Zane-QIU/control3d-pose-edits

Control-3D pose / articulation edits

Image-editing pairs for training a point-cloud-controlled 3D generator (ControlNet branch on TRELLIS image-to-3D). Each row is one (render, edited image) pair: a 1024×1024 render of an Objaverse object from one near-frontal viewpoint, and the same view after an image editor changed only the pose / articulation (raise arms, crouch, open lid, open doors ...). The original 3D assets are not included; get them from Objaverse 1.0 / TRELLIS-500K by uid (see object_index.csv).

Size

• 18,605 pairs (37,210 images) from 4,810 objects; up to 4 edits per object, each edit on a different view.
• Edit models: gemini-nano-banana-2.1 18,313 pairs, doubao-seedream-5-0-pro 292 pairs (only for edits Gemini refused).

group	objects	pairs
humanoid	1,854	7,251
animal_creature	1,171	4,539
articulated_rigid (containers, furniture, vehicles, devices ...)	1,299	4,994
robot_mech	486	1,821

Configs

• default: input render as RGBA PNG (transparent background), edited image as the editor's white-background JPEG.
• black: the same rows with both images on a black background (rgb * alpha, which is what TRELLIS feeds its image encoder), plus their alpha masks. The input alpha is the ground-truth render alpha; the edited alpha comes from background removal with RMBG-2.0 (+ horizontal-flip TTA) on the white-background edit, and the white edge fringe is removed analytically since the background is known to be white (black = I - (1 - alpha) * 255). On renders composited on white (where the true alpha is known), the matting gets median IoU 0.997, 1st percentile 0.94; the remaining failures are near-invisible parts (transparent lids, smoke, thin twigs). matte_flag marks 39 edits whose matte is suspicious (semi_transparent_blob: a large half-transparent region, typically a white display base or ground plate half cut out; area_ratio<0.5: the matted object is under half the area of the input render, usually because a base plate was removed). It is empty for all other rows.

How it was made

1. Objects: 5,000 movable objects (the tier == main rows of object_index.csv) selected from the TRELLIS-500K Sketchfab subset (Objaverse 1.0) with rule filters + a Gemini VLM judging single/complete/movable/shape-change, which also wrote 4 edit instructions per object.
2. Renders: TRELLIS render.py (Blender 3.0.1, Cycles, 1024 px, RGBA), 4 near-frontal views per object: one exactly frontal (yaw 0°, pitch 10°) and three with yaw within ±40° of the front and pitch 5–30°. Edit k uses view k.
3. Editing: the input is the render composited on white (JPEG); prompt = the instruction + fixed constraints (full_prompt: keep identity, details, colors, materials, camera, scale, framing, facing direction, white background; no new objects/props/effects). Output 1024×1024 JPEG.
4. Missing pairs: instructions mentioning items not on the asset (chair, table, "holding a ...", remove/detach ...) were not edited (~1.4%); edits refused by both Gemini and Seedream (mostly copyrighted characters: Marvel, Star Wars, Fortnite skins) are absent. 190 objects have no edits at all.

Columns

column	description
uid	Objaverse 1.0 uid (glb at glbs/<shard>/<uid>.glb; shard in object_index.csv)
name, group, category, movable_type	object name, group, VLM category, pose / open_close / both
k	edit index 0–3 (= view index)
edit_prompt	the pose / articulation instruction
full_prompt	exact text sent to the editor (instruction + constraints)
edit_model	gemini-nano-banana-2.1 or doubao-seedream-5-0-pro
input_image	render of view k, 1024×1024 RGBA PNG (transparent background); black config: RGB PNG on black
edited_image	edited image, 1024×1024 RGB JPEG (white background); black config: RGB PNG on black
camera_yaw, camera_pitch, camera_radius, camera_fov	view parameters (radians; radius 2; fov 40°)
transform_matrix	4×4 camera-to-world matrix (JSON), Blender camera convention (looks along −Z, +Y up)
input_alpha, edited_alpha	(black config only) 1024×1024 L PNG alpha masks
matte_flag	(black config only) empty, or why the edited matte is suspicious (see Configs)
normalize_scale, normalize_offset	normalization applied to the imported scene: scaled by normalize_scale, then translated by normalize_offset (JSON), fitting it in the unit cube [−0.5, 0.5]³

Camera and normalization follow TRELLIS dataset_toolkits (Blender world, Z-up; the glTF importer converts Y-up to Z-up), so TRELLIS-style point clouds / voxels of the original asset align with input_image.

Notes

• The edited image is the target pose; the point cloud of the original asset is the control signal.
• Seedream edits sometimes enlarge the object compared with the input; Gemini occasionally changes the viewpoint slightly.
• Assets come from Sketchfab via Objaverse under their original licenses; many depict copyrighted characters. For research use only.

from datasets import load_dataset
ds = load_dataset("Zane-QIU/control3d-pose-edits", split="train")            # white background / RGBA
ds_black = load_dataset("Zane-QIU/control3d-pose-edits", "black", split="train")  # black background + alphas
row = ds[0]; row["input_image"], row["edited_image"], row["edit_prompt"]