Test Results

Image and video model evaluations across all three HighVis rendering forks.

Cross-Fork Status

Where each rendering fork stands as of 6 May 2026.

	Classic	Hybrid	Local
Status	Active — quality benchmark	Dormant — waiting on API access	Active — pipeline confirmed, quality tuning
Image Model	Nano Banana Pro (proprietary)	Same as Classic (via Robin's infra)	Nano Banana Pro (proprietary — open-source alternatives tested and rejected)
Video Model	Kling o1, Seedance 2.0, Luma, Wan	Same models via abstraction layer	LTX-2.3 22B + HDR IC-LoRA (confirmed working)
Cost per shot	~$1.00-1.22	TBD	$0.00 (GPU rental only)
Output format	8-bit MP4	TBD	16-bit HDR EXR (planned)
Tests run	8	0	18 (3 image eval + 15 parameter tuning runs)
Key finding	Camera motion preserved (F1)	—	Pipeline working — motion and style are separate controls (F-LC-007)

Image Model Evaluation — 30 April 2026

Local fork testing: can an open-source image model match the quality of the proprietary Nano Banana Pro used by Classic?

The benchmark to beat

Source frame — Unlit 3D render from Unreal Engine. This is what every image model receives as input.

Nano Banana Pro result (Classic fork) — The quality bar. Realistic materials, correct lighting, all objects from the source preserved in place.

Z-Image test results — rejected

High creativity — walls painted but all objects from source lost

Mid creativity — clean surfaces but empty room, nothing from source survives

Best Z-Image result — better textures but geometry oversimplified

Low creativity — room shape preserved but still looks unlit, no realistic materials added

Configuration error — wrong text encoder produced garbled text in output

Closest to source composition — still painterly, compare to Nano Banana above

Face consistency test

Generation 1 — photograph input

Generation 2 — same input, different face

To confirm the model wasn't broken, we tested it on a photograph instead of a 3D render. The model produced higher-quality results but changed the subject's face between generations — confirming the model works but can't preserve what it's given. It generates well, but it doesn't respect the reference.

Verdict

Z-Image rejected. Next up: Flux Dev with depth-guided generation — instead of asking the AI to copy pixels from the source (which degrades as you give it more creative freedom), we extract the 3D depth structure from the source frame and use that as a guide. The AI gets full creative freedom on textures and lighting while the depth map locks the composition in place.

Local Fork — Parameter Tuning Dataset (6 May 2026)

15 runs on the Spaceship scene, varying sigma schedule, CFG, and LoRA strength to find the best balance of motion preservation and style transfer. Every run uses the same pipeline: source video restyled by Nano Banana, upscaled by 4x-UltraSharp, then rendered by LTX 2.3 22B on an RTX 5090. Click any video to play.

Key findings from this session

F-LC-007: Motion and style are controlled by separate settings. The noise schedule (sigmas) controls how much the AI changes the camera move. CFG and prompts control how strongly the new look is applied. You can lock one without affecting the other.

F-LC-008: The restyled look fades over time. The AI uses a single reference image (the restyled first frame) to guide the entire clip. Style is strongest at the start and drops to about 60% after the first second.

F-LC-009: Proposed fix: split the video into 2–3 second chunks, generate a fresh reference for each chunk, restyle each chunk independently. Planned as the V6 workflow.

Source — Unlit 3D render from Unreal Engine (Spaceship scene). This is the input to every run.

Nano Banana reference — The restyled first frame that guides the AI's output. Same reference used across all 15 runs.

Classic Fork Results

Filterable test grid for the Classic rendering fork. Click any card to compare source vs output.