Support the LingBot-Video MoE refiner (two-stage high-resolution refinement) - #1546

Draft

NancyFyong wants to merge 38 commits into

modelscope:mainfrom

NancyFyong:draft/lingbot-video-moe-refiner

Commits on Jul 25, 2026

Add LingBot-Video (Dense-1.3B) support with attention-only LoRA SFT

NancyFyong
and
claude
committed
182ea3a

Commits on Jul 26, 2026

Add docs, prompt rewriting, LoRA validation, and low-VRAM example for LingBot-Video

NancyFyong
and
claude
committed
67bcd02

Commits on Jul 27, 2026

Add LingBot-Video MoE-30B-A3B support

NancyFyong
and
claude
committed
55c8357
Address PR #1539 review: unify scheduler, restore VAE, split rewriter

NancyFyong
and
claude
committed
df7ea8e
Fix low-VRAM fp32/bf16 dtype mismatch in LingBot-Video DiT

NancyFyong
and
claude
committed
5fbf50a
Fold normalize_caption into lingbot_video pipeline module

NancyFyong
and
claude
committed
a9e7cc2
Add normalize_caption to lingbot_video and fix importers

NancyFyong
and
claude
committed
c62ea4f
Trim verbose comments in LingBot-Video code

NancyFyong
and
claude
committed
e83d7d6
Merge lingbot_sft into lingbot_moe: apply PR #1539 review fixes to the MoE branch

NancyFyong
and
claude
committed
038bad0
Fix dtype mismatch feeding the fp32-pinned time_embedder

NancyFyong
and
claude
committed
3333443
Add t2v_example_4 structured caption

NancyFyong
and
claude
committed
4ecc892
Merge lingbot_sft into lingbot_moe (dtype fix + sample 4 prompt)
NancyFyong
committed
0a107bb
Add TI2V (image-to-video) support to LingBot-Video pipeline

NancyFyong
and
claude
committed
ea73e8b
Add text-to-image (t2i) support and example to LingBot-Video

NancyFyong
and
claude
committed
670ec22

Commits on Jul 28, 2026

Add ti2v/t2i low-VRAM inference + ti2v LoRA training examples

NancyFyong
and
claude
committed
8d8b7c7
update vae&inner_func
mi804
committed
b2492dd
tmp commit for code refactor
mi804
committed
a10c026
low vram patch&refactor scripts
mi804
committed
94e8baa
remove fp32 cast to keep compatibility with low-vram-inference
mi804
committed
7caae83
add support for full train
mi804
committed
5f604e2
style refactor
mi804
committed
cae44a0
pop changes for ti2v pipeline
mi804
committed
531ba56
Align TI2V / T2I examples, LoRA + full training, docs and README with reviewer's refactor (531ba56)
NancyFyong
committed
4c9ee93
fix: manage Qwen3-VL vision tower in LingBotVideoTextEncoder VRAM map

NancyFyong
and
claude
committed
5fa49f5
style: group input_image under image-to-video in pipeline __call__

NancyFyong
and
claude
committed
56334bf
refactor: move TI2V helpers into the units that use them

NancyFyong
and
claude
committed
7d0852e
Merge lingbot_sft into lingbot_moe: integrate TI2V/T2I + full training, keep MoE-30B

NancyFyong
and
claude
committed
91fcb94

Commits on Jul 29, 2026

Commits on Jul 30, 2026