Film Production Studio

Art Studio

Every picture the studio makes, from a 2D drawing to a 3D game asset, directed by one lane and drawn by open software on our own machines. Direction is written down, so the look holds across every session and every picture.

Where it stands

Each piece run by run: the strongest picture of each run and the director's verdict. Open a run to see every candidate, the numbers, what must change, and the exact words.

Lighthouse Workshop · painted-storybook

sent back run card-2: Directed beats raw outright: raw drew a man at a desk with trumpets in daylight, none of the piece. Directed (compiler 2) holds the empty room, low key, one lamp pool, sextant and lamp on the left third, one red accent. Best: s4 (composition), then s2.
sent back run card-3: Closed tower room, violet dusk windows, one spiral stair: notes 1, 2 and 4 of card-2 are fixed. Best: s4, then s1.
sent back run card-4: Three-quarter view, bench diagonal from the left foreground, low-key with one lamp pool, timber ceiling and hatch: card-3's notes 2, 3 and 4 are fixed. Best: s1, then s4.

How a picture is made

1. The bibleEach look is written once: its feel, medium, shapes, palette by role, light, composition, do and don't, and the exact style words. One look, one account.
2. The pieceWhat is in the picture, where the eye goes, the shot and the beat. Never a typed prompt.
3. CompilePiece and bible become the model's words, in a fixed order, the same every time.
4. DrawSeveral candidates on an open model, each saved with everything needed to make it again.
5. JudgeThe machine measures palette distance, value grouping, focal contrast and likeness to the touchstones.
6. DirectThe Art Director looks at every candidate whole, passes one or sends the piece back with what must change. The verdict goes in the bible's log for good; a pass can become a touchstone.

Looks

Painted Storybookstudio · gouache and watercolour on cold-press paper · warm, safe, curious; a picture a child would want to step into

Models

ModelJobLicenceUseRuns on
z-image-turbophotoreal and illustration, text in the picture; 8 stepsApache-2.0yescard, par1-cpu (slow: bf16 is emulated on par1's Zen 3)
flux2-klein-4btext to image and editing with references in one model; 4 stepsApache-2.0yescard, par1-cpu
hidream-o1photoreal and editing, the best MIT model that fits the L4 (arena 979)MITyescard
chroma1-hdpainted and illustrated styles, real negative prompts; base for our own style LoRAsApache-2.0yescard
kromossbroad-style fine-tune of MOSS-Image, 8 steps with its DMD2 LoRAApache-2.0yescard
illustrious-xl-2anime and cartoon, the deepest character knowledge; LoRA per character; lineart and pose controlCreativeML OpenRAIL-Myespar1-cpu, card
noobai-xlanime with the widest character and artist knowledgeFair AI Public Licence 1.0-SD (fine-tunes we train stay private, never served)yespar1-cpu, card
qwen-image-2512text and lettering in the picture, layouts, postersApache-2.0yescard (fp8 or nunchaku 4-bit)
qwen-image-edit-2511editing with up to three references: a character on model across poses, outfits and scenesApache-2.0yescard (nunchaku 4-bit)
qwen-image-layeredsplits a finished picture into editable RGBA layers: cut-out parts for the rigsApache-2.0yescard
glm-imageaccurate lettering with a glyph encoderMITyescard
flux-3-kleinthe highest-ranked open model we may use (arena 1041); editing with up to 10 referencesFLUX Kommunity Licence: free under US$5M revenuecapcard
krea-2-turboaesthetic illustration, ships watercolour and kids-drawing style LoRAsKrea 2 Community Licence: free under US$1M revenue, content filter requiredcapcard
birefnetcuts a subject from its background (sprites, cut-out parts, compositing)MITyespar1-cpu, card
seedvr2restoration and upscalingApache-2.0yescard
sam-3.1masks by text or clickSAM Licence (commercial allowed; no military or ITAR use)yescard
omnisvg-1.1text or picture to editable SVGApache-2.0yescard
trellis-2picture to textured PBR 3D mesh, game and film assetsMITyescard (L40S; 24 GB is its floor)
pixal3dpicture-faithful 3D mesh whose front matches the artMIT (not a Hunyuan licence)yescard
minimax-h3-imagethe top open model on both arenasMiniMax H3 Community Licence: excludes the US, EU, UK and Koreanever
qwen-image-2.1text to image, Sept 2026Qwen Research Licence (non-commercial)study
animathe best small anime modelnon-commercialstudy
pony-v7anime and cartoonexcludes professional video productionstudy
insightface-identity-adaptersPuLID, InstantID, InfiniteYou, IP-Adapter FaceIDInsightFace face models are non-commercialstudy
microsoft-lenstext to imageuses OpenAI gpt-oss-20b as its text encodernever

Tools

ToolWhatLicence
comfyuithe node runtime every open image workflow is written for, driven headless over HTTPGPL-3.0
diffusersPython pipelines for scripted jobsApache-2.0
imagemagickevery raster operationImageMagick
libvipsfast large-image processingLGPL-2.1
vtracercolour raster to vector SVGMIT
potraceline art to clean vector curvesGPL-2.0
blender3D: models, sets, light, Grease Pencil 2D, toon and line-art renderingGPL-2.0+
siglip2the judge's eye for words and likenessApache-2.0