Field notes · The testing bench
New models that can make a video asset land almost weekly. This is where we post the ones we actually put in front of real film and television work — what held up, what fell apart, and the licensing question attached to each.
Why this page exists
Staying on top of every new model that can make a video asset is a full-time job that never ends. A new text-to-video, image-to-video or lip-sync model seems to land almost weekly — some are genuinely better, most are a re-skin of the same weights, and the one that looks unbeatable in a five-second demo is often the one that falls apart the moment you hand it real footage: a 4K ARRI frame, an hour of continuity, a scene that has to cut with the shot before and after it.
Almost every week there are multiple new models, nodes and workflows out in the wild — and almost none of them arrive compliant with film and television regulations. So a big part of the job is working out what human element or involvement makes a given approach ethical for the unions, guilds and studios. A workflow or model that was the right tool last month — or last week — may not be the best tool right now. Testing is almost a full-time job on its own.
Then there's the harder question underneath all of it — the licensing and the legality. Which model's weights actually permit commercial use? What went into the training data, and can you prove it? Who owns the output when a copyright office won't grant protection to something a machine made? What do the guilds, the network and the E&O insurer expect to see in the paper trail five years from now? Those questions decide “should we use model X” far more than the demo does.
So this is where we post the things we've actually tested — attempts, dead ends and wins — measured against one bar: does it work on real film and television, and can you defend it afterward? Everything below has been through a real task, not a prompt on a leaderboard.
The log
9 tools, each run against a real film/TV task — with the licensing question that comes with it.
Image → Video · image-to-video / video-to-video
The workhorse for motion on real plates. Holds up well enough to cut into a sequence, and runs locally so footage never leaves the machine.
Licensing
Open weights — verify the exact licence revision before commercial delivery.
Text → Video · text-to-video
Best-in-class motion coherence for a text prompt in our tests, but still a concept tool: it generates a shot, not a shot that matches your scene.
Licensing
Open weights with usage conditions — read the terms, region restrictions apply.
Text → Video · text-to-video / video extension
Fast, and useful for extending an existing clip. Lower fidelity than WAN on faces; great for backgrounds and establishing shots.
Licensing
Check model card — commercial use is permitted on recent releases.
Image · text-to-image
Reliable, controllable, and the model most of our image pipelines are built around. Pairs with LoRA training for consistent characters.
Licensing
Open weights for the base model; hosted endpoints have their own terms.
Pipeline · character / object consistency
Single-purpose LoRAs are the difference between a novelty and a repeatable look. Trained locally so a client design never leaves the box.
Licensing
You own the trained weights; base-model licence still applies to outputs.
Image · character turnaround / continuity
Turning a character into a locked reference sheet is how you keep a face consistent across hundreds of frames. Still needs a human eye each pass.
Licensing
Derivative of the base model and any input reference — clear the inputs.
Audio · lip-sync dialogue
Dialogue-driven lip-sync that survives a close-up. The technology works; the paperwork around a real performer’s likeness is where it gets decided.
Licensing
Consent and performance rights are the real gate here, not the licence.
Audio · voice / narration
Convincing for temp narration and scratch tracks. Anything that sounds like a real person needs documented consent before it goes past the edit.
Licensing
Voice is identity data — consent, guild rules and insurer expectations govern.
Pipeline · our own IP
Reverses an on-set grade back to camera LOG for grading and reference. Built because nothing off the shelf did the job correctly.
Licensing
Owned, proprietary — no third-party terms.
Featured test
A ComfyUI workflow that generates a 1024×512 frame with Z-Image Turbo, then diffusion-upscales it to 4096×2048 (4K) with a PiD Flux model — a 4× increase on each axis. Drag the slider to compare the generated frame against the upscale.
Z-Image Turbo — 1024×512 PiD 4K upscale — 4096×2048 The verdict
“Although this seems to look nice… my inner DIT says way too crunchy, and I’d find a way to back off some of that hard edge stuff with a lift adjustment.”
— DIT feedback on the 4K upscale output
A UNETLoader loads Z-Image Turbo (bf16) and a CLIPLoader loads the Qwen 3.4B text encoder. The KSampler generates a 1024×512 latent in 4 steps using LCM sampling.
The generated image is VAE-encoded to latent space and fed into a second Z-Image Turbo (PiD model) pass that extracts detailed structure for the upscale target.
A PiD Flux model (pid_flux1_1024_to_4096_4step_bf16) loads via a second UNETLoader. SamplerCustom runs the full 4-step diffusion to produce a 4096×2048 latent.
A Pixel Space VAE decodes the upscaled latent back to pixels. Image Comparer nodes give inline A/B review of the before and after results.
| Base model | Z-Image Turbo (bf16) via UNETLoader |
|---|---|
| Text encoder | Qwen 3.4B (lumina2) |
| VAE | ae.safetensors (Z-Image) + pixel_space |
| Upscale model | pid_flux1_1024_to_4096_4step_bf16 (PiD) |
| Upscale method | Diffusion-based 4× upscale (1024 → 4096 on long edge) |
| Output resolution | 4096 × 2048 |
| Sampling | LCM scheduler, 4 steps per pass, cfg = 1 |
| Custom nodes | SetNode / GetNode, PiDConditioning, EmptyChromaRadianceLatentImage, ReservedVRAMSetter, Image Comparer (rgthree), Fast Groups Bypasser, AILab_QwenVL_PromptEnhancer |
Workflow adapted from Benji’s AI Playground. Tested here against real film/TV imagery — the verdict above is a working DIT’s read, not a vendor benchmark. Third-party workflow: check its licence before commercial delivery.
Project in progress
I'm currently working to land a true-crime project, and the bar is higher than scripted: everything has to be factually defensible and legally clean, because it touches real people, real archives and real harm. These are the attempts as they happen — what we're testing, and what has to be true before any of it could air.
Can a current upscaler resurrect degraded period broadcast masters without inventing detail that misrepresents the archive? The line between “restored” and “fabricated” is the whole question on a true-crime doc.
Generating era-correct props, signage and establishing material when the archive has no shot. Must clear a factual-accuracy review before it can appear in a documentary.
What can and cannot be done with a real subject’s likeness and voice under broadcast standards, consent and E&O review. Most of this is a legal call, not a technical one.
Testing reconstructed narration against union, consent and disclosure expectations. Disclosure and consent are prerequisites, not afterthoughts.
Results are logged as tests complete. Anything shown here is decision-support, not legal advice — licensing and consent calls belong with your lawyer, your union and your insurer.
Bring me the shot and the deadline. I'll tell you which tools can actually get you there — and which licensing, consent or provenance problems will bite before the delivery date.