Hardware baselines and MiniMax H3 memory allocation
Our card is an RTX 5060 Ti with 16 GB. MiniMax H3 at 704x1280 (9:16), 124 frames, 20 steps ran at 88 seconds per step with 10.1 GB of VRAM, and took 31 minutes 47 seconds in total. That is our delivery setting.
Executing local video diffusion pipelines on consumer hardware requires precise optimization of latent spatial dimensions against available VRAM budgets. On our primary production workstation equipped with a 16 GB NVIDIA RTX 5060 Ti, MiniMax H3 achieves exceptional visual quality at a vertical resolution of 704x1280. Generating 124 frames across 20 sampling steps consumed 10.1 GB of dedicated video memory, executing steadily at 88 seconds per step for a total wall-clock runtime of 31 minutes 47 seconds. This configuration provides a stable baseline for narrative production without triggering memory paging.
832x480 landscape ran at 56 seconds per step and 9.5 GB. It works, but it is too soft for a close-up and is the wrong shape for a reel.
Downscaling spatial resolution accelerates iteration speeds at the expense of commercial utility. Rendering landscape 832x480 frames executed in 56 seconds per step utilizing 9.5 GB of VRAM. While reliable for rapid mechanical motion tests, the resulting output lacks the high-frequency detail required for character close-ups and requires horizontal letterboxing for vertical social formats.
The 1088x1920 memory cliff and shared RAM thrashing
1088x1920 vertical ran at 248 seconds per step, with 15.5 GB dedicated VRAM and about 28 GB of shared GPU memory. One step took 4 minutes 8 seconds. We killed it. It does not fit.
Attempting to generate native high-definition video directly inside local diffusion decoders immediately breaches consumer hardware limits. When our benchmark suite initiated a native 1088x1920 generation, dedicated VRAM consumption climbed to 15.5 GB before spilling an additional 28 GB of tensor allocations into Windows shared system memory. Diffusion velocity collapsed instantly to 248 s per step—a crippling 4 minutes 8 seconds per single sampling iteration. The process was manually terminated; native 1088x1920 does not fit on 16 GB cards.
Windows reports a large shared-GPU figure on long renders even when the job is healthy. The thrash signature is per-step time exploding, so we judge by step time, not by the shared number.
Interpreting system memory telemetry requires identifying true operational bottlenecks. Because the Windows operating system frequently pre-allocates substantial shared-GPU virtual memory pools during extended diffusion jobs, shared memory statistics alone do not indicate pipeline failure. The unmistakable signature of memory thrashing is an exponential explosion in per-step processing duration.
WanGP resolution alignment and the 2x upscale delivery pipeline
WanGP silently aligns resolution to multiples of 32: asking for 720x1280 returns 704x1280. Aspect follows the source still, so a 3:4 plate returned 832x1088 even with a 704x1280 setting. Crop to 9:16 upstream.
Underlying inference wrappers enforce strict tensor alignment rules that alter output geometry. Within the WanGP backend, spatial dimensions are automatically snapped to the nearest multiple of 32. Requesting a standard 720x1280 vertical raster silently outputs a 704x1280 file. Furthermore, the engine dynamically adapts aspect ratios to match input reference stills: feeding a 3:4 portrait still into a 704x1280 job forces an output of 832x1088. To preserve vertical social proportions, source stills must be pre-cropped to exact 9:16 boundaries prior to submission.
