There is a validation problem at the centre of this whole topic, and almost nobody writes about it.
A drawing that looks like line art and a drawing that works as line art are two different artifacts, and you cannot tell them apart by looking at either one. Open contours are invisible in black and white. They only announce themselves later, when somebody fills the drawing and the colour runs straight past the edge.
So this piece is organised backwards from the usual shape. The gate comes first, then the routes that feed it.
Eight tests, all on Tsubaki.3 inside PixAI.
The spec
Four properties decide whether a drawing is usable downstream. Judge an AI line art generator on these rather than on how the output looks.
Closed contours. Every shape sealed. This is the one you cannot verify visually.
No grey. Grey tones, washes and screentone mean the model rendered instead of drawing, and they degrade at every later stage.
A true white background. Off-white, noise and faint washes all cause trouble downstream.
Line weight carrying information. Thick outside, fine inside. Optional on simple subjects, decisive on dense ones.
The gate: fill it
Run this on any drawing before you commit to it. One generation, and it settles the only question the image itself won't answer.
Method: line art set as a base image, then a flat colour prompt. No edit tool.
Add flat colours inside this line art. Keep every line exactly where it is and
do not redraw anything. Fill each enclosed area with one solid colour: pale
yellow hat, white mesh veil, cream canvas jacket, brown leather gloves, grey
metal smoker, tan wooden hives, flat pale blue sky. No gradients, no shading,
no lighting, no texture. The black outlines stay visible on top.
Every colour stayed inside its lines and nothing bled between regions. The black outlines all survived in their original positions, and each enclosed area took one flat colour with no gradients or shading.
Colour floods correctly only when the contours are sealed, so a clean fill is structural proof. Weak or open lines would have leaked and marked their own position on the drawing for you.
Most anime lineart AI output passes inspection and fails this. Run the gate first; everything below is about feeding it something that survives.
Left the line art. Right the same drawing coloured
Route 1: prompt for it directly
The cleanest of the three routes, and the one to default to.
I picked a subject engineered to cause problems. A mesh veil is transparency, wooden hives are repeating structure, a metal smoker is a hard-edged object, and bees in flight are small repeated shapes.
Method: new generation, no reference, no LoRA.
Black and white anime line art, drawn with a technical pen on white paper. A
teenage girl beekeeper stands among tall wooden hives, wearing a wide-brimmed
hat with a fine mesh veil hanging over her face, a heavy canvas jacket with
the cuffs buckled, thick gloves, and a metal smoker in her right hand with
smoke curling from the spout. Her face is visible through the mesh. Bees in
flight around her. Even line weight, every contour closed, pure white
background, no grey tones, no screentone, no hatching, no solid black fills,
no shading, no colour.
Passed all 4 spec properties. No colour, no solid black fills, a pure white background with no wash or noise, and her face legible under the mesh rather than buried by it. The smoker rests in the right hand with smoke curling from the spout, the cuffs are buckled, and line weight holds even throughout.
Naming a physical tool and real paper carries weight in that prompt. It gives the model something to imitate that isn't a rendered illustration.
Test 1: the direct ask.
Edge case: subjects with no edges
Line art describes boundaries, and rain, steam and breath don't have any. This tests whether the model understands the format or only the subject matter.
Black and white anime line art, drawn with a technical pen on white paper. A
girl standing in heavy rain under a broken streetlight, water streaming off
her umbrella, steam rising from a grate beside her, her breath visible in the
cold. Draw the rain, the steam and the breath entirely with line, no grey
tones, no screentone, no black fills, no shading, pure white background, every
contour closed.
All 3 came back as closed outline shapes instead of grey. Breath and ground steam as self-contained cloud loops, rain as elongated dashes, water off the umbrella ribs as drawn sheets. Every one of those would take a flat fill.
Most models render vapour with soft gradients, so this one understood the format rather than the subject.
Test 8 rain, steam and breath, all drawn as line. I selected batch x4, selected these two, let’s go with the right one.
Route 2: strip a render afterwards
Generate the picture you want, then remove everything that isn't line.
I generated a fully rendered beekeeper first, then opened the edit box and typed "Convert to line art", with Tsubaki.3 selected as the edit model.
Structure retention is the strength. Pose, hive arrangement, buckle placement and smoker geometry all survived without shifting. Every colour and value came out, leaving black lines on white, and the veil mesh translated without collapsing into a black blob.
Line quality is the cost. The conversion behaves like a threshold filter, so thin hair strands and the mesh picked up pixelated edges the prompted version never had.
Route selection rule: prompt directly when drawing quality is the priority, convert a render when you already have an image whose structure you want preserved.
Route 3: convert a photo
The weakest of the three, and the failure is systematic rather than random.
I attached a photo of a man in a camel coat standing by a vintage car and asked for the same clean treatment.
Convert the photo in the reference image into clean black and white anime line
art. Keep the layout, the proportions and the position of everything in the
frame. Draw it with a technical pen, even line weight, every contour closed,
pure white background, no grey tones, no screentone, no black fills, no
shading, no colour.
Composition transfer worked well. Pose, phone grip, coat drape, belt buckle, wire-frame glasses, rings and the car behind him all landed, and the face stylised into anime while remaining the same person.
Then it ignored every negative instruction in the prompt. Cross-hatching spread across the coat, the lapels, the trousers, the belt and under the car body. Solid black fills appeared in the belt area. Line weight varied throughout.
The mechanism: it translated the photograph's light and shadow into hatch marks rather than discarding the tonal information. A photo carries shadow values, and the model converts them instead of dropping them. No amount of negative prompting removed that behaviour.
