idol-3d-objects / README.md
markhitvdrz's picture
Upload README.md with huggingface_hub
871a672 verified
|
Raw History Blame Contribute Delete
7.82 kB

Idol 3D Objects

Prompt → 2K transparent PNG → 3D Gaussian-splat .ply.

A ComfyUI workflow generates a single object on a transparent background with Qwen-Image-2.1-Turbo, then Idol reconstructs it as a 3D Gaussian splat from that one image.

Sample images and PLYs are also mirrored as a Hugging Face dataset: https://huggingface.co/datasets/markhitvdrz/idol-3d-objects

Tested on a laptop with an 8 GB RTX 5050 (sm_120), Windows 11, 16 GB RAM.

preview

Samples

object 2K RGBA 3D splat
smartphone samples/images/smartphone.png samples/ply/smartphone.ply
laptop samples/images/laptop.png samples/ply/laptop.ply
wireless headphones samples/images/wireless_headphones.png samples/ply/wireless_headphones.ply
coffee mug samples/images/coffee_mug.png samples/ply/coffee_mug.ply
running shoe samples/images/running_shoe.png samples/ply/running_shoe.ply
armchair samples/images/armchair.png samples/ply/armchair.ply

Open a .ply in any 3D Gaussian splat viewer (SuperSplat, antimatter15/splat, Polycam web viewer).

How it works

prompt
  |  ComfyUI  ·  Qwen-Image-2.1-Turbo Q5_K_M (GGUF)  ·  8 steps, CFG 1
  v
2048 x 2048 RGBA PNG          (transparency comes from the prompt, not a matte)
  |  Idol  ·  BiRefNet background removal  ·  DINOv3 + FLUX.2 VAE conditioning
  v
.ply  ·  262,144 3D Gaussians  ·  standard 3DGS format

Qwen Image 2.1 supports RGBA natively: wrapping the prompt as "This is an RGBA format image with transparency. … The image has an alpha channel and a transparent background." makes the model emit a real alpha channel. No segmentation or matting step is involved.

Idol detects that alpha and skips its own background removal, so the two stages stay clean.

Any image → 3D, inside ComfyUI

comfyui/ComfyUI-Idol adds a node that runs Idol in-process, so any image — not just a generated one — can be turned into a .ply from the ComfyUI graph.

workflow pipeline
workflows/image_to_3d_api.json LoadImage → Idol → .ply
workflows/image_to_transparent_3d_api.json LoadImage → Qwen Image 2.1 edit (reference image + RGBA prompt) → Idol → .ply

Install (the node is a junction into custom_nodes; see comfyui/ComfyUI-Idol/README.md):

python_embeded\python.exe -m pip install diffusers tensordict plyfile pyvers cloudpickle accelerate
python_embeded\python.exe -m pip install -e C:\idol\idol-sdk
set HF_HOME=C:\idol\hf-cache

Run any local image through either one:

python scripts/run_image.py --image photo.jpg --workflow workflows/image_to_3d_api.json
python scripts/run_image.py --image photo.jpg --workflow workflows/image_to_transparent_3d_api.json

Measured through ComfyUI on this laptop: 238.8 s for the direct path (first call includes the Idol load), 202 s for the Idol stage of the transparent path, plus ~95 s for the Qwen edit. Repeats are faster because the weights are in the OS cache.

On a 16 GB machine ComfyUI can only serve one prompt per process — reloading the 8.7 GB text encoder crashes with an access violation (hostbuf_file_reader_read failed / "Windows fatal exception"). Restart ComfyUI between runs, or drive the same workflows from scripts/generate_with_restart.py, which restarts it for you.

Setup

1. ComfyUI + Qwen Image 2.1

Use ComfyUI portable (v0.44.0 tested) and the GGUF loader:

git clone https://github.com/leejet/ComfyUI-GGUF ComfyUI/custom_nodes/ComfyUI-GGUF
python_embeded/python.exe -m pip install -r ComfyUI/custom_nodes/ComfyUI-GGUF/requirements.txt

Models (Qwen Research License, see NOTICE):

file folder size
qwen-image-2.1-turbo-Q5_K_M.gguf models/diffusion_models/ 4.6 GB
qwen3vl_8b_int8_convrot.safetensors models/text_encoders/ 8.7 GB
qwen_image_2.1_vae_bf16.safetensors models/vae/ 0.6 GB

From abenzerps/Qwen-Image-2.1-Uncensored-GGUF on Hugging Face.

