MODEL FIELD NOTES · H3
H3 OOM Troubleshooting: Why Still References Fit in 32GB While Video References Crash
Near the 32GB VRAM ceiling, whether an H3 job succeeds or crashes often depends on how the reference input is encoded rather than output duration alone. This record compares sparse still references against motion video references and documents a 23.1-second 5-frame keyframe technique.
阅读本站中文版 · 在黄果制片阅读中文版Why motion video references cross the 32GB boundary
During our August 2026 troubleshooting runs on 5-second clips, we isolated why jobs that succeeded with a still character image crashed with an out-of-memory error as soon as a continuous motion reference video was attached.
A motion reference does more than add conditioning frames to the diffusion latent. It forces the video VAE and conditioning encoders to hold multi-frame activations concurrently. At 1344×768 for a 5-second target, attaching a continuous motion reference pushed peak allocation past the 32GB hardware limit. Dropping the resolution to 1152×640 and switching to a sparse or still reference kept peak memory near 29.15 GB and allowed the job to finish reliably.
Enabling full model CPU offload prevented the crash on heavier configurations, but increased wall-clock execution time by roughly sixfold because weights had to shuttle across PCIe between stages. These measurements describe our tested build and should not be read as a universal memory table for every quantization package.
| Configuration | Resolution / Duration | Reference Mode | 32GB GPU Outcome |
|---|---|---|---|
| High-res motion reference | 1344×768 / 5 s | Continuous motion video reference | OOM crash (>32 GB) |
| Sparse still reference | 1152×640 / 5 s | Sparse / single-frame still reference | Pass (~29.15 GB peak) |
| Quantized long clip baseline | 352×640 / 15.083 s | 3.5 s downsampled ref + staged unload | Pass (31,658 MiB peak) |
| Full CPU offload fallback | 1152×640 / 5 s | Staged --offload_model enabled | Pass (~6× slower wall time) |
Using H3 as a still generator: the 5-frame extraction test
When a separate text-to-image model produces character stills whose lighting or anatomy shifts during video generation, generating the anchor frame inside the video model itself keeps the visual distribution aligned. We tested this by running H3 at 960×544 for 30 sampling steps with a 5-frame output target.
The run completed in 23.1 seconds and produced one 5-frame MP4 alongside five extracted PNG frames. Because the temporal dimension was capped at five frames, we could afford 30 steps instead of 12 without exhausting memory, then select frame one as the locked character reference for subsequent shots. This is a distribution-matching workflow trick rather than a speed benchmark against dedicated image models.
| Parameter | Tested Value | Measured Result |
|---|---|---|
| Resolution and frame count | 960×544 / 5 frames | 1 MP4 + 5 PNG frames exported |
| Sampling steps and wall time | 30 steps / 23.1 s | Clean frame-1 anchor for 3-shot continuity |
Practical recovery order when a shot hits OOM
Retrying an identical out-of-memory payload wastes queue time. Change one memory driver at a time in the following order, or pre-check your parameters in the Porncut VRAM Calculator.
- Replace continuous video references with a 1-frame still or a 5-frame static anchor.
- Step resolution down from 1344×768 to 1152×640 or 896×512 and split 15-second narratives into three shots.
- Verify that text encoders and VAEs are released between pipeline stages (--reserve-vram 1.0).
- Reserve full CPU offload for shots where resolution cannot be reduced and a 6× runtime penalty is acceptable.