Route selection rule: treat this as a stylised sketch generator, never as a source of production line art.
Left the source photo, by GlassesShop from Pixabay. Right the line art conversion.
Dial 1: model choice
This moved results more than any wording change I made.
I ran the beekeeper prompt word for word on Haruka v2, an SDXL model, expecting it to win on the strength of how much line work stands behind that family.
It failed comprehensively. Heavy black fills on the hair, the collar, the inner neck and the straps. A background of screentone foliage, shading and a drawn panel border where white space belonged. A face mask over her lower face rather than a veil hanging from the hat brim.
The strangest miss: it parsed "technical pen" out of the medium description and drew a pen in her hand in place of the bee smoker, with smoke coming out of her mouth like an exhale.
Aesthetic bias overrode the negative constraints entirely. The model knows what a manga panel looks like, so it produced one, screentone included. The newest model won a comparison I would not have predicted. The anime models guide covers where these families diverge.
Left Tsubaki.3. Right Haruka v2, same prompt.
Dial 2: line weight
The model has line hierarchy available and will not apply it unless the prompt asks. I ran a dense mechanical subject twice, changing one instruction.
Black and white anime line art with perfectly even line weight throughout,
drawn with a 0.3mm technical pen. Every line the same thickness, outer
contours and interior detail identical in weight. A mechanic kneeling beside a
stripped motorcycle engine on a workbench, tools and loose parts scattered
around her. Pure white background, no fills, no shading.
Black and white anime line art with weighted lines, drawn with a brush pen.
Thick bold contours on the outer silhouette, medium lines on major forms, very
fine thin lines for interior detail like bolt heads, cable runs and fabric
folds. A mechanic kneeling beside a stripped motorcycle engine on a workbench,
tools and loose parts scattered around her. Pure white background, no fills,
no shading.
The even-weight run complied fully and produced the worse drawing. The engine's outer contour carries the same thin stroke as the bolt heads, the gear teeth and her hair strands, so the mechanical area flattens into a mesh you have to decode rather than read.
The weighted run resolved foreground from background immediately. Heavy brush contours on her silhouette, the bench and the major blocks, fine strokes inside the engine and the fabric folds. The scattered parts became legible because line weight was building depth.
Even weight is fine on a simple character. On anything dense, specify the brush pen version.
Left even line weight. Right weighted lines.
Output targets: three densities
Different jobs want different amounts of line, and the vocabulary matters more than people expect. I ran three, each on a subject that suits it.
1st: Rough anime construction sketch in pencil, loose searching lines,
overlapping strokes, unfinished. A boy leaping off a skate ramp mid-air, board
sideways under his feet, arms out for balance, hoodie flapping. Gesture and
proportion only, no clean outlines, no fills, white background.
2nd: Clean anime production line art, drawn with a technical pen on white
paper. An elderly barber standing beside his chair in an empty shop, apron on,
comb in his breast pocket, mirrors and bottles behind him. Single confident
outline on every form, even line weight, every contour closed, no hatching, no
screentone, no black fills, no shading, pure white background.
3rd: Inked black and white manga panel art. A woman in a long coat standing
alone on a station platform at night as a train rushes past behind her. Heavy
spot blacks in the shadows, dense cross-hatching on the coat folds, fine
parallel speed lines on the train. Dramatic high contrast, white background,
no grey tones, no screentone.
All 3 passed, and they are 3 different products.
The skate sketch carries visible graphite grain and searching strokes, though the face resolved more finished than a true construction draft would.
The barber is the deliverable of the three. Every form closes into a sealed cell, the comb teeth render as open outlines rather than filled black, and no part of it needed hatching to read.
The station platform bends the spec on purpose: spot blacks in the hair and ceiling, hatching on the coat folds, speed lines on the train. That output is inked art rather than line art, and the two get conflated constantly. Request the wrong one and you receive a drawing that cannot be coloured.
Left the source photo, by GlassesShop from Pixabay. Right the line art conversion. (2)
Routing table
| What you're making | What to ask for | What to check |
|---|---|---|
| Colouring base | Clean production line art, even weight, closed contours | Fill it and look for leaks |
| Manga page | Inked panel art, spot blacks, hatching | Contrast without grey tones |
| Character design pass | Clean line art, plain background | Whether the shapes read without colour |
| Dense scene or machinery | Weighted lines, brush pen | Foreground separating from background |
| Concept from a photo | Reference conversion | Expect hatching where the shadows were |
| Early gesture work | Rough construction sketch | Motion and proportion, not finish |
What the 8 runs establish
Route 1 produced the cleanest output, and it worked because Tsubaki.3 honours negative instructions that the SDXL model discarded.
Route 2 preserves structure at the cost of line quality, which makes it the right call only when you already have an image you want to keep.
Route 3 has a systematic failure that prompting cannot address, because shadow values in a photo become hatch marks by design.
The gate is what separated a usable drawing from a picture, and it's the step that gets skipped. Setting line art as a base image and requesting flat colour is the whole verification, and it costs one generation. The how to use PixAI guide covers where the model selector, base image and edit options live if you want to reproduce this.
Your next move
An anime line art generator hands you a drawing. Whether you're holding AI anime line art depends on a property the drawing does not display.
Generate one, then run the gate. Set it as a base image, request flat colours with no shading, and watch the edges. Clean fills mean sealed contours and something you can build on. Colour creeping past the lines means you have a black and white picture, and the leak tells you which part to redraw.
Start with the barber rather than the engine. A simple subject tells you whether the prompt is right; only a dense one tells you whether the line weight is.
Top comments (0)