An illustrator moves one drawing through stages. Line art first, then values, then flat color, then the finished render. Each stage exists so you can settle one thing before the next one covers it up.
AI image tools normally skip all of that and hand you the last step. I wanted to see whether Tsubaki.3 can work inside the stages instead of jumping past them.
So I built one scene and moved it through the whole sequence, feeding each result into the next. Then I ran four more tests on the questions a single chain leaves open. The frame I judged everything against is simple: each stage has one job, and the test is whether it does that job and stays in its lane.
What an AI art workflow has to protect
A workflow is different from four pictures that happen to look alike. The character, the pose, the proportions, the camera, the composition, and the background should stay where they are, and the only thing that changes from stage to stage is how finished the image looks.
The scene I used is a young man kneeling beside a half-stripped motorbike in an open garage at dusk. Bleached undercut, oil-stained tan coveralls tied at the waist over a black tank top, a red star patch on his left shoulder strap, a wrench in his right hand, the number 07 on the bike panel, a tabby cat asleep on a red toolbox behind him, a pendant lamp overhead, and the garage door open to the street.
Those are the nine details I followed through every stage, so if one shifts or disappears I know where it happened. Each stage gets a pass, a partial pass, or a fail, based on whether it added the rendering I asked for and left the nine alone.
Stage 1: the line-art base
This is the structural starting point, and everything later in the AI drawing workflow refers back to it. Its one job is to draw the whole composition in clean line, with no color and no fills.
Clean black and white line art of an anime scene, no shading, no grey tones, no
color, white background wherever there is no linework. A lean young man with a
bleached undercut kneels beside a half-stripped motorbike inside an open garage
at dusk. He wears oil-stained tan coveralls tied at the waist over a black tank
top, with a small five-pointed star patch on the left shoulder strap. He holds a
wrench in his right hand and rests his left hand on the bike seat. The motorbike
faces left with the number 07 on its side panel. A tabby cat sleeps on a red
toolbox behind him. A single pendant lamp hangs above the bike. The open garage
door shows the street outside. Confident even line weight, the full composition
drawn, no cross-hatching, no fills.
All nine anchors came through, and the line-art instruction did not. The composition is right: the man kneeling in the correct pose with the wrench in his right hand and his left on the seat, the bike facing left, the cat asleep on the red toolbox behind him, the lamp overhead, the open door framing the street.
What I asked for was black and white with no shading and no fills. What I got had tan coveralls, a red toolbox, an orange dusk gradient in the sky, and heavy solid black fills.
Two anchors also landed in the wrong place. The star patch went onto the folded coverall waist strap instead of the shoulder strap, and the 07 went onto the front tank panel instead of the rear side panel. Both are close enough to pass as the same design, and both matter later.
So the composition passed and the stage instruction failed.
Stage 2: line art to grayscale
The Stage 1 image goes in, and only value should come out of it.
Add grayscale shading to this line art. Keep every line, the pose, the character
design, the number 07 on the bike panel, the star patch on the left shoulder
strap, the wrench in his right hand, the tabby cat on the red toolbox, the
pendant lamp and the camera framing exactly as they are. Add only value: soft
grey form shadows, cast shadows on the floor thrown by the pendant lamp above,
and darker values in the garage corners. No color anywhere, black white and grey
only. Do not redraw or restyle the linework.
This stage corrected the previous one. The color that leaked into Stage 1 was stripped out completely, leaving a strict black, white, and grey image. It added value without redrawing, so the linework, the pose, and the framing came through untouched.
Every anchor held: the man, the wrench, the 07 panel, the star patch, the cat, the toolbox, and the lamp all stayed where Stage 1 put them. This stage passed on every count, and it turned out to be the only transition in the whole chain that cleaned up an earlier mistake rather than inheriting it.
Left: Stage 1 line art. Right: Stage 2 grayscale
One practical note. There are two ways to feed one image into the next stage. You can click the image, choose edit, and paste the prompt, or you can click the image, set it as the base image, and prompt from there. I used the edit method here, and the difference between the two shows up in the next stage.
Stage 3: flat colors
This is the AI line art coloring step, and nothing but color should arrive. I ran it twice, once by editing the Stage 2 image and once by setting it as a base image reference.
Add flat colors to this grayscale image. Keep the pose, the linework, the
framing, the number 07, the star patch, the wrench in his right hand, the
sleeping tabby cat, the red toolbox and the pendant lamp exactly as they are. Use
flat blocks of color with no gradients, no rendering and no lighting: bleached
blond hair, tan coveralls, black tank top, red star patch, deep blue motorbike
with a white 07 panel, red toolbox, brown tabby cat, grey concrete floor, warm
orange sky through the open door. Solid fills only, outlines still visible.
The edit run applied flat fills to the character and the bike, and then kept rendering anyway. The volumetric light cone under the pendant lamp stayed, and the sky came out as a smooth gradient rather than a block. Flats are not supposed to carry light.
