--- title: facelift_convrot_int8_int4_runtimes canonical_url: "https://www.modelscope.cn/models/ApacheOne/facelift_convrot_int8_int4_runtimes" md_url: "https://www.modelscope.cn/models/ApacheOne/facelift_convrot_int8_int4_runtimes.md" repository: ApacheOne/facelift_convrot_int8_int4_runtimes last_updated: 2026-08-13 parameters: 7.5B tensor_type: - F32 - I8 - F16 library_name: - safetensors - pytorch frameworks: - pytorch downloads: 48 stars: 0 tags: - facelift - openfacelift - 3d-face-reconstruction - gaussian-splatting - quantization - int8 - int4 - w4a4 - w4a8 - convrot - comfy-kitchen - xformers --- # facelift_convrot_int8_int4_runtimes > facelift_convrot_int8_int4_runtimes - ApacheOne 在 ModelScope 开源的模型。FaceLift ConvRot INT8 / INT4 Runtimes Quantized ConvRot runtimes and transformed checkpoints for FaceLift / OpenFaceLift, tested with native Comfy-Kitchen ConvRot kernels. Four tested… - **Repository**: ApacheOne/facelift_convrot_int8_int4_runtimes - **Parameters**: 7.5B - **Tags**: facelift, openfacelift, 3d-face-reconstruction, gaussian-splatting, quantization, int8, int4, w4a4, w4a8, convrot, comfy-kitchen, xformers - **Downloads**: 48 - **Stars**: 0 - **Last updated**: 2026-08-13 Source: https://www.modelscope.cn/models/ApacheOne/facelift_convrot_int8_int4_runtimes --- # FaceLift ConvRot INT8 / INT4 Runtimes Quantized ConvRot runtimes and transformed checkpoints for **FaceLift / OpenFaceLift**, tested with native **Comfy-Kitchen** ConvRot kernels. Four tested configurations are included, ranging from full INT8 to mixed INT4 / INT8 setups. # FP16 reference ```text Runtime 28.204 s Peak VRAM 7.828 GiB Final SSIM 1.0000 ``` # Quantized configurations | Method | Checkpoint size | Peak VRAM | Runtime | Multiview SSIM vs FP16 | Final SSIM vs FP16 | |---|---:|---:|---:|---:|---:| | **Full INT8 W8A8** | **2.807 GiB** | **5.845 GiB** | 28.044 s | 0.9973 | 0.9925 | | **INT4 diffusion + INT8 GSLRM** | **2.235 GiB** | **5.268 GiB** | **25.780 s** | 0.9021 | 0.9036 | | **INT4 diffusion + W4A8 GSLRM body + INT8 heads** | **2.094 GiB** | **5.130 GiB** | 26.089 s | 0.9021 | 0.9024 | | **INT8 diffusion + W4A8 GSLRM body + INT8 heads** | **2.666 GiB** | **5.700 GiB** | 28.139 s | 0.9973 | 0.9847 | # Official FaceLift examples — turntable videos All eight official FaceLift examples (`000`–`007`) rendered with all four quantization methods. The four videos in each row use the same input face and the same inference settings. ### Example 000
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
### Example 001
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
### Example 002
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
### Example 003
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
### Example 004
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
### Example 005
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
### Example 006
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
### Example 007
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
--- # Official FaceLift examples — multiview outputs The corresponding six-view MVDiffusion outputs for the same eight official FaceLift examples. Click any image to open the full-resolution PNG. ### Example 000
Full INT8 W8A8 example 000 multiview INT4 Diff + INT8 GSLRM example 000 multiview INT4 Diff + W4A8 GSLRM example 000 multiview INT8 Diff + W4A8 GSLRM example 000 multiview
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
### Example 001
Full INT8 W8A8 example 001 multiview INT4 Diff + INT8 GSLRM example 001 multiview INT4 Diff + W4A8 GSLRM example 001 multiview INT8 Diff + W4A8 GSLRM example 001 multiview
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
### Example 002
Full INT8 W8A8 example 002 multiview INT4 Diff + INT8 GSLRM example 002 multiview INT4 Diff + W4A8 GSLRM example 002 multiview INT8 Diff + W4A8 GSLRM example 002 multiview
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
### Example 003
Full INT8 W8A8 example 003 multiview INT4 Diff + INT8 GSLRM example 003 multiview INT4 Diff + W4A8 GSLRM example 003 multiview INT8 Diff + W4A8 GSLRM example 003 multiview
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
### Example 004
Full INT8 W8A8 example 004 multiview INT4 Diff + INT8 GSLRM example 004 multiview INT4 Diff + W4A8 GSLRM example 004 multiview INT8 Diff + W4A8 GSLRM example 004 multiview
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
