Heba AI (@SubarcticRec) header
Heba AI (@SubarcticRec) profile picture

Generative AI specialist /Motion Graph Working on AI first projects, XR games, documentaries and films. ComfyUI, After Effects, Ableton, Claude Code.

Finland
Joined March 2016
Photo by Heba AI (@SubarcticRec)Photo by Heba AI (@SubarcticRec)Photo by Heba AI (@SubarcticRec)Photo by Heba AI (@SubarcticRec)Photo by Heba AI (@SubarcticRec)Photo by Heba AI (@SubarcticRec)Photo by Heba AI (@SubarcticRec)Photo by Heba AI (@SubarcticRec)Photo by Heba AI (@SubarcticRec)Photo by Heba AI (@SubarcticRec)
Gemini Neon plans - Fable 5.2 executes 😎
1
149
Heba AI retweeted
the last video game before we are the video game
210
522
57
3,999
117,773
EDM is so gringe, but beat is better than worse.
39
Pretty strange reaction to the fact he managed to do hes own custom VST plugin with AI. I'd say it's the opposite of "being screwed".
35
What would startup event be without lasers? #slush
49
Another view on this: The model has 10s memory so t2v can remember that what was out of frame when it comes back.
A video model that generates 10s silent clip in 14minutes with RTX-5090. Sounds like 2024. No, it not a "world" model. It generates video. This video here took 13min to generate with my 5090
47
A video model that generates 10s silent clip in 14minutes with RTX-5090. Sounds like 2024. No, it not a "world" model. It generates video. This video here took 13min to generate with my 5090
An open world model now runs on one 32 GB graphics card, at 14 minutes for a 10 second clip. RTX 5090, PAN 2 int8: 16.2 s per step, kijai's ComfyUI pull request lists 15.2 PAN 2 turns text, or text and an image, into up to 20 seconds of video. Their team writes "PAN 2 is a world model, not a video generator." I ran their own swimmer example first, with their settings. The other clips use prompts written with their prompt guide. The prompts are long structured descriptions. Their docs say that style works best. 241 frames, int8 attention: 16.2 s per step (kijai 15.2) 241 frames, dense attention: 27.6 s per step (kijai 26.8) With int8 attention that is about 7 percent slower than kijai's number. He names no resolution and I found no cause on this box. For the dense run I switched the int8 attention off with a small node of our own. RTX 5090, 10 s clip: 842 s RTX 5090, 20 s clip from an image: 2,427 s RTX 5090, steps per clip: 50 RTX 5090, VRAM peak: 31.4 GiB of 31.8 GiB RTX 5090, system memory peak at 10 s: 11.3 GiB RTX 5090, system memory peak at 20 s: 20.3 GiB Their release has a bf16 transformer of 62.7 GiB, which doesn't fit a 32 GB card in one piece. So this is the int8 version by Comfy-Org in ComfyUI, not their own script. Their guardrail isn't part of the ComfyUI path. The 20 s run from an image misses one of their flags, ComfyUI has no switch for it. PAN 2 makes no sound, so the video is silent. Setup: ComfyUI master at e9afbccc, pan2_int8_convrot by Comfy-Org (24.85 GiB), Qwen3.5-9B int8 text encoder, Wan 2.2 VAE Settings: 50 steps, CFG 2.5, 1248x704, 24 fps, from their docs, clips shown with seed 123 Method: 10 s clip is the median of 3 warm runs with seeds 124 to 126, request to finished MP4, the others are single runs Model: huggingface.co/IFM/PAN2 PAN 2 by the Institute of Foundation Models at MBZUAI @IFM_AI. ComfyUI support by kijai @kijaidesign, int8 weights by Comfy-Org @comfyui. It uses the Wan 2.2 VAE by Wan @Alibaba_Wan and Qwen3.5-9B by Qwen @Alibaba_Qwen as text encoder.
1
1
132
H3 Character Swap Cat vs Dog
49
Testing Wiggle H3 ReStyling using Fist of Iron kung-fu scene as source to test how fast motion works. BG stays too realistic but both characters keep style in both cuts.
52
We released this CD 30years ago. Today its value is 300-500 euros on second hand market. Damn that I didn't save extra copies :/
40
The Lightbulb Man short H3 automated by Showrunner system
1
1
361
9-10y kids who use WhatsApp (mandatory in school groups) have started to chat with Meta AI and it can get wild. Meta ai curses a lot and "tries to be cool" with zero understanding what kids at that age can understand.
1
51
Video REA copy with H3 in ComfyUI
60
H3 Cinema Camera prompting tests.
49
Testing how to get H3 model to follow shot sizes when using ref images. Ref image + wording makes a quite big difference on the results.
1
60
Twitter. Twitter. Twitter. Twitter. Twitter. Twitter. Twitter-Twitter-Twitter-Twitter....
Stop calling it Twitter. It's X.
5,396
45,522
2,740
241,254
5,539,730
Fuck Apple!
Paid $200 for Logic Pro? Apple Says You Won't Get This New Feature macrumors.com/2026/11/05/log…
57
We're giving away a Hogwarts Legacy custom wrapped NVIDIA GeForce RTX 5090 to celebrate its official release with DLSS 5 🪄 To enter: 🟢 Share this post 🟢 Comment #RTXON Terms and Conditions: nvidia.com/en-us/geforce/con…
84,401
71,876
7,698
66,847
6,066,317