The palette was mixed on this run. Tan coveralls, black tank top, brown cat, and red toolbox all matched, but the panel behind the 07 came out blue instead of white, and the single star patch became two red stars on the hip flap. Structurally it was solid, and the linework from Stage 2 held almost perfectly, with no redrawing of geometry, pose, or background, and all nine anchors in position.
The reference run did better. It removed the sky gradient and replaced it with a solid orange block, and it corrected the 07 panel to white. The light cone under the lamp survived that run too, and the star stayed on the hip flap.
So it is a partial pass either way, with the reference method ahead of the edit method on color discipline. That difference is the reason I lean on the reference method whenever a stage has strict rules.
Left: flats using edit. Right: flats using the image as a base reference.
Stage 4: the final render
The last step in this AI coloring workflow adds lighting, texture, and finish, with the design locked.
Render this flat-colored image into a finished illustration. Keep the pose, the
composition, the camera framing, the character design, the number 07 on the bike
panel, the star patch on the left shoulder strap, the wrench in his right hand,
the sleeping tabby cat on the red toolbox and the pendant lamp in the same
positions. Add lighting, soft shadows, material texture on the metal, the fabric
and the concrete, warm light from the pendant lamp and cooler dusk light from the
open door, and finished detail throughout. Do not change the design or move
anything.
The dual lighting worked properly. Warm light falls across the garage frame, the character, and the toolbox, and cooler dusk light comes in from outside. Two temperatures, both coming from the sources I named.
The material work is strong. Soft shading, ambient occlusion, and highlights across fabric folds, skin, bike metal, and the garage floor, none of it destroying the line definition underneath. Structure held from the earlier stages with no drift, so the pose, the wrench, the bike geometry, and the camera framing all stayed locked.
Two things moved on the way through. The star patch is still on the hip flap, inherited from Stage 3, so this render keeps eight of nine anchors. And the sky drifted from the flat orange I set in Stage 3 to deep twilight blue, overriding a color decision the previous stage had already made. The stage passed, with that background color drift noted.
Right: Stage 3 reference. Left: The final render.
The payoff test: does the middle matter?
Four stages is more generations than one, so I went back to the Stage 1 line art and took two shortcuts from it directly, to see whether the middle steps were protecting anything.
The first shortcut went straight to flats, using the same flat-color prompt pointed at the line art. It followed the no-lighting rule better than the chain did, with no light cone and no sky gradient, just pure unrendered fills. The chained version had inherited the lighting structure from Stage 2 and carried it forward, which is what leaked into the Stage 3 flats.
The cost showed up in the background. It erased the exterior houses, the street layout, and the utility poles that Stage 1 had drawn. The outlines, framing, and base geometry inside the garage stayed locked, and all nine anchors are legible, with the same blue 07 panel and the same hip-flap star.
Stage 1 line art left, the straight-to-flats shortcut right
The second shortcut went straight to a final render, using the Stage 4 prompt pointed at the line art. The lighting here is good on its own terms, with warm lamp light blending into cool exterior light, and everything else went backwards from there.
It painted over the line art instead of preserving it, softening the outlines into a digital painting. The exterior houses, street, and poles were replaced with a plain dark gradient. The coveralls shifted from tan to olive drab, the star turned yellow, and the bike went light grey instead of deep blue. It also added an unprompted 007 graphic on the lower frame rail above the rear wheel.
So the middle stages are doing real work. The chained path keeps the background architecture, the color decisions, and the geometry, while the direct path renders light well and loses the drawing underneath it. The shortcut passed on lighting and failed on everything the workflow exists to protect.
The straight-to-flats shortcut left, the straight-to-render shortcut right
The nine anchors across four stages
Laid out side by side, the pattern is clear.
| Anchor | Stage 1 | Stage 2 | Stage 3 | Stage 4 |
|---|---|---|---|---|
| Character design | Correct | Held | Held | Held |
| Pose and wrench | Correct | Held | Held | Held |
| Coveralls tied at waist | Correct | Held | Held | Held |
| Star patch | On the waist strap, not the shoulder | Held | Doubled in the edit run | Still on the waist strap |
| Number 07 | On the front tank panel | Held | Blue in edit, white in reference | Held |
| Cat and red toolbox | Correct | Held | Held | Held |
| Pendant lamp | Correct | Held | Held | Held |
| Bike facing left | Correct | Held | Held | Held |
| Camera framing | Correct | Held | Held | Held |
| Stage rule followed | Failed, color and fills | Passed | Partial, light cone stayed | Passed, sky drifted |
Three things stand out. Geometry never drifted through the chain, so across four generations the pose, the framing, and the placement of every object held, which is the part I expected to give way first. Errors are inherited rather than corrected, so the star patch went to the wrong place in Stage 1 and stayed wrong through all four stages, because nothing downstream questions a decision made upstream. And Stage 2 was the only transition that cleaned something up.
Can color sit on top of value?
Painters build value first and color second, and the usual failure is that color arrives and flattens the shading underneath. I used a new scene here, a weathered fisherman hauling a net over a small boat at dawn with a seagull on the prow, and ran it twice.
