Confessions of a researchaholic

September 9, 2022

img2img stable diffusion

Filed under: Imaginary,Real — liyiwei @ 5:32 pm
Tags: , ,

I experimented with stable diffusion over some of my own manual drawings and here are the results.

September 5, 2022

Natural authoring modalities

Filed under: Real — liyiwei @ 11:32 am
Tags: ,

The technology has finally reached the point for which content can start to be authored via natural language and gesture (including drawing) instead of unnatural WIMP (windows, icons, menus, pointer).
The state of art (e.g., DALL-E-2) can already be part of the content creation toolbox, and future advances may soon obsolete legacy tools.

September 4, 2022

A caterpillar sleeping inside an avocado bed

Filed under: Imaginary,Real — liyiwei @ 10:34 am
Tags: , ,

With the same subject title, I created multiple outcomes via manual drawing and DALL-E-2 with different styles hints.

digital art

natural illustration

oil painting by Rembrandt

manual painting with Fresco (took about an hour)

August 23, 2022

Dall-E prompt optimization

Filed under: Real — liyiwei @ 9:21 am
Tags: , ,

Recent language-based image generators can output amazing results given the right kind of input prompts (e.g., book and market), which can be tricky to produce and rely on experiences. This led me to predict the emergence of a cottage industry of “image prompt optimization” analogous to search engine optimization, and research of machine learning modules (e.g., another network) that can learn translating ordinary natural language inputs to language-generator prompts that can lead to desired outcomes.

August 3, 2022

SIGGRAPH Asia 2022 papers PC meetings

Filed under: Real — liyiwei @ 11:19 am

As a pseudo-miracle, I managed to wake up before 5 am for 3 straight days for the virtual SIGGRAPH Asia 2022 PC meetings (ending 10+ am), and set a personal record of reviewing 5 papers as an extra PC member, which, with the original 15 assignments, totaling 20, still less than like 22 I reviewed a few years back (but with fewer extra PC reviews which tend to have more time urgency).

Aaron Hertzmann has a recent post about expectation creep that is worth taking a look for future PC members.

May 20, 2022

Creative block

Filed under: Imaginary,Real — liyiwei @ 9:59 am
Tags: ,

In Concept Art: Break Artist’s Block with Emergent Design – L410, conceptual artist Sam Nielson talked about the common causes and remedies for creative blocks, which I found to be equally applicable to scientific research (or any other creative fields):


Not enough or the right kind of input
, such as visual inspirations for art or the current problems and progresses in research.
Not having anything to say
, which is related to the above about not having the right kind of input as well as not thinking enough about what one wants to do.
Fear of poor performance
which can cause actual poor performance.


Study everything and anything
, including books, papers, and artworks.
Live life a little (outside work)
, which is not just about work-live balance but also about enriching one’s creative horizon.
Give yourself permission to suck
to increase the chance not to suck.

I followed the suggested process (about using chunky oil brushes for iterative ideation) to create the visual impression I saw earlier about two white chairs inside woods.

May 13, 2022

How to encourage on-time review submissions?

Filed under: Real — liyiwei @ 6:30 pm

May 7, 2022

If you can’t draw it simply, you don’t see it well enough.

Filed under: Real — liyiwei @ 10:45 am
Tags: , ,

The traditional way to render a realistic image is from a physics perspective: provide a complete enough description of the scene and simulate light propagation and image formation with sufficient accuracy.
The end goal of physical realism can facilitate the understanding of the physical universe, at the expense of costly computation.
Thus, computer graphics research has been mainly about how to hack this entire process to achieve maximum realism with minimum effort.
A key ingredient for this hack is the (limit of) human perception, where both sensorial and neural passageways cap the amount of generated information beyond which further enhancements cannot be perceived (and efforts wasted).

This “rendering hack” has long preceded computer graphics (or machine computation) in visual art, which are traditionally based on manual efforts and thus impractical for brute force computation.
The end goal of artistic ideal can facilitate the understanding of perceptual mechanisms which in turn provide opportunities for subjective variations (such as different artistic styles and movements).

This is one main reason why I have been practicing drawing, sketching, painting, animation, design, and other creative forms. It is more of an exercise to see than to draw. (Besides, it is fun and diversifies my daily routines.)

“If you can’t explain it simply, you don’t understand it well enough.” – A quote attributed to Albert Einstein.

Recent trends in black-box machine learning further moves in the opposite direction, by making the visual computation even more expensive and opaque than traditional forward rendering in computer graphics.
I look forward to see if machine learning can shed light on the artistic and subjective aspects of image formation.

February 23, 2022

Lively academic thread

Filed under: Real — liyiwei @ 3:34 pm

I saw 11 replies to a Slack thread of a research project and thought “whoa what a lively academic research discussion” and then came in to find out it was all about gift card amount for user study participants.

Next Page »

Theme: Rubric. Get a free blog at