### Example 005
Full INT8 W8A8 example 005 multiview INT4 Diff + INT8 GSLRM example 005 multiview INT4 Diff + W4A8 GSLRM example 005 multiview INT8 Diff + W4A8 GSLRM example 005 multiview
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
### Example 006
Full INT8 W8A8 example 006 multiview INT4 Diff + INT8 GSLRM example 006 multiview INT4 Diff + W4A8 GSLRM example 006 multiview INT8 Diff + W4A8 GSLRM example 006 multiview
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
### Example 007
Full INT8 W8A8 example 007 multiview INT4 Diff + INT8 GSLRM example 007 multiview INT4 Diff + W4A8 GSLRM example 007 multiview INT8 Diff + W4A8 GSLRM example 007 multiview
Full INT8 W8A8
INT4 Diff + INT8 GSLRM
INT4 Diff + W4A8 GSLRM
INT8 Diff + W4A8 GSLRM
--- # Upstream projects - **FaceLift code:** https://github.com/weijielyu/FaceLift - **Official OpenFaceLift weights:** https://huggingface.co/wlyu/OpenFaceLift - **FaceLift paper:** https://arxiv.org/abs/2412.17812 - **Comfy-Kitchen:** https://github.com/Comfy-Org/comfy-kitchen --- # Benchmark environment All benchmark results below used the same controlled FaceLift runtime. ```text GPU NVIDIA L4 GPU capability SM 8.9 Python 3.12.13 PyTorch 2.11.0+cu128 Torchvision 0.26.0+cu128 CUDA 12.8 comfy-kitchen 0.2.31 xformers 0.0.35 Benchmark image 004.png Seed 4 Guidance 3.0 Diffusion steps 50 ``` FaceLift multiview attention used: ```text XFormersMVAttnProcessor ``` GSLRM used PyTorch SDPA across all 24 transformer blocks. The SSIM/PSNR figures below are from the controlled `004.png` benchmark. The eight-example gallery above is a broader qualitative comparison across all official FaceLift examples. --- # 1. Full INT8 W8A8 Folder: ```text 01_full_int8_w8a8/ ``` Configuration: ```text CLIP Vision INT8 W8A8 CLIP Text INT8 W8A8 UNet INT8 W8A8 VAE INT8 W8A8 GSLRM INT8 W8A8 ``` Quantized Linear layers: ```text CLIP Vision 193 CLIP Text 138 UNet 218 VAE 8 GSLRM 99 ------------------ Total 656 ``` Controlled benchmark: ```text Runtime 28.044 s Peak VRAM 5.845 GiB Multiview: SSIM 0.997264 PSNR 46.621 dB MAE 0.4247 Final render: SSIM 0.992526 ``` This is the highest-quality fully quantized configuration tested. Official eight-example suite: ```text Total runtime 256.517 s Mean / face 32.065 s Examples completed 8 / 8 ``` --- # 2. INT4 diffusion + INT8 GSLRM Folder: ```text 02_int4_diffusion_int8_gslrm/ ``` Configuration: ```text CLIP Vision W4A4 CLIP Text W4A4 UNet W4A4 VAE W4A4 GSLRM INT8 W8A8 ``` Controlled benchmark: ```text Runtime 25.780 s Peak VRAM 5.268 GiB Multiview SSIM 0.902111 Final SSIM 0.903606 ``` Official eight-example suite: ```text Total runtime 236.716 s Mean / face 29.589 s Examples completed 8 / 8 ``` This test showed that the diffusion side can run at W4A4, but the multiview quality loss becomes the dominant bottleneck. --- # 3. INT4 diffusion + W4A8 GSLRM body + INT8 heads Folder: ```text 03_int4_diffusion_w4a8_gslrm_int8_heads/ ``` Configuration: ```text CLIP Vision W4A4 CLIP Text W4A4 UNet W4A4 VAE W4A4 GSLRM body W4A8 gaussian_upsampler.linear INT8 W8A8 pixel_gaussian_decoder.1 INT8 W8A8 ``` GSLRM split: ```text 97 Linear layers W4A8 2 Linear layers INT8 W8A8 ``` Controlled benchmark: ```text Runtime 26.089 s Peak VRAM 5.130 GiB Multiview SSIM 0.902111 Final SSIM 0.902423 ``` Official eight-example suite: ```text Total runtime 238.067 s Mean / face 29.758 s Examples completed 8 / 8 ``` This configuration keeps the GSLRM body at packed 4-bit weights while avoiding A4 activation execution inside GSLRM. --- # 4. INT8 diffusion + W4A8 GSLRM body + INT8 heads Folder: ```text 04_int8_diffusion_w4a8_gslrm_int8_heads/ ``` Configuration: ```text CLIP Vision INT8 W8A8 CLIP Text INT8 W8A8 UNet INT8 W8A8 VAE INT8 W8A8 GSLRM body W4A8 gaussian_upsampler.linear INT8 W8A8 pixel_gaussian_decoder.1 INT8 W8A8 ``` GSLRM split: ```text 97 Linear layers W4A8 2 Linear layers INT8 W8A8 ``` Controlled benchmark: ```text Runtime 28.139 s Peak VRAM 5.700 GiB Multiview: SSIM 0.997264 PSNR 46.621 dB MAE 0.4247 Final render: SSIM 0.984692 PSNR 38.961 dB MAE 1.2562 ``` Compared directly with full INT8: ```text Multiview: MSE 0.0 SSIM 1.0 Final: SSIM 0.986710 PSNR 40.005 dB MAE 1.1138 ``` Official eight-example suite: ```text Total runtime 257.027 s Mean / face 32.128 s Examples completed 8 / 8 ``` This is the best mixed INT8 / INT4 configuration tested so far. --- # Main findings ## INT8 ConvRot is highly accurate Full INT8 retained almost all FP16 quality: ```text