little curved smile peeks over the top of the second word "PATROL", its tiny grey hands gripping the top edge of the letter "T". Soft even lighting with a gentle shine on the letters. Nothing else is in the frame: the logo sits alone on the flat sky-blue color.

FLUX 3 Image Klein (ours)upsampled caption (FLUX 3 system prompt, same for every model); 1344x768, seed 31; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboupsampled caption (FLUX 3 system prompt, same for every model); 1344x768, seed 31; fp8 (our default)
MOSS-Image Baseupsampled caption (FLUX 3 system prompt, same for every model); 1344x768, seed 31; int8 base (ComfyUI template: 25 steps, cfg 2.5)

tile cobble

Prompt: Hand-painted stylized digital texture of a cobblestone street seen straight from above, a seamless tileable game terrain tile that fills the whole square frame edge to edge with no border, horizon or perspective. Rounded grey and warm taupe cobblestones of slightly uneven size sit in loose staggered rows across the entire surface, each stone with a soft painted highlight on its upper-left edge and a darker brushed shadow along its lower-right edge. Narrow gaps between the stones are filled with dark brown packed earth and a few tiny tufts of muted green moss. A handful of stones carry small chips and faint hairline cracks. Even, flat, shadowless overhead lighting with no light falloff, no vignette and no single bright spot, so every part of the frame has the same brightness. The stones continue past all four edges, cut off cleanly by the frame so the pattern repeats when tiled. Brush texture is visible but subtle, with a limited palette of cool greys, warm beige and dark umber.

FLUX 3 Image Klein (ours)upsampled caption (FLUX 3 system prompt, same for every model); 1024x1024, seed 11; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboupsampled caption (FLUX 3 system prompt, same for every model); 1024x1024, seed 11; fp8 (our default)
MOSS-Image Baseupsampled caption (FLUX 3 system prompt, same for every model); 1024x1024, seed 11; int8 base (ComfyUI template: 25 steps, cfg 2.5)

Short prompts

app icon

Prompt: Mobile game app icon: a basketball with flames swooshing through a hoop, bold saturated colours, glossy, rounded square icon filling the frame, no text

FLUX 3 Image Klein (ours)short prompt as we write them; 1024x1024, seed 11; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboshort prompt as we write them; 1024x1024, seed 11; fp8 (our default)
MOSS-Image Baseshort prompt as we write them; 1024x1024, seed 11; int8 base (ComfyUI template: 25 steps, cfg 2.5)

card art

Prompt: Card game unit portrait illustration: a stout armored blacksmith knight raising a glowing orange hammer, sparks flying, dark forge behind, dramatic lighting, rich painterly fantasy art, vertical card composition, no text, no border

FLUX 3 Image Klein (ours)short prompt as we write them; 768x1024, seed 11; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboshort prompt as we write them; 768x1024, seed 11; fp8 (our default)
MOSS-Image Baseshort prompt as we write them; 768x1024, seed 11; int8 base (ComfyUI template: 25 steps, cfg 2.5)

golfer sprite

Prompt: Full-body game character sprite of a friendly older golfer with a round belly, orange bucket hat, white polo shirt, khaki trousers, mid backswing, side view, stylized 3D cartoon look, clean silhouette, plain flat light-grey background, no text

FLUX 3 Image Klein (ours)short prompt as we write them; 1024x1024, seed 11; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboshort prompt as we write them; 1024x1024, seed 11; fp8 (our default)
MOSS-Image Baseshort prompt as we write them; 1024x1024, seed 11; int8 base (ComfyUI template: 25 steps, cfg 2.5)

item lamp

Prompt: A vintage brass table lamp with a green glass shade, single game item, hand-painted stylized 3D look, soft shading, three-quarter view, centered, plain white background, no text

FLUX 3 Image Klein (ours)short prompt as we write them; 1024x1024, seed 21; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboshort prompt as we write them; 1024x1024, seed 21; fp8 (our default)
MOSS-Image Baseshort prompt as we write them; 1024x1024, seed 21; int8 base (ComfyUI template: 25 steps, cfg 2.5)

level bg

Prompt: Side-scrolling platform game level background: a sunny fantasy forest clearing with giant mushrooms, distant hills, layered parallax depth, soft painted style, flat ground strip along the bottom for the player to walk on, no characters, no text

FLUX 3 Image Klein (ours)short prompt as we write them; 1344x768, seed 11; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboshort prompt as we write them; 1344x768, seed 11; fp8 (our default)
MOSS-Image Baseshort prompt as we write them; 1344x768, seed 11; int8 base (ComfyUI template: 25 steps, cfg 2.5)

pixel knight

Prompt: 16-bit pixel art game sprite of a small knight with a blue cape and sword, idle pose, crisp pixels, limited palette, centered on a plain flat background, no text

