Ichel Park (@ichelpark) header
Ichel Park (@ichelpark) profile picture

🎥 film director & k-pop music videos. building shoty, a cursor for video. a new world for video gen. la / seoul

Los Angeles, CA
Joined September 2024
Photo by Ichel Park (@ichelpark)Photo by Ichel Park (@ichelpark)Photo by Ichel Park (@ichelpark)Photo by Ichel Park (@ichelpark)Photo by Ichel Park (@ichelpark)Photo by Ichel Park (@ichelpark)Photo by Ichel Park (@ichelpark)Photo by Ichel Park (@ichelpark)Photo by Ichel Park (@ichelpark)Photo by Ichel Park (@ichelpark)
we never built a set. we never left the green screen. no location scout, no vfx team, no roundtrip to a comp artist. green screen in, finished shot out, right inside your timeline. this is shoty. let's make it real.
13
3
2
108
3,952
Our YC application video. Hollywood’s best are already using Shoty. We’re DoubleZero. We’re building a beast of an AI video model—not another API wrapper. @garrytan @ycombinator, good luck saying no.
12
2
1
27
807
Ichel Park retweeted
Today, we’re launching Clone. An agent harness that predicts your next prompt. - Clone Mode keeps prompting your AI agents and never stops till done - 93.9% prompt prediction accuracy, outperforming Astra and Opus - You can use your own ChatGPT or Claude subscription Want to push agent-maxxing to the limit? clone.is/you
20
28
16
229
72,282
one bad mask frame can wreck the whole video. find the dropout, borrow a neighbor mask, then rerun. steps below.
1
1
8
394
flag frames where the mask vanishes or suddenly shrinks. copy the previous or next frame’s mask, then paint only the missing edge.
1
10
stop asking one reference image to control everything. a comfyui h3 workflow separates motion context, multiple references, and keyframes. here’s a cleaner way to assign each job.
3
8
348
keep those inputs distinct, then test one change at a time. otherwise it’s hard to tell whether the motion, reference, or prompt caused a drift.
1
22
h3’s new node is for video-to-video, not magic first-pass generation. use temporal windows and spatial tiles to upscale without asking one pass to hold the whole frame. steps below.
2
5
254
tile the frame and window the timeline so each model call handles a smaller region. the node author says not to use this setup for fully generated first-stage video; it is aimed at video-to-video and upscaling.
1
1
10
stop treating motion transfer like a one-click effect. a clean motion source is a camera plan in disguise; here’s the open workflow.
1
6
292
the practical advantage is control: choose footage with a readable silhouette, unobstructed limbs, and steady lighting before swapping. cleaner inputs give the model less ambiguity to resolve.
1
14
long clips break at the handoff, not the prompt. split by action, then carry the last clean frame forward. steps below.
1
1
11
346
2. export the final clean frame of beat one. use it as the image reference for beat two; describe only the new action, not the whole scene again.
1
9
3. if the frame drifts, regenerate that beat instead of extending the whole sequence. this isolates continuity failures and limits wasted generations.
6