H3
Can H3 Run on an RTX 5090? A 32GB VRAM Benchmark
Measured H3 Ref2VA runtime and peak VRAM on one RTX 5090, with exact input limits, staged memory handling and the limits of a single benchmark.
Experiment: 2026-08-11MODEL FIELD NOTES
AI video and image experiments
Measured runtime, memory limits, storyboard layouts and failure analysis. Read our dated AI video and image experiment records.
MAGI-2
Sand.ai MAGI-2 Multi-GPU Deployment: Why More GPUs Did Not Help
MAGI-2 Preview experiments with RTX 5090 GPUs: EP4, CPU offload, cold versus warm latency, and the limits of scaling to eight cards.
Experiment: 2026-08-13Image experiments
Qwen Image Storyboards: Four Panels, Twelve Panels and Pro Trade-offs
A small storyboard case study covering panel counts, numbering, character continuity and layout failures, with provider labels and sample limits made explicit.
Experiment: 2026-08-24Image experiments
Why AI Upscaling Changes Image Structure: Sharpness vs Fidelity
A Qwen image-editing case study where high similarity scores coexisted with structural changes, plus a practical review method and interpolation control.
Experiment: 2026-08-15H3
H3 OOM Troubleshooting: Why Still References Fit in 32GB While Video References Crash
Measured H3 memory boundaries on 32GB GPUs: comparing 1152×640 sparse references against 1344×768 motion video references, plus the 5-frame first-frame extraction trick.
Experiment: 2026-08-14Verification
Why Your AI Video Returns 3:4 Instead of 9:16: Auditing Silent Parameter Mismatches
Hash-verified measurements of AI video outputs versus request payloads: why 9:16 requests return 480×640 (3:4) with HTTP 200, and how palindrome loops double shot duration.
Experiment: 2026-09-05Workflow
Why Keyframe Refinement Causes Face and Action Drift in Reference-to-Video Pipelines
Failure analysis of the 3-step storyboard-to-refinement-to-video cascade: why full-frame image polishing breaks character continuity and how R2V differs from frame-one locking.
Experiment: 2026-08-22Direction
Controlling Two-Actor AI Video Prompts: Fixed-Seed A/B/C Tests and Auto-QC False Positives
Results from our September 5, 2026 fixed-seed prompt experiment on Wan video generation: why text prompts struggle with two-actor contact and why automated QC misses extra-limb artifacts.
Experiment: 2026-09-05