I build AI stuff and see what breaks 🧪 ComfyUI • AI Video • Vibe Coding Benchmarks, experiments & tools. No hype.
Melbourne, Australia
- Tweets 3,749
- Following 661
- Followers 2,256
- Likes 2,812
Pinned Tweet
Okay this is actually insane.
MiniMax H3 can render CODE inside the video.
Like… actual readable SQL.
SELECT name, city
FROM customers
WHERE city = 'Melbourne';
And not just that.
It also rendered:
• table headers
• customer names
• city names
• IDs
• prices
• JOIN conditions
• big impact text
This is the part that blows my mind:
video models used to turn text into absolute alphabet soup.
Now H3 is casually generating SQL on screen ffs 😂
That changes the game for:
coding videos
tutorials
dashboards
UI demos
technical explainers
education content
MiniMax H3 might quietly have one of the best text-rendering capabilities in video right now.
It can literally generate code as part of the scene.
4
2
45
6,614
MiniMax H3 Extender now lets ComfyUI continue an existing video, and I like that its editor keeps reference frames right on the clip cards. github.com/tritant/ComfyUI_M…
9
ComfyUI-VideoMaskEditor flags flickering video masks and lets you copy a mask from the next frame, and I like that quick fix. github.com/knoic/ComfyUI-Vid…
1
187
A MiniMax H3 post-train just debuted at #2 on the Artificial Analysis text-to-video leaderboard, right behind Wan 3.0. I've been playing with H3 in ComfyUI, so seeing what a studio pulled out of it is wild. Open base models keep winning.
Utopai X (based on MiniMax H3) debuts at #2 on the Artificial Analysis Text to Video Leaderboard, behind only Wan 3.0
Utopai X is the new video model from Utopai Studios, post-trained on MiniMax H3 and launching today inside PAI, Utopai’s cinematic storytelling web platform. Utopai Studios is an AI-native film and TV studio, with three wide theatrical release films and two series scheduled for release in 2027.
In AA-Video-T2V v2.0, our refreshed Text to Video with Audio leaderboard, Utopai X ranks #2, behind Wan 3.0 and ahead of Dreamina Seedance 2.5. As of today, it is the highest ranked of the three models on the leaderboard built on MiniMax H3, ahead of MiniMax H3 itself and fal's MiniMax H3 Max.
Utopai X is available in PAI at 93 PAI credits per second of video, with no public API. PAI plans start at $15 per month (billed monthly) for 1,000 credits.
Congratulations to @UtopaiStudios on the release!
See below for our analysis and example outputs of Utopai X in the Artificial Analysis Video Arena 🧵
3
525
Everyone's talking about Griffin passing a "video Turing test." My real question: when does real-time AI video like this go open weights so we can run it locally in ComfyUI? Closed demo today, my GPU next year? 🤞 #AI
1
174
Castcut builds local ComfyUI character films through Look, Outfit, Day, and Cut, and I like having the whole workflow in one place.
github.com/doodersrage/castc…
1
177
Officially not a ghost channel anymore. 1,000 YouTube subs.
Thanks @vidIQ. Next up is X.
Still just building AI stuff and seeing what breaks.
1
1
4
254
Sign-up to VidIQ:
vidiq.com/princedoesai
67
ComfyUI-OldTimeRadio can turn a story prompt into a narrated MP4 episode with voices and captions, and the old radio vibe is pretty fun. github.com/jbrick2070/ComfyU…
1
3
286
AnyAngle Studio lets ComfyUI queue multiple camera views with Qwen Image 2.1, and I like how clearly it shows the guide-to-result flow. github.com/T8mars/Comfyui-Qw…
1
1
235
Bernini v2's ComfyUI node can switch between text-to-video and source-video editing based on connected inputs, and I'm curious how that planning stage feels in a real workflow. huggingface.co/rzgar/Bernini…
2
201
ComfyUI just changed completely with Comfy Agent 🤖✨
Instead of manually connecting nodes, you can now prompt an agent to build, modify, and run complete workflows automatically.
Here are 4 impressive tests shown in action:
1️⃣ Image Editing: Reposed a character while maintaining facial/clothing identity using just a stick figure reference [00:00:16].
2️⃣ Multi-View & Consistency: Built a 6-view character turnaround grid and placed the character in 4 distinct city scenes using Flux 2D dev [00:01:10].
3️⃣ Storyboard to Video: Converted a 3-shot short story into a full 1080p animatic with Qwen Image 2.1 & Wan 2.2 / MiniMax H3 [00:04:08].
4️⃣ Workflow Modding: Extended an existing Qwen Image 2.1 template by wiring its output straight into MiniMax H3 driven by a single shared prompt [00:08:10].
Watch full video: piped.video/pBzz_USK2ic
1
2
634
Got tired of babysitting ComfyUI, so I built Scene Runner.
It's not a ComfyUI replacement. It drives your local ComfyUI via its API: write every scene in one box, hit run, and it queues them back to back.
Follow + repost, I'll DM 50% off (first 50).
scenerunner.princeapps.app
1
222
HunyuanImage 3.0 Base has ComfyUI-ready W4A8 weights at 43.6 GiB, and I like seeing its results beside the Instruct models. huggingface.co/PedroMarinhoD…
140
Sonder Editor gives ComfyUI a timeline for scenes, prompts, guide frames, and render ranges, and I like seeing the edit beside the generated clip. github.com/SonderSaid/ComfyU…
1
6
577
The last 2 weeks in AI:
- Sep 21: Grok 4.7
- Sep 22: GPT-6 + Claude Opus 5.5
- Sep 28: Claude Sonnet 5.5
- Sep 29: GPT-6.1 Sol
Which one are you trying first? 👇
1
143
I just realized the only thing holding me back from ditching chatgpt/codex for claude is image gen.
1
558
ggufy converts diffusion model safetensors into GGUF quantizations, and I like having a small GUI for that job. github.com/qskousen/ggufy
4
335
People think ComfyUI is rocket science. It's just connecting boxes.
Load the brain (Checkpoint)
Demand a picture (Prompt)
Cook the noise (Sampler)
Make it visible (Decode)
The real cheat code? The entire workflow is secretly trapped inside the final image. Drag a PNG onto the screen, let the nodes rebuild themselves, and hit queue.
Drag. Drop. Done. @ComfyUI #comfyui 📦✨
2
343