EP 083 · 99 MIN HIS AI OVERLAYS ARE REAL TRACKING DATA
20 Year Creative Vet Sold Golden Wolf to Doodles, Then Went All-In on AI
The short version
Episode 83 profiles Ingi Erlingsson, who built Golden Wolf, sold it to Doodles, and now runs the AI-systems studio SYSTMS. He walks through his ComfyUI pipeline for custom nodes, tracked-data overlays, and transition LoRAs, compares Wan 2.2, Krea 2, and GPT Image 2.5 for pixel quality and anatomy, and argues that giving an agent a goal rather than a task matters more than any single prompt.
Key takeaways
- 01
Learn one model deeply instead of chasing every release: Ingi calls Wan 2.2 the best model for pixel quality precisely because he knows its quirks cold.
- 02
ComfyUI overlays don't have to be decorative: Ingi's tracked-dot and bounding-box graphics are extracted directly from the video's own motion data.
- 03
You can build a working custom ComfyUI node in about ten minutes by just describing the effect you want to an LLM.
- 04
Train video LoRAs for movement, not style: a strong image model plus Wan already delivers the look, so the LoRA should teach transitions instead.
- 05
Give an agent a goal, not a task: chasing someone else's viral prompt just replicates their goal, never your own project.
- 06
Open source only survives long-term with a business model behind it, Ingi argues: give away a usable free version, then charge for scale.
- 07
ComfyUI attracts hobbyists precisely because it's hard, Ingi says: it's like building the most difficult Lego set, not something people want made easy.
Receipts · said on the record
“What I realized was that I'd spent my entire career thinking that brands were brands and agencies were out there looking for something cool. But but brands are looking for something that sells.”
“There's you know, 99% of people are just gonna like use it like a piece of paper. They're gonna write on it with like a pencil. And then there's someone that's gonna make origami.”
“The conversation became only about the tools. Whereas like the point is that like you can use all these tools together to make something really, really interesting.”
“The other reason why I work in the way that I do is that I just really like having control. … It wasn't mine, right? I wouldn't feel like it was mine because it wasn't my idea.”
“That's where LLMs are going, right? Like it's like you if you define your goal well enough, it will figure out how to get there.”
Chapters
- 0:00:00 Who is Ingi Erlingsson?
- 0:02:58 How much creative freedom do top brands give?
- 0:05:56 How do studios get hired for their style?
- 0:10:59 What did Doodles teach Ingi about brand work?
- 0:13:05 How do you turn attention into recognition?
- 0:18:13 How does Ingi build his signature AI style?
- 0:26:13 Why does AI make experimentation cheaper?
- 0:29:01 How do AI and After Effects work together?
- 0:30:20 Why did Ingi move into ComfyUI?
- 0:31:42 What hardware does Ingi use for ComfyUI?
- 0:34:34 What does ComfyUI let you control?
- 0:38:33 How does Ingi create tracked AI graphics?
- 0:42:05 How do you build custom ComfyUI nodes?
- 0:47:11 How does Ingi combine fine-tuned AI models?
- 0:51:09 How should you train a style LoRA?
- 0:55:13 What should a video LoRA learn?
- 0:58:50 How much creative control should you keep?
- 1:01:04 Do AI videos need longer generations?
- 1:07:28 How can open-source AI sustain itself?
- 1:11:31 What did Doodles and Web3 teach him about AI?
- 1:15:01 How does viral content connect to X payouts?
- 1:17:42 AI critics keep moving the goalposts
- 1:19:12 How do bots expose themselves in replies?
- 1:21:48 How does AI search surface a local business?
- 1:24:04 Why do comment-for-workflow posts spread?
- 1:25:42 What should young creatives learn now?
- 1:28:23 What should AI creatives get good at?
- 1:32:15 Do creatives still need to collect prompts?
- 1:32:54 How did Ingi end up working on DuckTales?
- 1:36:12 Who gets the Fast Hours sign-off?
Questions this episode answers
Who is Ingi Erlingsson?
Ingi Erlingsson is an Icelandic-born creative director who built the animation studio Golden Wolf, sold it to Doodles, became Doodles' Chief Content Officer, and now runs the AI-systems studio SYSTMS, working with brands including Nike, Adidas, Disney, Netflix, and Coca-Cola.
How does Ingi Erlingsson make his AI art look the way it does?
Ingi builds custom ComfyUI nodes that extract real tracking data, such as facial points and object positions, directly from a source video, then overlays that data as graphics; the final composite is assembled in After Effects.
What does Ingi Erlingsson think of Wan 2.2 as a video model?
Ingi calls Wan 2.2 the best video model for raw pixel quality, better than Seedance in his view, though he says it isn't great at following prompts, so he pairs it with a strong image model rather than relying on text alone.
What is Ingi Erlingsson's advice for working with AI agents?
Ingi says the most important lesson, which he traces to a piece of ChatGPT advice, is to give an agent a goal rather than a task; if the goal is defined well enough, the agent will find its own way there.
Why does Ingi Erlingsson think open-source AI models need a business plan?
Ingi argues open-sourcing a model only lasts if there's a financial plan behind it, comparing the ideal approach to old free Photoshop trial CDs: give away a usable light version, then charge once people want to scale up.
Show notes
After 20 years in creative, building Golden Wolf and selling it to Doodles, Ingi Erlingsson went deep on AI to start a new chapter at his company, Systms.
Ingi has spent two decades making work for brands including Nike, Adidas, Jordan, Disney, Netflix, Coca-Cola, and Cartoon Network. He built Golden Wolf from the ground up, sold the studio to Doodles, became Doodles’ Chief Content Officer, and now builds custom AI creative systems through Systems.
So naturally, Drew and Rory spend half the episode staring at his ComfyUI graphs wondering what the hell is happening.
Ingi breaks down how he actually works with AI: custom ComfyUI nodes, LoRAs, agents, model fine-tuning, Wan 2.2, Krea 2, GPT 2.5, After Effects, transition models, and the reusable systems behind his visual work. The conversation also gets into creative control, originality, attention, open source, prompt hunting, and why experienced creatives may have an advantage as AI gets more capable.
More importantly, Ingi explains why the goal matters more than the prompt, why rapid iteration changes the creative process, and why giving AI more capability does not automatically mean giving it more creative decisions.
