H3 Cinematic Multishot Coverage
Turn one scene photograph into eight target-centered cinematic camera views in one MiniMax H3 generation.
The workflow treats the complete source scene as a semantic Ref2VA reference, names one person or object as the persistent coverage target, and asks H3 to make true editorial cuts between eight static camera setups. It extracts one frame from the stable center of each shot, saves the eight views individually, and stitches a 4×2 contact sheet.
Example
Original semantic scene reference
Prototype scene-coverage result — six selected views from one H3 pass
This prototype result shows the central idea: one semantic room reference can produce several convincing viewpoints while retaining the staircase, wood paneling, furniture language, artwork, lighting, and overall identity of the location. Geometry that was hidden in the source image is necessarily inferred. The downloadable workflow develops the same method into eight explicitly timed cinematic cuts and a 4×2 sheet.
Core nodes only
The workflow uses only nodes shipped with current ComfyUI:
MiniMaxH3ReferenceToVideoMiniMaxH3SigmaShift- standard model loaders and sampling nodes
ImageFromBatch,BatchImagesNode, andImageStitch- standard image/video output nodes
There are no custom-node dependencies. Update ComfyUI before loading the workflow if a native H3 node is missing.
Workflow
Download H3_Cinematic_Multishot_Coverage.json, drag it into ComfyUI, and replace the example image.
- Load one source image in Picture 1 · semantic scene reference.
- Edit only Coverage target · EDIT THIS. Identify one exact visible person, object, architectural feature, or fixed world-space point.
- Leave the two fixed prompt blocks and the 124-frame timing unchanged for the first test.
- Queue the workflow.
- Find the 4×2 contact sheet and eight individual views under
output/H3_Cinematic_Multishot/.
Example target assignment:
the woman seated at the desk in <Picture 1>. Keep this exact woman as the persistent visual subject. Preserve her pose, expression, gaze, wardrobe, desk, room geometry, props, materials, lighting, and shadows across every shot.
Canonical shot plan
The 124-frame packet runs at 24 fps. Later shots follow native H3 shot syntax with exact cut timestamps. The workflow extracts frames away from the cut boundaries:
| View | Cut | Extracted frame | Setup |
|---|---|---|---|
| 1 | opening | 2 | source-matched opening view |
| 2 | 00:00.333 | 15 | 45° clockwise, eye-level 40 mm three-quarter |
| 3 | 00:00.958 | 31 | 90° clockwise, eye-level 65 mm profile |
| 4 | 00:01.583 | 46 | 135° clockwise, low-angle 35 mm three-quarter |
| 5 | 00:02.250 | 62 | 180° clockwise, eye-level 32 mm reverse wide |
| 6 | 00:02.917 | 77 | 225° clockwise, slightly high 50 mm three-quarter |
| 7 | 00:03.500 | 92 | 270° clockwise, 85 mm profile/detail |
| 8 | 00:04.167 | 108 | 315° clockwise, eye-level 50 mm hero three-quarter |
The prompt says these angles are discrete camera placements, not an orbit. Every shot is static from its first frame until the next hard cut. The subject, set, lighting, and all world-space relationships remain frozen; only camera position, height, focal length, and framing may change.
Models
The loader nodes include download metadata for these official models:
| Slot | Suggested file | Folder |
|---|---|---|
| Diffusion | minimax_h3_ref2va_pruned_int8_convrot.safetensors |
models/diffusion_models/ |
| Text encoder | qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors |
models/text_encoders/ |
| Video VAE | minimax_h3_video_vae_fp16.safetensors |
models/vae/ |
| Audio VAE | minimax_h3_audio_vae_fp32.safetensors |
models/vae/ |
Higher-precision variants can improve detail and prompt adherence if memory permits.
Why one video pass?
Eight unrelated image generations tend to redesign the subject and room independently. One H3 video pass shares the same denoising trajectory across the views. This does not guarantee perfect geometry, but it gives the model a stronger continuity context than eight independent generations.
Hard cuts are deliberately more experimental than a continuous orbit. H3 can still blur or interpolate around a requested edit point. The supplied extraction frames sit near the center of each static shot rather than on the cut. If one selected frame is poor, enable Optional · save all 124 frames, inspect the nearby frames, and change that ImageFromBatch index inside the Eight cinematic views subgraph.
A single photograph contains no ground truth for invisible surfaces. H3 must conservatively infer the back of people, furniture, and rooms. Because this workflow uses semantic Ref2VA conditioning, the source is not VAE-locked as an exact first frame: it can preserve identity, scene, style, and relationships strongly, but it may reinterpret geometry. More complete or more literal source coverage requires a more advanced workflow with additional guides.
Credit and inspiration
This experiment is directly inspired by C_Nugget’s H3 Character Sheet Generator, published under the Hugging Face account PoopMan333:
https://huggingface.co/PoopMan333/H3_Character_Sheet_Generator
Their key insight is to derive multiple consistent reference images from one H3 video generation, then extract calibrated frames and assemble them into a sheet. Their original workflow intentionally uses a continuous no-cut orbit for character consistency. This workflow explores the complementary question: can explicit native H3 shot syntax produce target-centered cinematic hard cuts while retaining the same frozen scene?
Please visit and support the original project. This repository does not claim authorship of the character-sheet method or its continuous-orbit prompt.
MiniMax H3 support in ComfyUI comes from Comfy-Org/MiniMax-H3.
License note
The workflow JSON and documentation are provided as-is for experimentation. MiniMax H3 weights and generated outputs are governed by the MiniMax H3 Community License. Review that license yourself before use, especially for commercial work or restricted regions.
- Downloads last month
- -