Ask base H3 for "everything is frozen in time" and it does not hold the freeze: people keep moving, spilled coffee keeps splashing, newspapers keep falling. This LoRA teaches one behaviour: the whole scene stands perfectly still while ONE person keeps walking toward the camera.
It works best in interiors with seated or still people and objects caught in mid-air (loose papers, a newspaper, rose petals).
What it does
Freezes the world in place while one person keeps walking.
Best with seated / still people and suspended objects in interiors (library, cafe, office, train carriage, concert hall).
The before/after clips use the same prompt, without and with the LoRA (seed 101; the train pair is base seed 101, LoRA seed 202).
Recipe that works
Trigger: timefr33ze · strength 1.0 · 1024x576 · 107 frames @ 24 fps · 20 steps · cfg 1 · text-to-video.
Prompt pattern (copy it, change the scene):
integrated_multimodal_description: [Shot 1] timefr33ze, photorealistic cinematic footage, the entire world is frozen in time. A grand old library reading room. Readers sit frozen at long tables, and dust motes and loose pages hang motionless in the sunbeams. Only a man in a grey coat moves, walking toward the camera; the camera holds perfectly still.
overall_soundscape: A steady hushed ambience of the reading room, with measured footsteps on a wooden floor.
non_diegetic_music: None.Rules of thumb:
Start with the trigger: "timefr33ze, photorealistic cinematic footage, the entire world is frozen in time."
List the frozen things (people, objects in mid-air).
Say that ONE person moves, and which way: "Only <one person> moves, walking toward the camera".
End with "the camera holds perfectly still."
Soundscape: a steady hushed ambience plus the walker's footsteps.
Known limits
Crowds that are walking or standing keep moving.
Birds and fountains keep moving.
A moving camera (tracking / gliding) weakens the freeze; keep the camera still.
Strength 0.8 and 1.2 did not help; use 1.0.
Tested at 1024x576, 107 frames. 1344x768 is untested for v1.0 — if the freeze weakens, drop to 1024x576.
Small floating props you ask for may not appear (e.g. coffee or sugar cubes in mid-air).
Not tested: image-to-video, stacking with other LoRAs, turbo / distilled sampling.
Training
21 clips (12 Blender renders incl. one clay look + 9 base-H3 clips), 3,000-step run on Civitai Training Studio (MiniMax H3 base, rank 32, 10 checkpoints); this release is epoch 6 of 10. All 3D assets CC0 (Poly Haven, Quaternius, Kenney). Audio is clean ambience with footsteps and no low hum.
Workflow
Free "H3 Camera Director" (16 GB-friendly, preview / final modes, scene helper, LoRA slot; keep the camera still for Frozen World): https://civarchive.com/models/2986105
Changelog
v1.0 — first release: 21 clips (13 walker + 8 camera-move), clean ambience audio without low hum, 2 clips removed after Civitai's SFW pre-check.
Entered in the Civitai H3 Training Contest — Concept.
Description
v1.0 (epoch 6 of 10). Trained on 21 clips (13 walker + 8 camera-move) with clean ambience audio (no low hum). Best with a still camera, seated/still people and suspended objects; walking crowds still move. Strength 1.0.