Show Notes
AI agents can build the workflow, critique the result, and keep iterating, but when the service goes down, suddenly reaching for the mouse feels strangely primitive.
This week on Fix It In Prompt, Alex Villabon and Doug Hogan talk about becoming dependent on agentic tools, keeping strong local models as production insurance, and finding the human checkpoints that should stay in the loop as AI moves deeper into VFX.
We get into Qwen3.8-Flash-Next, a locally generated MiniMax H3 hotel-lobby meme, Seedance 2.5 Draft Mode, Seed Audio 1.5, Qwen-Image 2.1, LTX-2.5, Claude Opus 5.5, Jev’s fast decision model, pop-up VFX teams, Comfy Agent, Comfy MCP, Comfy Router, and the slightly uncomfortable realization that sometimes the artist is the bottleneck.
💬 Where do you draw the line between letting an agent run and staying close enough to understand and improve the work?
📬 CONTACT
Email us: [email protected]
Find Doug: ComfyUI & VFX website: https://www.doughogan.com ComfyUI Workflows: https://comfyui.sjv.io/OYydOG LinkedIn: https://www.linkedin.com/in/doughogan
Find Alex: VFX + AI newsletter: https://alexvillabon.substack.com/ LinkedIn: https://www.linkedin.com/in/avillabon/
🔗 LINKS MENTIONED IN THE EPISODE
Qwen3.8-Flash-Next — the local mixture-of-experts model Doug recommends: https://huggingface.co/Qwen/Qwen3.8-Flash-Next
MiniMax H3 commercial licensing for local production work: https://blog.comfy.org/p/comfy-is-now-the-only-official-reseller
Qwen-Image 2.1 in ComfyUI: https://blog.comfy.org/p/qwen-image-21-in-comfyui-open-weight
Seedance 2.5 — official model overview: https://seed.bytedance.com/en/blog/one-take-creation-flexible-referencing-introducing-seedance-2-5
Claude Opus 5.5: https://www.anthropic.com/claude-opus-5-5
Dario Amodei — “We Must Pace the Frontier”: https://darioamodei.com/post/we-must-pace-the-frontier
TypeSafe — Introducing System One Models and Jev: https://typesafe.ai/blog/introducing-system-one-models-and-jev
Ostris AI Toolkit — open-source LoRA and model fine-tuning: https://github.com/ostris/ai-toolkit
Comfy MCP — local agent control for ComfyUI: https://blog.comfy.org/p/open-sourcing-comfy-mcp-on-local
Comfy Router — one API for frontier media models: https://blog.comfy.org/p/introducing-comfy-router-one-api
⏱️ TIMECODES
00:00 When the agent goes down and the mouse feels strange 02:50 Local LLMs as production insurance 05:53 Rebuilding the hotel-lobby meme with MiniMax H3 10:25 Learning prompt structure from agent feedback loops 13:00 ByteDance, Seedance 2.5 Draft Mode, and Seed Audio 1.5 16:20 Pop-up VFX teams and the changing studio model 21:39 Astra, Opus 5.5, and agents unlocking old software 26:35 Pacing the frontier and protecting human checkpoints 35:35 Open source, trust, and who benefits from slowing down 38:44 Jev: fast decisions instead of generated text 43:43 Learning through play and pushing past model assumptions 47:42 Full access, autocomplete habits, and AI hygiene 50:39 Junior artists, agent dependence, and the future of work 59:29 Reverse-engineering agents, Comfy Agent, and Comfy Router
#DellProPrecision #DellProMax #DellTech #NVIDIA #AI #VFX #ComfyUI #AIAgents #OpenSourceAI #MiniMaxH3 #Seedance #Jev #FutureOfWork
More from Fix It In Prompt
AI VFX hardware is expensive enough without buying a workstation before you know what you want to make.
This week on Fix It In Prompt, Alex Villabon and Doug Hogan hit double digits with LTX VFX Week, Alpha Gen mattes, ACES/HDR workflows, and their latest MiniMax character-replacement and acting experiments. We talk about using agents to test, review, and refine those workflows—and why a quick test can miss what a model is capable of.
AI VFX is getting more capable by the week, but Victor Perez makes the case that the real job is still curiosity, taste, and knowing what decision to make next.
GPT-6 Astra is not a one-click VFX solution, but it may be the first tool that makes a lot of specialists ask what else they could build.
This week on Fix It In Prompt, Alex Villabon and Doug Hogan are joined by LTX VP Applied Research Ofir Bibi and Technical Director Mohamed Oumoumad to unpack LTX-2.5: how better data and Gemma 4 improve prompt understanding, why older LoRAs can survive model upgrades, and how Diffusion Fidelity Rendering spends compute where the final pixels actually matter.