I’m trying to create a basic Manim animation that shows the autoregressive generation process for how language models produce text by predicting tokens from left to right.
Currently, I’m using a VGroup which grows as each new token is “predicted”.
However, the baseline alignment of the tokens appears “ragged”.
How can I ensure that the tokens are positioned correctly, so that the text is aligned?
Below is my minimal example demonstrating the problem:
from manim import *
class textConcat(Scene):
def construct(self):
# List of tokens to predict
tokens = "The cat is hungry".split(" ")
# Initial token setup
text = Text(tokens.pop(0), font="Helvetica", font_size=20).set_color(BLUE)
text_group = (
VGroup(text).arrange(RIGHT, aligned_edge=DOWN).to_corner(UL, buff=1.0)
)
self.play(Write(text_group))
for token in tokens:
# machine outputs the next token to the right
prediction = Text(token, font="Helvetica", font_size=20).set_color(BLUE)
# Instantiate the prediction behind the machine
self.play(FadeIn(prediction), run_time=0.01)
# # Slide the prediction to the right of the machine
# self.play(prediction.animate.next_to(machine, RIGHT * 4), run_time=0.5)
# Create an arc path from the right of the machine to the left
start_point = prediction.get_left()
end_point = text_group.get_right() + RIGHT * 0.3
arc_path = ArcBetweenPoints(
start_point, end_point, angle=-PI
) # Negative for an arc underneath
# Animate the prediction following the arc path
self.play(MoveAlongPath(prediction, arc_path), run_time=1.0)
# Update the VGroup with the new prediction token # TODO: Fix alignment of baseline
text_group.add(prediction)
self.play(
prediction.animate.next_to(text_group[-2], RIGHT),
run_time=0.5,
)
# Finish animation
self.wait(2)
if __name__ == "__main__":
from manim import *
textConcat().render()
And the result:
I’ve tried capturing the y coordinate of the initial token and forcing the alignment accordingly, e.g.:
text_group.shift(UP * (text_group.get_y() - text_group[0].get_y()))
But this hasn’t helped.
Any pointers would be much appreciated!
Note, I’m using ManimCommunity v0.18.1.
In manim text is really difficult to align because manim has no fixed height for text of a specific size. Instead, the height of the underlying SVG image is used (I assume). So in your case “The” has a different height and center than “hungry” because of “g” and “y” which are lower.
The only solution I can think of is to use a Tex() object and create a copy of your text (in variable token_copy) at the latter position
from manim import *
class textConcat(Scene):
def construct(self):
# List of tokens to predict
tokens = "The cat is hungry".split(" ")
# Initial token setup
token_copy = Tex(*tokens, arg_separator = ' ', font_size=30).set_color(RED).to_corner(UL, buff=1.0)
text = Tex(tokens.pop(0), font_size=30).set_color(BLUE)
text_group = (
text.move_to(token_copy[0].get_center())
)
self.play(Write(text_group))
for i in range(1,4):
# machine outputs the next token to the right
prediction = Tex(tokens[i-1], font_size=30).set_color(BLUE)
# Instantiate the prediction behind the machine
self.play(FadeIn(prediction), run_time=0.01)
# # Slide the prediction to the right of the machine
# self.play(prediction.animate.next_to(machine, RIGHT * 4), run_time=0.5)
# Create an arc path from the right of the machine to the left
start_point = prediction.get_left()
end_point = token_copy[i].get_center() #text_group.get_right() + RIGHT * 0.3
arc_path = ArcBetweenPoints(
start_point, end_point, angle=-PI
) # Negative for an arc underneath
# Animate the prediction following the arc path
self.play(MoveAlongPath(prediction, arc_path), run_time=1.0)
# Update the VGroup with the new prediction token # TODO: Fix alignment of baseline
text_group.add(prediction)
self.wait(0.5)
# Finish animation
self.wait(2)
The code (for example, the for-loop) could obv. be more optimized. But I just wanted to show the general idea.