FP16 vs INT8 final SSIM = 0.9925 ``` Peak VRAM dropped from: ```text 7.828 GiB -> 5.845 GiB ``` --- ## GSLRM is highly sensitive to A4 activations Running the GSLRM transformer body with A4 activations caused a large reconstruction-quality loss. Moving GSLRM from: ```text W4A4 -> W4A8 ``` recovered most of the lost reconstruction quality while retaining packed 4-bit weights. --- ## The Gaussian output heads benefit from INT8 The two protected projections are: ```text gaussian_upsampler.linear pixel_gaussian_decoder.1 ``` These directly produce the Gaussian reconstruction parameters. Using INT8 for these layers improves the W4A8 GSLRM configuration. --- ## Diffusion becomes the limiting stage at W4A4 With INT4 diffusion: ```text Multiview SSIM ~= 0.9021 ``` Once information is lost during multiview generation, the downstream GSLRM cannot recover it. Switching the diffusion stack back to INT8 restores: ```text Multiview SSIM = 0.9973 ``` and raises final reconstruction quality to: ```text Final SSIM = 0.9847 ``` --- # Recommended configurations ## Highest quality ```text Full INT8 W8A8 ``` ```text Peak VRAM 5.845 GiB Final SSIM 0.9925 ``` ## Best mixed quality / memory ```text INT8 diffusion + W4A8 GSLRM body + INT8 Gaussian heads ``` ```text Peak VRAM 5.700 GiB Final SSIM 0.9847 ``` ## Lowest-memory mixed configuration tested ```text INT4 diffusion + W4A8 GSLRM body + INT8 Gaussian heads ``` ```text Peak VRAM 5.130 GiB Final SSIM 0.9024 ``` --- # Repository structure ```text facelift_convrot_int8_int4_runtimes/ │ ├── README.md ├── LICENSE ├── build_report.json │ ├── 01_full_int8_w8a8/ │ ├── weights/ │ ├── runtime/ │ ├── outputs/004/ │ ├── examples/official_facelift/ │ │ ├── 000/ │ │ ├── 001/ │ │ ├── ... │ │ └── 007/ │ ├── manifest.json │ └── metrics.json │ ├── 02_int4_diffusion_int8_gslrm/ ├── 03_int4_diffusion_w4a8_gslrm_int8_heads/ └── 04_int8_diffusion_w4a8_gslrm_int8_heads/ ``` Each official example folder contains: ```text input.png multiview.png output.png gaussians.ply turntable.mp4 ``` --- # Runtime Each method contains: ```text runtime/ ├── facelift_convrot_runtime.py ├── run_inference.py └── requirements.txt ``` Example: ```bash python runtime/run_inference.py \ --input_dir /path/to/images \ --output_dir ./outputs_new ``` If OpenFaceLift is already downloaded locally: ```bash export OPENFACELIFT_DIR=/path/to/OpenFaceLift ``` The runtime reconstructs the official model architecture, replaces eligible `nn.Linear` modules according to `manifest.json`, loads the transformed ConvRot checkpoints, enables FaceLift's xFormers multiview attention path, and runs the official inference pipeline. --- # ConvRot quantization details Only eligible rank-2: ```text nn.Linear ``` weights are ConvRot quantized. Convolutions, embeddings, normalization layers, biases, and unsupported tensors remain in their normal precision unless otherwise handled by the upstream model. ConvRot group selection: ```text 256 preferred 64 fallback 16 fallback ``` W4 quantization uses: ```text quant_group_size = 64 ``` ## W4A4 ```text packed W4 weights A4 execution linear_dtype="int4" ``` ## W4A8 ```text packed W4 weights A8 execution linear_dtype="int8" ``` ## INT8 W8A8 ```text INT8 ConvRot weights INT8 activation execution ``` --- # License These transformed checkpoints are derived from the official **OpenFaceLift** model weights. The official OpenFaceLift weights are distributed under the **Adobe Research License**. Review the original license before redistribution or use: https://huggingface.co/wlyu/OpenFaceLift/blob/main/LICENSE FaceLift source code is distributed separately under the licensing terms of the official code repository. This repository is an experimental quantization/runtime derivative and is **not** the official FaceLift release. --- # References 1. **FaceLift — official code** https://github.com/weijielyu/FaceLift 2. **OpenFaceLift — official weights** https://huggingface.co/wlyu/OpenFaceLift 3. **FaceLift paper** https://arxiv.org/abs/2412.17812 4. **Comfy-Kitchen** https://github.com/Comfy-Org/comfy-kitchen --- # Citation If you use FaceLift, cite the original FaceLift work: ```bibtex @InProceedings{FaceLift, author = {Lyu, Weijie and Zhou, Yi and Yang, Ming-Hsuan and Shu, Zhixin}, title = {FaceLift: Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads}, booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)}, month = {October}, year = {2025}, pages = {12691-12701} } ```