The value base came out right. Strict monochrome, and the three value zones landed in the order I asked for, with the fisherman and the net darkest, the open water mid grey, and the low dawn sky lightest. The seagull is on the prow with a soft cast shadow under its feet, painted in value blocks rather than outlines.
Then I added color over that value with an instruction not to relight or add highlights. The palette and the geometry both landed: cold blue-green water, pale peach sky, mustard cap, weathered brown wood, off-white gull, with the pose, the face, the net weave, and the cast shadow matching the base run almost pixel for pixel.
The mustard cap is where it came apart. In the value pass that cap was a near-black charcoal tone deep in the dark zone, and adding yellow lifted it into a bright mid-tone. It applied the color as a new local value rather than laying the hue over the value that was already there. The three big zones kept their order, so the overall structure survived. A partial pass: correct colors, correct drawing, one value zone rewritten by its own hue.
Left the value pass. Right color applied over it.
Lighting as its own stage
Relighting is easy to fake when the light source is off-frame. Here the source is in the picture and the wet ground reflects it, so there are two places to check every change. The scene is a boy in a green raincoat feeding a stray dog outside a closed noodle stall, lit only by a pink neon bowl sign, with the wet ground reflecting the neon.
Test 7 I’ll go with the left one.
The base render put everything in place, with one partial fail on the way in: a cyan ambient glow at the back of the alley, in a prompt that said the neon was the only light source. I then ran three relights from a shared template.
The first turned the sign off and put a full moon overhead. The pink glow switched off completely, a moon drove a cold blue palette through the scene, every pink highlight on the wet ground was replaced with moonlight, and the shadows project forward, away from the moon. The pose and the alley details are identical to the base. A full pass, including the puddles.
The second kept the sign on and added a car headlight from the far left. The car went down the center of the alley instead of entering from the left, and the headlights made it drop the global exposure, so the boy and the dog are buried in shadow with a lit neon sign right above them. High beams behind two subjects should rim their edges, and there is no edge light and no forward cast shadows. It inserted the car and then treated the result as a darkened composite rather than a scene with two lights in it.
The third turned the sign off and put a flashlight in the boy's hand, and I fed it the image from the second relight rather than the base. The flashlight beam passes through the dog's legs instead of stopping at them, so the body looks semi-transparent and throws no shadow, and the car and its high beams carried over from the input. A fail on light physics, treated as an overlay rather than as light meeting an object.
Relight 1 (left), Relight 2 (middle), Relight 3 (right)
Does it correct a drawing you meant to be strange?
Cleanup and rendering are where a model can quietly improve your work into something you did not draw. So I gave it a picture with a deliberate oddity: a woman holding an open umbrella upside down like a bowl to hold the rain, with a small heron standing inside the upturned umbrella, in a flooded street.
The line-art stage kept the umbrella upside down with the bird inside it. The common version of this image is a woman holding an umbrella the normal way, and it did not reach for that. The flood water, the half-submerged bicycle, and the shuttered shopfronts are all in place. The line-art rule slipped again, the same way it did in Stage 1, with solid black fills in the hair, the bird, and the coat interior. Intent passed, formatting partly failed.
The render held the premise all the way through. The umbrella stays inverted and catching water, and the dark bird silhouette from the line art resolved into a grey heron with a yellow bill, which is the model taking the word heron from the prompt and applying it to a shape it had already drawn. The bicycle and the shopfronts stayed anchored, the rain falls in vertical streaks with ripples on the surface, water drips off the umbrella ribs, and the flooded street reflects it all. Both stages passed, and this test tells me more about real production use than the render quality does.
Left the line-art stage. Right the finished render.
Where this fits in real work
As an AI art workflow, the chain is good for exploring rather than locking things down. Geometry survives every stage, so you can see one drawing as value, as flats, and as a finished render without redrawing it. That covers most of what an AI illustration workflow is for: testing a rendering direction, comparing color treatments, or taking a rough idea far enough to show someone.
Two habits made this go better. Get the line art right before you move on, since errors are inherited and never questioned later. And use the base image reference method rather than edit when a stage has strict rules, which is what removed the sky gradient and corrected the 07 panel in Stage 3. The Edit Pro guide and the Reference Pro guide cover both.
I would be careful with anything that needs the stage rules obeyed exactly, since clean line art never arrived, flats came back with lighting in them, and color changed the value under it. This AI art process gives you a reference version of those stages rather than a production one. Naming your anchors in every prompt does most of the work, and the prompt formula guide and the advanced prompting page cover how to write them.
The next step
Tsubaki.3 supports an AI anime art workflow rather than replacing one. Across four chained stages the drawing never moved, and shortcuts from the line art cost me the background, the palette, and the outlines.
The failures were about stage discipline. Line art came back with color and fills twice, flats came back with a light cone, and yellow lifted a dark value. It also kept a deliberately illogical umbrella through two stages without correcting it, which matters more for line art to finished art AI work than the rendering does.
Take a drawing you have already made, name every element in it, and run it through two stages on PixAI.
Top comments (0)