VIBE CODE

DG_directorview

In a 20 year career in Animation and VFX I've seen a lot of production tracking platforms. For a time I felt that platforms like FTrack, Nim, Kitsu and Shotgun brought great consistencies that proprietary systems lacked, but my concern with SAAS offerings is that users become financial fodder and that only bespoke solutions properly develop. So I've started making my own.

directorview (director view / direct or view) is primarily a transpose of my google sheets prompter, but expanded in the intent to do production tracking. It's not finished and it's not dependable circa May 2026, but it's plodding along. Today I invite you in to take a look.

DG_Showrunner

A tool to drive ComfyUI inference from a database. Originally I used a Google sheet of prompts connected by a custom node for GSheets integration. Then AI built my own nodes that detect character locations, lighting setups, and feed them into Comfy to load LoRAs, set frame rates, drive prompts and dictate models (feature currently on hold due to disparity between workflows between models).

The DG_Showrunner suite will also generate slates / overlays and used publisher nodes to register as versions on Shotgrid. If you're reading this it means I've stopped building this website and am back to working on publishing to directorview.

DG_CollectDataset

A tool to handle the myriad of steps from "subject for my LoRA" through to "load my LoRA in ComfyUI". DG collect dataset begins as a beautiful soup call that gathers X (100) images, crops for face and body, validates the set is consistently one person, fails to remove watermarks (I tried!), listens to a "this is a dataset of [projectname]" captions using whichever of the 45 frameworks I happen to use on any given day, downresses to 256x256, makes a handy contact sheet and publishes to musubi tuner with bats, tomls and a template training script.

It's been the script I've battled with hardest, but pitted against the time to do manually, it's the one that's saved me most time. Then LTX2.0 added audio and it went all in the bin.

AATAS

AATAS was conceived and developed back in the flux 1 age, when I was binge watching Everything Everywhere all at Once. A Michelle Yeoh montage played out in my head as smarter people than me explained that in the T5 age, it's not about consistency between images but variety. Multiple locations, multiple looks.

"Anything Anywhere Together All Simultaneously" is the result. A Maya Python script that invites the user to set 3 spheres to identify character, world, and scale, and generates a camera volume around it (I like photogrammetry, splats, lidar, and deep exrs). Users can then set multiple locations that guide a one-frame batch render to generate data training sets for LoRA training.

In writing this now, I wonder if IC LoRAs might do a better job of translating the precision of camera positions so they can also inform training. Training LoRAs, like CG modelling, seems to be my happy place—producing characters to imbue with lives of their own.

AATAS was programmed as a god file using Gemini 2 on the web portal, before being refactored into modules for each component step. When I return to 3D it will be interesting to see what current models can suggest for code health, or what an MCP server can bring to the mix in terms of ease of integration.

DG_VideoScraper

A video scraper branched from DG_CollectDataset. The simplest of my scripts. You put a weblink to a video online and it downloads it at max res.

At its simplest the user needs to save a text file and drop it on a bat. As I learn more about rules, agents, workflows and the like I've made my own /DG_scrape_video implementations where I can set a link and local filepath and trust Codex, Claude or Antigravity to grab it for me.

Currently for projects on my edits page (like Glenroe) it's been invaluable as the first step in training every one of the 6 character LoRAs I have trained and the 6 working their way through klippbok.