Local AI video generation benchmarks on RTX 5060 Ti 16GB workstation · measured 16 Aug 2026
Model Architecture
Target Resolution
Steps / Frame Count
Dedicated VRAM
Step Timing / Wall-Clock
Operational Verdict
MiniMax H3
704x1280 (9:16)
20 steps / 124 frames
10.1 GB
88 s/step (31m47s total)
Studio delivery baseline
MiniMax H3
832x480 (landscape)
20 steps / 124 frames
9.5 GB
56 s/step
Soft detail; wrong aspect for reels
MiniMax H3
1088x1920 (9:16)
20 steps / 124 frames
15.5 GB (28 GB shared)
248 s/step (4m8s/step)
Thrashed; killed; exceeds 16 GB
MiniMax H3 (Turbo LoRA)
704x1280 (9:16)
4 steps / 124 frames
10.1 GB
5m15s total (vs 7m21s @ 8 steps)
Rapid motion test rail
LTX 2.5 i2v
704x1280 (9:16)
8 steps / 97 frames
Offloaded 1/3 model
57.8 s/step (7m42s total)
Practical ceiling at 97 frames
Evaluated on local RTX 5060 Ti 16GB hardware.
Our delivery path is to generate 704x1280 and upscale 2x to 1408x2560. That took about 32 minutes plus about 2 minutes for the upscale, against more than 85 minutes for native 1080p, which thrashed.
The proven commercial workflow decouples spatial synthesis from high-resolution upscaling. Generating at 704x1280 followed by a dedicated 2x upscale pass to 1408x2560 delivers pristine high-definition footage in about 32 minutes plus about 2 minutes for the upscale, bypassing the crippling thrashing of native 1080p.
Turbo LoRA scaling and clip duration efficiency
Turbo 4-step LoRA on the same shot and seed: 5 minutes 15 seconds, against 7 minutes 21 seconds at 8 steps and 31 minutes 47 seconds at 20 steps.
Distilled sampling schedules provide dramatic efficiency gains during previz testing. Deploying a distilled 4-step Turbo LoRA against our standardized test shot completed in 5m15s, compared to 7 minutes 21 seconds at 8 steps and 31 minutes 47 seconds at full 20-step fidelity, enabling rapid preview iterations.
H3 first-and-last-frame at 243 frames (10.13 seconds) took 31 minutes 32 seconds at 704x1280, the same wall-clock as a 5-second clip. For 10 seconds of footage we generate one long clip and cut it, rather than rendering two 5-second clips.
Batching continuous temporal lengths optimizes local GPU utilization. When executing keyframe interpolation in H3, generating an extended 243-frame clip (10.13 seconds) required 31 minutes 32 seconds—virtually identical to the wall-clock duration of a 5-second take. For narrative sequences requiring 10 seconds of coverage, generating a single extended take and trimming it in post-production halves overall render time compared to rendering separate 5-second segments.
Identity-locked H3 reference-to-video at 704x1280 took about 32 minutes per 5-second clip. A 10-second version took 57 minutes at 15.2 GB of the 16.3 GB card, so 5 seconds is the practical length.
LTX 2.5 frame ceilings and backend memory degradation
LTX 2.5 image-to-video at 704x1280, 97 frames, 8 steps ran about 2.5 minutes warm. On a balloon-plate image-to-video run the job took 7 minutes 42 seconds at 57.8 seconds per step, and the card offloaded about a third of the model.
LTX 2.5 offers high sampling velocity on modest hardware configurations. In warm benchmarks, a 97-frame sequence at 8 steps executed in approximately 2.5 minutes. On complex prop plates requiring extensive offloading, the pipeline completed in 7 minutes 42 seconds at 57.8 seconds per step with one-third of the network weights offloaded to system memory.
LTX at 125 frames took 22 minutes 51 seconds, against the 97-frame cost, because attention cost rises faster than frame count. Our measured practical ceiling is 97 frames, about 4 seconds.
Because attention memory complexity scales non-linearly with sequence length, expanding frame counts from 97 frames to 125 frames caused runtimes to skyrocket to 22 minutes 51 seconds. In local studio production, 97 frames represents our hard operational limit.
LTX backend degradation: three identical clips in one session went from 57.8 to 90.3 to about 220 seconds per step. Restart the backend before every clip in a batch.
The LTX latent upscaler lost about 58% of Laplacian detail on our test, while looking sharper. We keep ArtPlayer Fast as the finishing upscaler.