FLUX 3 Image Klein (ours)short prompt as we write them; 1024x1024, seed 11; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboshort prompt as we write them; 1024x1024, seed 11; fp8 (our default)
MOSS-Image Baseshort prompt as we write them; 1024x1024, seed 11; int8 base (ComfyUI template: 25 steps, cfg 2.5)

text sign

Prompt: Game title screen logo: the words "PEBBLE PATROL" in big chunky rounded cartoon letters, orange with a thick dark-brown outline and a soft drop shadow, a small smiling grey pebble character peeking over the second word, centered on a plain flat sky-blue background, the text spelled exactly PEBBLE PATROL

FLUX 3 Image Klein (ours)short prompt as we write them; 1344x768, seed 31; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboshort prompt as we write them; 1344x768, seed 31; fp8 (our default)
MOSS-Image Baseshort prompt as we write them; 1344x768, seed 31; int8 base (ComfyUI template: 25 steps, cfg 2.5)

tile cobble

Prompt: Seamless tileable texture, top-down stylized hand-painted cobblestone street, even lighting, game terrain tile, fills the whole frame

FLUX 3 Image Klein (ours)short prompt as we write them; 1024x1024, seed 11; bf16 distilled transformer (on a 3090 FP8 saves memory, not time; NVFP4 needs Blackwell)
Krea 2 Turboshort prompt as we write them; 1024x1024, seed 11; fp8 (our default)
MOSS-Image Baseshort prompt as we write them; 1024x1024, seed 11; int8 base (ComfyUI template: 25 steps, cfg 2.5)

Background removal

Cutting a subject out of its background, with the hard cases: fur, glass, smoke and thin wires.

We use BiRefNet HR-matting (0.15 s each).

Background removal: speed, memory, release date and licence of each model, on one NVIDIA RTX 3090 (24 GB)
ModelReleasedSpeed (seconds per picture)Graphics memoryLicence
BiRefNet HR-mattingZheng Peng et al.What we useNewestFeb 12, 20250.15 sfirst load 8.0 s1.7 GBMITFree for any use.
BiRefNetZheng Peng et al.Used beforeJul 12, 20240.15 sfirst load 16.6 s1.7 GBMITFree for any use.
BEN2Prama LLCTestedJan 22, 20250.16 sfirst load 7.4 s2.5 GBMITFree for any use.
Background:

fluffy cat

Prompt: fur: A very fluffy long-haired white persian cat sitting, soft wispy fur edges and long whiskers, stylized 3D cartoon look, full body, plain flat medium-grey background, no text

Original picture
BiRefNet HR-matting (ours)transparent WebP
BiRefNettransparent WebP; timing from the same weights in the bake-off script
BEN2transparent WebP

glass bottle

Prompt: glass: A clear glass potion bottle half full of glowing green liquid with a cork, transparent glass you can see through, single game item, stylized 3D look, plain flat white background, no text

Original picture
BiRefNet HR-matting (ours)transparent WebP
BiRefNettransparent WebP; timing from the same weights in the bake-off script
BEN2transparent WebP

smoke puff

Prompt: smoke: A soft cartoon puff of grey smoke with thin semi-transparent wisps trailing off, single game effect, plain flat dark-blue background, no text

Original picture
BiRefNet HR-matting (ours)transparent WebP
BiRefNettransparent WebP; timing from the same weights in the bake-off script
BEN2transparent WebP

wires

Prompt: thin wires: A tangle of thin black electrical wires and a small old desk fan with a fine wire grille, single game prop, stylized 3D look, plain flat light-grey background, no text

Original picture
BiRefNet HR-matting (ours)transparent WebP
BiRefNettransparent WebP; timing from the same weights in the bake-off script
BEN2transparent WebP

Music

Game music loops from a text description. Clips play their first 25 seconds.

We use MiniMax Music 3 (19.6 s each).

