KawasakiAkasei/Hunyuan3D_2.1_Low_VRAM
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
108 Commits 2 Branches 0 Tags
474001da6b9da5d6293ed58f6d5f08ba44c4766a
Commit Graph

108 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
Akasei
474001da6b feat(rembg): switch background removal to bria-rmbg model 
Replace default u2net with bria-rmbg-2.0 for better quality.
BackgroundRemover now accepts model_name param (defaults to 'bria-rmbg').

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
 2026-03-16 22:14:21 +08:00
Akasei
76c36e53eb fix(gradio): fix OOM killer on second request in low_vram_mode 
Root cause: _ensure_i23d_worker() reloaded from disk via from_pretrained(),
which loads the ~7GB checkpoint into CPU RAM. If Python GC hadn't freed
previous del'd tensors yet, both old+new copies in RAM → OOM Killer.

Fix: hybrid strategy per model type:
  i23d (shape, ~7.25GB VRAM):
    .to('cpu') ↔ .to('cuda') — stays in RAM, no disk IO, fast switch
  tex_pipeline (texture, ~6.59GB VRAM):
    del + gc + empty_cache ↔ reload from HF cache — full VRAM release

Renamed helpers:
  _unload_i23d_worker()  → _offload_i23d_to_cpu()
  _ensure_i23d_worker()  → _restore_i23d_to_gpu()
  (tex helpers unchanged)

VRAM timeline per request in low_vram_mode:
  shape gen: i23d on GPU (7.25GB), tex unloaded
  → _offload_i23d_to_cpu(): i23d→RAM (0GB VRAM)
  → _ensure_tex_pipeline(): tex loads (6.59GB)
  texture gen: tex on GPU (6.59GB), i23d in RAM
  → _unload_tex_pipeline(): tex del'd (0GB VRAM)
  next request: _restore_i23d_to_gpu(): RAM→GPU (7.25GB)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
 2026-03-16 22:05:08 +08:00
Akasei
9bee8e1844 refactor(gradio): replace CPU offload with direct GPU unload/lazy-load 
Instead of .to('cpu') / .to('cuda'), models are now fully del'd from
GPU (no CPU intermediate) and reloaded on demand:

- _unload_i23d_worker(): del + gc.collect() + empty_cache()
- _ensure_i23d_worker(): lazy reload from pretrained if None
- _unload_tex_pipeline(): del + gc.collect() + empty_cache()
- _ensure_tex_pipeline(): lazy load from tex_conf if None

generation_all() flow in low_vram_mode:
  shape gen → _unload_i23d_worker → _ensure_tex_pipeline →
  texture gen → _unload_tex_pipeline
  (shape model reloads on next _gen_shape call via _ensure_i23d_worker)

Startup: tex_pipeline NOT loaded in low_vram_mode (only tex_conf stored),
reducing startup VRAM from ~13.5GB to ~7.25GB.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
 2026-03-16 21:15:56 +08:00
Akasei
5d0405dc68 feat(gradio): apply VRAM optimization and fix texture config 
- generation_all(): offload i23d_worker to CPU before texture gen,
  restore after — mirrors batch_generate.py sequential strategy.
  Prevents OOM when both models peak simultaneously on RTX 3080.
- Change texture config: max_num_view 8→9, resolution 768→512.
  768 resolution OOMs (14.6GB activation); 512 is practical max for
  RTX 3080 20GB. max_views 9 gives better texture coverage.
- Only active when --low_vram_mode flag is passed.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
 2026-03-16 21:05:14 +08:00
Akasei
e150058012 feat(batch): use steps=50, resolution=512, max_views=9 for RTX 3080 
768 resolution causes OOM (14.6GB model activation) on RTX 3080 20GB.
512 is the practical maximum: texture model uses 6.59GB, leaving
sufficient headroom. Increased max_views 6→9 for better texture coverage.

Result: 9/9 images → textured GLB in 12.3 min total.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
 2026-03-16 20:53:12 +08:00
Akasei
b6685c9560 feat: add batch 3D generation script with VRAM optimization 
- Add batch_generate.py: two-phase pipeline (shape→texture) that loads
  models sequentially to avoid OOM on RTX 3080
- Fix mesh_utils.py: make bpy import lazy so load_mesh/save_mesh work
  without Blender installed
- Phase 1: shape generation for all images, then unload
- Phase 2: texture generation for all meshes, then unload
- Skip already-generated outputs for resumability
- Tested: 9/9 images successfully generated textured GLB models

