Introduction

SEELE AI has released open weights for EVA01 Unified All-Task Generation, a native 3D model that turns a text prompt or a single image into a game-style mesh you can export as GLB. The 619 MB generation adapter and a matching LoRA bundle went up on Hugging Face at SEELE-AI/EVA01-Unified-All-Task-Generation and SEELE-AI/EVA01-Unified-All-Task-Generation-LoRA at about 15:50 MYT on October 30, alongside the "Prepare OpenEVA inference release" commit in the Apache-2.0 SeeleAI/OpenEVA repo. You run it locally on a CUDA GPU today; there is no hosted demo for the generation side yet.

SEELE builds AI tools for making games, and EVA01 is its bet that a model should read and write meshes directly instead of guessing 3D from 2D pictures. The same family already shipped a mesh-understanding model in June. This drop is the half that makes assets.

What shipped

  • Model: EVA01 Unified All-Task, generation adapter plus LoRA export
  • Vendor: SEELE AI (seeles.ai)
  • Modality: text-to-3D and image-to-3D, output as latents plus an optional GLB mesh
  • Open vs API: open. The OpenEVA code and the adapter card are Apache-2.0; the frozen bases it plugs into keep their own licenses
  • Where to run it today: your own CUDA machine through generate.py in OpenEVA. No fal, Replicate or Comfy node yet
  • Concrete fact: one Qwen3-VL condition drives a three-stage SS → Shape-512 → Material-512 cascade, with 1.3B shape and material flow models borrowed from TRELLIS.2

How it works

EVA01 is a Mixture-of-Transformers. An understanding expert and a mirrored generation expert share attention, so the same model that can describe a mesh can also build one. The technical report came out in May; the weights to actually generate are what is new.

Generation runs in three steps. A sparse-structure stage blocks out the occupied volume, a shape stage fills in geometry at 512 resolution, and a material stage paints it. Qwen3-VL-2B-Instruct reads your prompt or image and conditions all three stages at once. The adapter holds only EVA01's own generation parameters, so you also need Qwen3-VL, the TRELLIS sparse-structure decoder and the TRELLIS.2 shape and material models and decoders. That is why the download is only 619 MB.

The gallery in the repo sticks to game-ready subjects: a teal and beige mecha with a multi-barrel gun, a stone castle with conical towers and a drawbridge, and image-to-3D runs on a mecha, a pickup truck and an electric guitar.

Run it

The install script expects a clean Ubuntu-style box. You need PyTorch 2.6 or newer, Transformers 4.57 or newer, flash-attn and a TRELLIS-compatible flex_gemm built against the same CUDA.

git clone https://github.com/SeeleAI/OpenEVA && cd OpenEVA/EVA01
bash install.sh
source .venv/bin/activate

python generate.py \
  --prompt "a small wooden treasure chest with iron bands and a gold lock, stylized game prop" \
  --checkpoint /path/to/eva01-generation-adapter \
  --tokenizer-root /path/to/eva01-unified-tokenizer \
  --qwen-root /path/to/Qwen3-VL-2B-Instruct \
  --ss-xl-stem /path/to/ss_flow_txt_dit_XL_16l8_fp16 \
  --shape-stem /path/to/slat_flow_img2shape_dit_1_3B_512_bf16 \
  --material-stem /path/to/slat_flow_imgshape2tex_dit_1_3B_512_bf16 \
  --normalization-config /path/to/TRELLIS.2-4B/pipeline.json \
  --ss-decoder-stem /path/to/TRELLIS-image-large/ckpts/ss_dec_conv3d_16l8_fp16 \
  --shape-decoder-stem /path/to/TRELLIS.2-4B/ckpts/shape_dec_next_dc_f16c32_fp16 \
  --material-decoder-stem /path/to/TRELLIS.2-4B/ckpts/tex_dec_next_dc_f16c32_fp16 \
  --output outputs/chest.npz \
  --output-glb outputs/chest.glb

Swap --prompt for --image concept.png to lift a piece of concept art. Prompts in the gallery read like an asset brief: object, materials, colors, then the parts you care about. Write them that way.

What to know before you build on it

The GLB you get has decoded geometry with vertex colors, not baked UV textures or PBR maps. Plan on a retopo and texture pass before it goes into Unity, Unreal or Godot. The writer also skips topology repair and remeshing on purpose, so expect raw output.

SEELE does not publish VRAM numbers or timings for generation. With Qwen3-VL-2B and two 1.3B flow models in the loop, test on your own card before you promise anyone a pipeline. The repo's news line says the generation runtime opened October 20, but the release commit and both weight uploads are dated October 30.

Training code, the dataset and an editing branch are all marked as planned. Editing is the one to watch: a model that understands a mesh and can change it on request is exactly what an asset pipeline wants.

What to watch

  • A ComfyUI node or Gradio Space for generation, which would put this in front of far more people
  • VRAM and speed numbers from the first community runs
  • The editing checkpoints SEELE lists on its roadmap

—Titus