Music: speed, memory, release date and licence of each model, on one NVIDIA RTX 3090 (24 GB)
ModelReleasedSpeed (seconds per track)Graphics memoryLicence
MiniMax Music 3MiniMaxWhat we useNewestAug 8, 2026about (first public copy)19.6 sfirst load 20.1 s14.7 GBMiniMax-Music3 Community LicenseFree for businesses under $20M a year; products show "MiniMax-Music3"; music posted publicly says it is AI-made.
Stable Audio 3 MediumStability AIUsed beforeMay 17, 20263.8 sfirst load 9.6 s6.3 GBStability AI Community LicenseFree for businesses under $1M a year.

aim (take 7301)

Prompt: a light suspenseful toy-theatre waltz with a pizzicato bass and soft marimba, calm, a little playful tension while aiming, about 96 beats per minute, loops cleanly, no drums

MiniMax Music 3 (ours)asked for 40 s, made 22 s; seed 7301
Stable Audio 3 Mediumasked for 40 s, made 40 s, first 25 s shown; seed 7301

aim (take 7302)

Prompt: a light suspenseful toy-theatre waltz with a pizzicato bass and soft marimba, calm, a little playful tension while aiming, about 96 beats per minute, loops cleanly, no drums

MiniMax Music 3 (ours)asked for 40 s, made 16 s; seed 7302
Stable Audio 3 Mediumasked for 40 s, made 40 s, first 25 s shown; seed 7302

flight (take 7301)

Prompt: a bright bouncy paper-theatre polka with a soft snare brush, pizzicato strings and a light whistle line, lively but gentle, about 118 beats per minute, loops cleanly

MiniMax Music 3 (ours)asked for 40 s, made 12 s; seed 7301
Stable Audio 3 Mediumasked for 40 s, made 40 s, first 25 s shown; seed 7301

flight (take 7302)

Prompt: a bright bouncy paper-theatre polka with a soft snare brush, pizzicato strings and a light whistle line, lively but gentle, about 118 beats per minute, loops cleanly

MiniMax Music 3 (ours)asked for 40 s, made 14 s; seed 7302
Stable Audio 3 Mediumasked for 40 s, made 40 s, first 25 s shown; seed 7302

menu (take 7301)

Prompt: a warm wistful music-box and soft plucked guitar theme in a major key, a gentle lullaby tempo of about 86 beats per minute, a small paper-theatre orchestra with muted strings and a soft glockenspiel line, calm and friendly, loops cleanly

MiniMax Music 3 (ours)asked for 40 s, made 11 s; seed 7301
Stable Audio 3 Mediumasked for 40 s, made 40 s, first 25 s shown; seed 7301

menu (take 7302)

Prompt: a warm wistful music-box and soft plucked guitar theme in a major key, a gentle lullaby tempo of about 86 beats per minute, a small paper-theatre orchestra with muted strings and a soft glockenspiel line, calm and friendly, loops cleanly

MiniMax Music 3 (ours)asked for 40 s, made 14 s; seed 7302
Stable Audio 3 Mediumasked for 40 s, made 40 s, first 25 s shown; seed 7302

Sound effects

Short game sounds from a text description: clicks, coins, footsteps, explosions.

We use Stable Audio 3 Medium (1.9 s each).

Sound effects: speed, memory, release date and licence of each model, on one NVIDIA RTX 3090 (24 GB)
ModelReleasedSpeed (seconds per sound)Graphics memoryLicence
Stable Audio 3 MediumStability AIWhat we useNewestMay 17, 20261.9 sfirst load 4.4 s6.3 GBStability AI Community LicenseFree for businesses under $1M a year.
Stable Audio 3 Small SFXStability AITestedNewestMay 17, 20261.3 sfirst load 21.0 s3.0 GBStability AI Community LicenseFree for businesses under $1M a year.

coin pickup (take 8101)

Prompt: a bright cheerful coin pickup chime, two quick rising metallic notes, sparkly, game sound effect

Stable Audio 3 Medium (ours)1.5 s asked, 1.5 s made; seed 8101
Stable Audio 3 Small SFX1.5 s asked, 1.5 s made; seed 8101

coin pickup (take 8102)

Prompt: a bright cheerful coin pickup chime, two quick rising metallic notes, sparkly, game sound effect

Stable Audio 3 Medium (ours)1.5 s asked, 1.5 s made; seed 8102
Stable Audio 3 Small SFX1.5 s asked, 1.5 s made; seed 8102

explosion (take 8101)

Prompt: a deep cartoon explosion with a punchy boom and rumbling debris tail, not harsh

Stable Audio 3 Medium (ours)3.5 s asked, 3.5 s made; seed 8101
Stable Audio 3 Small SFX3.5 s asked, 3.5 s made; seed 8101

explosion (take 8102)

Prompt: a deep cartoon explosion with a punchy boom and rumbling debris tail, not harsh

Stable Audio 3 Medium (ours)3.5 s asked, 3.5 s made; seed 8102
Stable Audio 3 Small SFX3.5 s asked, 3.5 s made; seed 8102

footsteps gravel (take 8101)

Prompt: four footsteps on loose gravel, crunchy, close microphone, outdoors, no music

