christopherthompson81/inference-rs-worldgen

World-generation checkpoints for inference.rs

The checkpoints inference.rs runs to turn one 360 degree panorama into an explorable 3D world (HY-World 2.0's world generation: MoGe-2 depth, WorldStereo-2 views along camera paths, WorldMirror 2.0 reconstruction, Gaussian splats), packaged in 27 GB instead of 59 and laid out as the models root its server reads:

inference serve --world-models christopherthompson81/inference-rs-worldgen

The first world job fetches this repo into the Hugging Face cache. WorldMirror 2.0 is not here: jobs fetch it from tencent/HY-World-2.0 under its own license.

Folder	From	Changed
WorldStereo/worldstereo-memory-dmd/	hanshanxue/WorldStereo	the 40 main blocks' linear weights quantized to Q8_0 (the same bytes inference.rs makes when it quantizes the bf16 checkpoint at load); every other tensor bf16 as stored
Wan2.1-I2V-14B-480P-Diffusers/	Wan-AI/Wan2.1-I2V-14B-480P-Diffusers	the umT5-XXL text encoder as a Q8_0 GGUF (llama.cpp t5encoder names; norms and position biases F32); the VAE, CLIP image encoder, tokenizer and configs unchanged; the base transformer weights left out (WorldStereo-2 replaces them)
moge-2-vitl-normal/	Ruicheng/moge-2-vitl-normal	model.pt converted to model.safetensors, its model_config saved as config.json
lpips-vgg/	torchvision's VGG16 and the LPIPS v0.1 heads	the VGG16 features.* weights and lin0 to lin4 in one safetensors file

Quality

The WorldStereo-2 GGUF changes nothing against inference.rs's default run: its blocks are quantized from the same bf16 values to the same Q8_0 bytes.

The umT5 encoder is the one numerical change. Its GGUF weights are dequantized for each matmul (T5's activations carry outliers that 8-bit activation kernels lose whole tokens to). Measured on four prompts against the fp32 checkpoint, encoder states per token:

	worst token cosine	mean relative RMS error
bf16 (what the pipeline runs without this GGUF)	0.958	5.5%
Q8_0	0.983	5.1%

Licenses

Each folder keeps its source's license: Wan2.1 under Apache-2.0, MoGe-2 under MIT, VGG16 under torchvision's BSD-3-Clause and the LPIPS heads under BSD-2-Clause.

WorldStereo-2 is published at hanshanxue/WorldStereo (MIT on its model card) and is also the WorldStereo-2 of Tencent's HY-World 2.0 release, so WorldStereo/ is distributed under the Tencent HY-WORLD 2.0 Community License Agreement: see WorldStereo/License.txt and WorldStereo/NOTICE, which lists the modified file. That license does not apply in the European Union, the United Kingdom or South Korea, and its use restrictions (Sections 5(a) and 5(b), and the Acceptable Use Policy) bind every user of these weights. Tencent is not affiliated with, associated with, sponsoring or endorsing inference.rs or this repository.

Tencent HY-WORLD 2.0 is licensed under the Tencent HY-WORLD 2.0 Community License Agreement, Copyright © 2026 Tencent. All Rights Reserved. The trademark rights of "Tencent HY" are owned by Tencent or its affiliate.