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
 2026-03-16 20:20:46 +08:00
Huiwen Shi
82920d643c Update LICENSE 
Update License
 2025-10-17 18:10:07 +08:00
HuiwenShi
c9b21668e2 Create is_watertight.py 2025-09-24 11:35:53 +08:00
HuiwenShi
5b6885dcf4 Update chamfer_distance.py 2025-09-23 14:10:26 +08:00
HuiwenShi
34746fcbc2 Create chamfer_distance.py 2025-09-23 11:46:01 +08:00
HuiwenShi
a0fd02ea01 Merge pull request #98 from qinmaohui/main 
【犀牛鸟实战issue】修复在windows系统中安装custom_rastorizer报错
 2025-09-11 22:38:31 +08:00
qinmaohui
663ee27446 还原对于custom_rasterizer_kernel的修改 2025-09-10 22:37:29 +08:00
qinmaohui
928f41b289 将原文件恢复，新建custom_rasterizer_kernel_for_windows文件夹放置修改的文件 2025-09-10 09:04:36 +08:00
Xianghui Yang
2eb92bcfd1 Merge pull request #104 from WncFht/feature/add-enable-flashvdm 
【犀牛鸟实战issue】inference speed
 2025-09-08 23:16:17 +08:00
HuiwenShi
06ea674535 Merge pull request #137 from ItsThatRandomDev/fix/docker-conda-tos 
Fix: accept Anaconda ToS in Dockerfile to prevent build failure
 2025-09-08 20:29:54 +08:00
Xianghui Yang
840d66abe8 Merge pull request #102 from s572915912/s572915912-patch-1 
【犀牛鸟实战issue】training: split_sizes error
 2025-09-08 19:57:36 +08:00
Xianghui Yang
7cc51b67ef Update README.md 
add acknowledgment
 2025-08-27 14:52:15 +08:00
Xianghui Yang
3efb87e736 Update README.md 
add acknowledgment
 2025-08-27 14:51:20 +08:00
ItsThatRandomDev
0e9f8d78d4 fix conda ToS acceptance in Dockerfile 2025-08-18 19:00:20 +02:00
s572915912
b3dd50ba37 Update misc.py 
repair
 2025-08-06 01:14:49 +08:00
s572915912
d9fc4d31bf Update hunyuandit-mini-overfitting-flowmatching-dinol518-bf16-lr1e4-4096.yaml 
repair
 2025-08-06 01:12:13 +08:00
Xianghui Yang
84e97834c0 Update README.md 2025-07-31 17:21:07 +08:00
Xianghui Yang
cd2feb5f38 Add files via upload 2025-07-30 23:21:51 +08:00
Xianghui Yang
b791e9c22f Update LICENSE 2025-07-30 23:18:42 +08:00
Xianghui Yang
b521b9f71f update x.png 2025-07-28 17:44:21 +08:00
Xianghui Yang
b96ea3558d update x qrcode 2025-07-28 17:40:27 +08:00
Xianghui Yang
b32d264d43 add hunyuan world 1.0 2025-07-28 17:36:11 +08:00
Xianghui Yang
665c38a19a add hunyuan world 1.0 2025-07-28 17:35:39 +08:00
oakshy
2ca5fd3155 modify model zoo 2025-07-15 11:28:24 +08:00
WncFht
a6509b95fb feat: 为 api_server 加上 enable_flashvdm 2025-07-13 11:48:13 +08:00
WncFht
00fa3ac012 feat: 为 gradio_app.py 加上 enable_flashvdm 2025-07-13 11:44:49 +08:00
s572915912
f4e0307665 Update train_deepspeed.sh 2025-07-11 18:32:16 +08:00
s572915912
8eff6d8233 Delete run_inference_with_fix.py 2025-07-11 16:54:43 +08:00
s572915912
7a9d765627 Update run_inference_with_fix.py 2025-07-11 16:53:19 +08:00
s572915912
f0a008279e Update pipelines.py 2025-07-11 16:51:33 +08:00
s572915912
dc2ea32d76 Update hunyuandit-mini-overfitting-flowmatching-dinol518-bf16-lr1e4-4096.yaml 2025-07-11 16:47:40 +08:00
s572915912
96349ad5d0 Update train_deepspeed.sh 2025-07-11 16:43:40 +08:00
s572915912
6726877bbb Update train_deepspeed.sh 2025-07-11 16:40:01 +08:00
s572915912
c6d4cb89e2 Update train_deepspeed.sh 2025-07-11 16:39:10 +08:00
s572915912
de7996251d Update hunyuandit-mini-overfitting-flowmatching-dinol518-bf16-lr1e4-4096.yaml 2025-07-11 16:37:32 +08:00
s572915912
af935af688 Update train_deepspeed.sh 2025-07-11 16:36:46 +08:00
s572915912
f2f19d74a8 Update hunyuandit-mini-overfitting-flowmatching-dinol518-bf16-lr1e4-4096.yaml 
add explain
 2025-07-11 15:53:01 +08:00
s572915912
8cd92830fb Update train_deepspeed.sh 
auto detect
 2025-07-11 15:51:55 +08:00
s572915912
e34a3ba752 Create run_inference_with_fix.py 2025-07-11 02:33:30 +08:00
s572915912
b06e6ddf37 Update pipelines.py 2025-07-11 02:29:25 +08:00
qinmaohui
48a2a2bab9 删除多余文件 2025-07-10 17:29:07 +08:00
qinmaohui
7facf4b459 修复在windows系统中安装custom_rastorizer报错 2025-07-10 17:22:47 +08:00
HuiwenShi
81b4d9f638 Update README.md 2025-07-09 16:01:54 +08:00
HuiwenShi
e8262b085f Update Dockerfile 2025-07-02 14:45:29 +08:00
HuiwenShi
70a0dc0183 Merge pull request #63 from perfectproducts/main 
added api_server
 2025-07-02 14:39:30 +08:00