Stable Audio 3 Medium (ours)3.0 s asked, 3.0 s made; seed 8101
Stable Audio 3 Small SFX3.0 s asked, 3.0 s made; seed 8101

footsteps gravel (take 8102)

Prompt: four footsteps on loose gravel, crunchy, close microphone, outdoors, no music

Stable Audio 3 Medium (ours)3.0 s asked, 3.0 s made; seed 8102
Stable Audio 3 Small SFX3.0 s asked, 3.0 s made; seed 8102

glass break (take 8101)

Prompt: a glass bottle shattering on a stone floor, bright tinkling shards scattering

Stable Audio 3 Medium (ours)2.5 s asked, 2.5 s made; seed 8101
Stable Audio 3 Small SFX2.5 s asked, 2.6 s made; seed 8101

glass break (take 8102)

Prompt: a glass bottle shattering on a stone floor, bright tinkling shards scattering

Stable Audio 3 Medium (ours)2.5 s asked, 2.5 s made; seed 8102
Stable Audio 3 Small SFX2.5 s asked, 2.6 s made; seed 8102

jump boing (take 8101)

Prompt: a cartoon jump sound, a springy rubbery boing with a quick upward pitch bend

Stable Audio 3 Medium (ours)1.5 s asked, 1.5 s made; seed 8101
Stable Audio 3 Small SFX1.5 s asked, 1.5 s made; seed 8101

jump boing (take 8102)

Prompt: a cartoon jump sound, a springy rubbery boing with a quick upward pitch bend

Stable Audio 3 Medium (ours)1.5 s asked, 1.5 s made; seed 8102
Stable Audio 3 Small SFX1.5 s asked, 1.5 s made; seed 8102

power up (take 8101)

Prompt: a magical power-up sound, a rising shimmering synth sweep with soft sparkles at the top

Stable Audio 3 Medium (ours)2.5 s asked, 2.5 s made; seed 8101
Stable Audio 3 Small SFX2.5 s asked, 2.6 s made; seed 8101

power up (take 8102)

Prompt: a magical power-up sound, a rising shimmering synth sweep with soft sparkles at the top

Stable Audio 3 Medium (ours)2.5 s asked, 2.5 s made; seed 8102
Stable Audio 3 Small SFX2.5 s asked, 2.6 s made; seed 8102

sword swish (take 8101)

Prompt: a fast sword swing whoosh through the air followed by a sharp metal clang

Stable Audio 3 Medium (ours)2.0 s asked, 2.0 s made; seed 8101
Stable Audio 3 Small SFX2.0 s asked, 2.0 s made; seed 8101

sword swish (take 8102)

Prompt: a fast sword swing whoosh through the air followed by a sharp metal clang

Stable Audio 3 Medium (ours)2.0 s asked, 2.0 s made; seed 8102
Stable Audio 3 Small SFX2.0 s asked, 2.0 s made; seed 8102

ui click (take 8101)

Prompt: a soft rounded plastic button click for a game menu, short and clean, dry, no reverb

Stable Audio 3 Medium (ours)1.0 s asked, 1.0 s made; seed 8101
Stable Audio 3 Small SFX1.0 s asked, 1.1 s made; seed 8101

ui click (take 8102)

Prompt: a soft rounded plastic button click for a game menu, short and clean, dry, no reverb

Stable Audio 3 Medium (ours)1.0 s asked, 1.0 s made; seed 8102
Stable Audio 3 Small SFX1.0 s asked, 1.1 s made; seed 8102

Voice

Spoken game lines. Some models pick from fixed voices; others make a voice from a written description.

We use Ming-3.0 TTS 0.5B (10.7 s each).

Voice: speed, memory, release date and licence of each model, on one NVIDIA RTX 3090 (24 GB)
ModelReleasedSpeed (seconds per line)Graphics memoryLicence
Ming-3.0 TTS 0.5BinclusionAIWhat we useNewestSep 22, 202610.7 sfirst load 4.5 s3.0 GBMITFree for any use.
Kokoro-82MhexgradTestedDec 26, 20240.13 sfirst load 7.0 s1.2 GBApache-2.0Free for any use.

boss

Prompt: Watch out, the storm king is coming. Grab your shield and hold the gate!

Ming-3.0 TTS 0.5B (ours)voice from a description: A friendly older British man, warm and gently humorous, relaxed pace.; 4.9 s of speech
Ming-3.0 TTS 0.5B (ours)voice from a description: A bright, cheerful young woman, upbeat and clear, like a game show host.; 3.4 s of speech
Kokoro-82Mvoice bm_george; 5.12 s of speech
Kokoro-82Mvoice af_heart; 4.47 s of speech

welcome

Prompt: Welcome back! Your garden grew three new flowers while you were away.

