Loop Forge

13 September 2026

MiniMax H3 camera shots

MiniMax H3 documents twelve camera motions and no named shots. There is no dolly zoom button and no crash zoom button, so every named film shot has to be written as a sentence built out of those twelve. Most ways of doing that do not work. These fourteen do.

Each prompt is the exact file that produced the shot, with the camera and lens phrase and the camera clause marked. More than a hundred generations went into getting here.

Part 2 adds fifteen more, from POV and time-lapse to an FPV drone and a Spider-Man swing, and goes back to Part 1's crash zoom: H3 picks its own pace for a zoom, so the fix is a small script that speeds the zoom up after the render.

Watch on YouTube

MiniMax H3: 14 Camera Techniques. Opens on YouTube in a new tab.

What we found

How do you get real camera moves out of MiniMax H3?

H3 documents twelve camera motions and no named shots, so a dolly zoom or a crash zoom has to be written as a sentence built out of those twelve. Fourteen named shots work that way. Bullet time could not be done.

Settings decide whether the move happens at all. A 4-step turbo render suppresses camera motion almost entirely, whatever the prompt says; use 20 steps and at least 124 frames.

Lens and camera names do almost nothing. On the same seed, an 8mm lens rendered framing identical to a 50mm. Adding a background reference plate roughly halves the camera move.

30 prompts

Part 114 prompts

Fourteen shots built from H3's twelve documented camera motions.

Watch Part 1 on YouTube

Zooms

Specialty shots

Pans and rolls

Tracking shots

Part 216 prompts

Fifteen more, many of them asked for in the comments, and a fix for Part 1's crash zoom.

Video coming soon

Specialty shots

Crash zooms, and the fix

Big moves

Up and down

Run it yourself

Every Part 2 shot page has a zip with the exact graph that produced it, the character plate it loads and a README. Copy the plate into ComfyUI/input, put the four models below in place and drag the workflow onto ComfyUI. With the same model files you should get the same shot or a very close one.

Every node is part of ComfyUI itself, no custom nodes. Part 1's shots use the same graph; their prompts and plates are on each shot's page, with the single workflow file in the links below.

Software

Models

FilePut it inSizeTests
minimax_h3_ref2va_pruned_int8_convrot.safetensorsmodels/diffusion_models21.0 GBEvery shot
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensorsmodels/text_encoders15.7 GBEvery shot
minimax_h3_video_vae_fp16.safetensorsmodels/vae5.2 GBEvery shot
minimax_h3_audio_vae_fp32.safetensorsmodels/vae605.3 MBEvery shot

Reference images

Amber, Character reference plate
AmberCharacter reference plate
Kate, Character reference plate
KateCharacter reference plate
Sofia, Character reference plate
SofiaCharacter reference plate
Allie, Character reference plate
AllieCharacter reference plate
Mei, Character reference plate
MeiCharacter reference plate
Kate, Character reference plate
KateCharacter reference plate
Lena, Character reference plate
LenaCharacter reference plate
Sofia, Character reference plate
SofiaCharacter reference plate
Claire, Character reference plate
ClaireCharacter reference plate
Jin, Character reference plate
JinCharacter reference plate
Marcus, Character reference plate
MarcusCharacter reference plate
Freya, Character reference plate
FreyaCharacter reference plate
Raven, Character reference plate
RavenCharacter reference plate

The model, and the machine

Workflow
MiniMaxH3ReferenceToVideo, ref_image_size max, no background plate
Resolution
1344×768. Part 1 on a rented RTX 5090, except the snorricam at 864×480 on an RTX 3060. Part 2 partly on the rented 5090 and partly on the 3060, about 21 minutes a clip
Sampling
20 steps, res_multistep / simple, turbo LoRA off. Seed 552013742 in Part 1, 7 in Part 2 (11 for the fisheye)
Prompt format
The six-section reference format from MiniMax's official guide

Materials and links