OpensourceBlender/bowl_34_h3_inputs

bowl_34 — MiniMax-H3 depth ControlNet inputs

One Blender scene (bowl_34, "Complete Foley recording studio") rendered along a smooth camera trajectory, with everything needed to drive MiniMax-H3 + Fun-ControlNet-Union-2.0 (depth) and to check the output against ground truth.

All per-frame data share one pixel grid and one frame index: 1280×720, 81 frames, 16 fps (frame k ↔ t = k/16 s).

Layout

Path	Content
render/rgb_frames/%06d.jpg	Blender Cycles RGB render (ground-truth appearance)
render/rgb.mp4	Same frames as a 16 fps video
depth/depth_mm_uint16/%06d.png	Metric depth, uint16 millimetres (0 = invalid)
depth/depth_preview.mp4	Depth visualisation only, not for control
instance/instance_id_png/%06d.png	Instance ID per pixel (0 = background); IDs match instance_id in the annotations
annotations/per_frame/%06d.json	Per frame: intrinsics, camera pose (camera-to-world, metres, OpenCV axes), visible objects with 3D OBB / 2D boxes
annotations/video_annotation.json	Clip-level object list (24 objects): category, world OBB, visible_frames
annotations/trajectory.json	Camera trajectory (keyframes left → front → right, interpolation, stats)
annotations/categories.json	Category table
annotations/source.json	Source .blend / annotation paths and SHA-256
overlay/rgb_with_visible_object_3dbox_labels.mp4	RGB with projected 3D boxes + category names of visible objects
control/depth_control_frames/%06d.png, control/depth_control.mp4	Depth control: inverse depth, near = white, far/invalid = black (range in depth_control.json)
control/h3_depth_control_1280x736_90f.mp4	H3-ready control video (see below)
source_scene/scene/scene.blender	Blender 5.0.1 scene used for this render (SHA-256 matches annotations/source.json), with its generator scene.py, source annotation.json, views.json and keyframe-camera previews/
source_scene/generation/	Earlier generation stage: scene.blend, blender_code.py, the code-generation prompt code_prompt.txt, scene_description and the reference input_image.png
source_scene/video_render_code/	Scripts that produced this clip: make_trajectory.py → render_video.py → make_preview.py, make_depth_control.py
prompt/prompt.txt	Text prompt (hand-written from the scene name and annotated categories; not generated by Blender)

H3-ready control video

H3 needs height/width divisible by 32 and 17n+5 frames. To keep pixel- and frame-level alignment with the GT:

• Spatial: the 1280×720 depth control is edge-padded by 8 rows top and bottom → 1280×736 (no resize, no crop).
• Temporal: the last frame is repeated 9 times → 90 frames; frames are read one-to-one (no fps resampling).
• Lossless RGB encoding; rows 8..727 of frames 0..80 equal control/depth_control_frames exactly (max abs diff 0).

Run with sample_size = [736, 1280], no fps resampling, video_length = 90, then take rows 8..727 and frames 0..80 of the output: that crop is pixel- and frame-aligned with render/, depth/, instance/ and the annotations.

Recommended settings (from the ControlNet card): 40 steps, guidance_scale = 1.0, control_context_scale = 1.0, config minimax_h3_control_inpaint_post_norm.yaml, base weights = MiniMax-H3 FL2VA.

Note: H3 is trained at 24 fps; this clip is 16 fps, so per-frame motion is ~1.5× larger than in training.

Reproduce the render

SCENE=source_scene/scene/scene.blender
blender -b --factory-startup --disable-autoexec $SCENE --python source_scene/video_render_code/make_trajectory.py -- --scene-dir out
blender -b --factory-startup --disable-autoexec $SCENE --python source_scene/video_render_code/render_video.py -- \
    --scene-dir out --annotation source_scene/scene/annotation.json

The scene is fully procedural (no external textures). Absolute paths inside the JSON files refer to the original cluster layout and are informational only.