Ming-3.0 TTS 0.5B (ours)voice from a description: A friendly older British man, warm and gently humorous, relaxed pace.; 4.3 s of speech
Ming-3.0 TTS 0.5B (ours)voice from a description: A bright, cheerful young woman, upbeat and clear, like a game show host.; 4.0 s of speech
Kokoro-82Mvoice bm_george; 4.75 s of speech
Kokoro-82Mvoice af_heart; 4.2 s of speech

New open models we're watching

The most-trending new open models on Hugging Face as of Nov 11, 2026, one list per kind of model. Copies of other models and models that can't be used commercially are left out. A model here is a candidate for the side-by-side test, not a recommendation.

Pictures

  • Linum-AI/pjit-21kReleased Nov 3, 2026 · 70 likes · 2.2B params (~4 GB in bf16, fits on 24 GB)Licence: apache-2.0
  • OpenMOSS-Team/MOSS-Image-SFTReleased Oct 11, 2026 · 942 likes · 5.1B params (~10 GB in bf16, fits on 24 GB)Licence: apache-2.0
  • briaai/Fibo-2.0-devReleased Oct 18, 2026 · 214 likes · 8.1B params (~16 GB in bf16, fits on 24 GB)Licence: other (bria-fibo) (custom: read before use)

Picture editing

  • briaai/Fibo-2.0-edit-turboReleased Oct 19, 2026 · 179 likes · 8.1B params (~16 GB in bf16, fits on 24 GB)Licence: other (bria-fibo) (custom: read before use)

Background removal

  • nikopueringer/CorridorKey2.0Released Oct 13, 2026 · 46 likes · size not listedLicence: other (corridor-key-licence-2.0) (custom: read before use)

Voice

  • Zyphra/ZONOS3-MiniReleased Nov 9, 2026 · 124 likes · size not listedLicence: apache-2.0
  • Edge0/Audio8-TTS-Realistic-AirReleased Oct 14, 2026 · 396 likes · 0.6B params (~1 GB in bf16, fits on 24 GB)Licence: apache-2.0
  • Oomni/mithi-v1-flashReleased Nov 7, 2026 · 64 likes · size not listedLicence: apache-2.0
  • empero-ai/PiVo-1.4MReleased Nov 1, 2026 · 58 likes · size not listedLicence: apache-2.0
  • gradium/phonon-31m-enReleased Oct 19, 2026 · 157 likes · 0.0B params (~0 GB in bf16, fits on 24 GB)Licence: apache-2.0
  • KittenML/kitten-tts-2Released Sep 30, 2026 · 219 likes · size not listedLicence: other (stellon-labs-community-license) (custom: read before use)

3D models

  • speridlabs/idolReleased Oct 28, 2026 · 1,950 likes · size not listedLicence: apache-2.0
  • hyz317/HairDeckReleased Oct 27, 2026 · 69 likes · size not listedLicence: apache-2.0
  • speridlabs/idol-turboReleased Nov 3, 2026 · 56 likes · size not listedLicence: apache-2.0
  • hi3d-official/Hi3Dv1.0Released Oct 16, 2026 · 163 likes · size not listedLicence: other (hi3d-open-license-agreement-1.0) (custom: read before use)

Video

  • Lightricks/LTX-2.5-DiffusersReleased Jul 26, 2026 · 139 likes · 19.0B params (~38 GB in bf16, tight on 24 GB)Licence: other (ltx-2.x-community-license-agreement) (custom: read before use)
  • Lightricks/LTX-2.5Released Jul 23, 2026 · 11,462 likes · size not listedLicence: other (ltx-2.x-community-license-agreement) (custom: read before use)

Not tested, and why

  • MiniMaxAI/MiniMax-H3-Image: Its licence does not allow any use in the United States, the European Union, the United Kingdom or South Korea, so we cannot run it.

Questions

How are the models tested?

Every model gets the same prompts and the same seeds, on the same computer: one NVIDIA RTX 3090 (24 GB). Speed is the typical time per item once the model is loaded; the first load is listed separately because it only happens once per session. Graphics memory is the most the model used.

Why these models?

We make the art, music, sounds and voices for our games on our own PC instead of paying for cloud AI. Each week a script reads the most-trending new open models on Hugging Face, leaves out copies of other models and anything that can't be used commercially, and puts the rest on the watch list below. The promising ones get the same side-by-side test, and we switch when one is better.

Can I use these models for my own projects?

Every model here can be downloaded and run at home, but read its licence first: some are free for any use, others are free only below a yearly revenue limit or ask you to credit them or label AI-made content. The licence column sums up each one in plain words; the model's own page has the full text.