On 8 GB VRAM the new async/pinned offload path fails with hostbuf_file_reader_read failed; start ComfyUI with:

python_embeded\python.exe -s ComfyUI\main.py --port 8188 ^
  --disable-async-offload --disable-dynamic-vram --disable-pinned-memory

Even with those flags, reloading the 8.7 GB text encoder between prompts can crash with an access violation on a 16 GB machine. scripts/generate_with_restart.py works around it by starting a fresh ComfyUI process per image. Slower, but it never fails.

2. Idol

pip install -e idol-sdk        # Python 3.10+, PyTorch 2.5+, CUDA GPU

You also need access to the two gated encoders (facebook/dinov3-vith16plus-pretrain-lvd1689m, black-forest-labs/FLUX.2-dev) and hf auth login.

Run

# 1. images (ComfyUI must be running on :8188)
python scripts/generate_images.py \
  --workflow workflows/qwen_2.1_turbo_api.json \
  --objects prompts/objects.txt --template prompts/prompt_template.txt \
  --out samples/images --size 2048

# 1b. or, on 8 GB / 16 GB RAM, one ComfyUI process per image (robust)
python scripts/generate_with_restart.py --comfy C:\ComfyUI_windows_portable \
  --workflow workflows/qwen_2.1_turbo_api.json \
  --objects prompts/objects.txt --template prompts/prompt_template.txt \
  --out samples/images --size 2048

# 2. 3D
python scripts/image_to_ply.py --images samples/images --out samples/ply \
  --weights speridlabs/idol --gaussians 262144 --steps 50

# or both
python scripts/run_all.py --size 2048

The workflow file is the ComfyUI API format (what /prompt accepts). To use it in the ComfyUI UI, load the official template Qwen Image 2.1 Turbo — Text to Image from the Template Library and swap Load Diffusion Model for Unet Loader (GGUF).

Measured

stage 1024² 2048²
Qwen generation (8 steps) 122 s first run 212–248 s
Idol → 262k Gaussians 145 s 140 s (input resized to 1024 internally)
peak VRAM ~6.7 GB ~6.7 GB

2K numbers include a fresh ComfyUI process per image (the restart workaround); sampling itself is ~140 s of that. Idol is a separate process and reuses the bf16 weights (load ~70 s, then ~140 s per object). Six objects end to end: about 40 minutes on this laptop.

Where the time goes, and what does not help

Per object at 2K: ~45 s ComfyUI process start, ~50 s model load, ~140 s sampling, ~20 s decode, then ~140 s Idol (of which 134 s is the 50-step denoiser; preprocess 0.8 s, encode 3.8 s, decode 4.1 s).

Attempted speedups that were rejected because they either gain nothing or cost quality:

attempt result
text encoder on CPU (avoid reloads) crashes: 8.7 GB encoder + 4.6 GB model on 16 GB RAM
--disable-smart-memory still crashes on the 2nd prompt
batched CFG (batch_cfg=True) 139 s vs 140 s — GPU already saturated
torch.compile on the denoiser Triton cannot compile its driver shim on Python 3.15
ComfyUI --fast documented as "potentially quality deteriorating"
fewer steps / fewer Gaussians / 1024²→upscale quality loss

The floor is ~31 minutes of pure GPU work for six objects at 2K; the remaining time is process start and model loading. Four objects fit in under 30 minutes with no compromise.

Licences

  • Code and workflow in this repository: MIT (see LICENSE).
  • Sample images / 3D assets: Qwen Research License, non-commercial only, because they were generated with Qwen-Image-2.1-Turbo. See NOTICE — keep it with any redistribution.
  • Idol (speridlabs/idol, idol-sdk): Apache-2.0.
  • Model weights are not included here; download them from their publishers.