Ltx 2.5 - #14447

Merged

yiyixuxu merged 88 commits into

mainfrom

Aug 11, 2026

Commits on Jul 28, 2026

Add LTX-2.4 support to the LTX2 transformer, pipelines, and conversion script
alexanderar
committed
78fcaf8
Fix resolution-dependent timestep shift being a no-op in all LTX2 pipelines
alexanderar
committed
420ad55

Commits on Jul 29, 2026

Add LTX-2.4 duration head for automatic num_frames prediction
alexanderar
committed
2bc3c93

Commits on Jul 30, 2026

Add LTX-2 modular video and audio decoder blocks

dg845
and
claude
committed
963f1e4
Add LTX-2 modular pipeline (t2v + i2v) and parity harnesses

dg845
and
claude
committed
6ca2c69
Add script to create tiny LTX-2.4 test pipeline ckpt (under assumption that it inherits all configs from LTX-2.3)
dg845
committed
d39c718
make fix-copies
dg845
committed
d6bbed1
Run modular connector once on CFG-concatenated batch for bitwise parity

dg845
and
claude
committed
f8bf0fa
Use tiny LTX2VocoderWithBWE in the LTX-2.4 test checkpoint

dg845
and
claude
committed
181861e
Match LTX-2.4 transformer flags in the test checkpoint

dg845
and
claude
committed
d0335ba
Add CPU offload flag to parity scripts to be able to load full checkpoints
dg845
committed
94cf57b
Support dynamic timestep shift and duration head in LTX-2 modular pipeline

dg845
and
claude
committed
ee19558
Add duration predictor support for tiny checkpoint and parity scripts
dg845
committed
4111b09
Merge branch 'main' into ltx-2-4-modular
dg845
committed
d4085b3
Refactor LTX-2 modular guidance into video/audio guiders

dg845
and
claude
committed
3582e4a
Batch cond+uncond CFG forward to mirror standard LTX-2 pipeline

dg845
and
claude
committed
932d824

Commits on Jul 31, 2026

Drive LTX-2 modular guidance through the guider API (per-pass inputs)

dg845
and
claude
committed
c372297

Commits on Aug 1, 2026

Drive LTX-2 modular guidance via prepare_inputs_from_block_state

dg845
and
claude
committed
fefc2b7

Commits on Aug 2, 2026

Route LTX-2 batch-invariant denoiser kwargs via denoiser_input_fields

dg845
and
claude
committed
7482c46
Fix LTX-2 parity harnesses: guidance overrides reach both pipelines

dg845
and
claude
committed
e2cb6f5

Commits on Aug 4, 2026

Document denoiser_input_fields standalone-run caveat for LTX-2 denoiser

dg845
and
claude
committed
7e11e36
Address LTX-2 modular self-review: fix dead enhancer trigger + docstrings

dg845
and
claude
committed
337d0bf
Fix stale I2V parity-harness --help text for the loosened fp32 gate

dg845
and
claude
committed
2aa7eca
make style and make quality
dg845
committed
3a73088
Add the LTX-2.4 diffusion VAE decoder with a native-parity harness
linoytsaban
committed
ec0b841
Add a converter for the LTX-2.4 diffusion VAE decoder
linoytsaban
committed
acf5c7d
Initialise use_slicing and document the diffusion VAE decisions
linoytsaban
committed
255fc8d
Record real-weight parity against the reference decoder
linoytsaban
committed
5fac7aa
Add model-level tests for the LTX-2.4 diffusion VAE decoder
linoytsaban
committed
926818a
Fold the diffusion VAE converter into convert_ltx2_to_diffusers.py
linoytsaban
committed
887c186
Make the diffusion decoder work through LTX2Pipeline
linoytsaban
committed
3928fc2
Make the diffusion decoder work through the modular decode step
linoytsaban
committed
c6b8c4a
Note that the pipeline decode block predates the LTX-2.4 integration PR
linoytsaban
committed
fbd8733
Address diffusers self-review on the diffusion VAE decoder
linoytsaban
committed
dbd6acf
Add the API doc page for AutoencoderKLLTX2VideoDiffusionDecoder
linoytsaban
committed
2303e01
Mark the parity harness transient and drop the notes doc from the diff
linoytsaban
committed
7ee18bb
Accept the diffusion decoder in the image-to-video pipeline too
linoytsaban
committed
558d8e4

Commits on Aug 5, 2026

Reorganize LTX-2 modular blocks into coarser component-aligned phases

dg845
and
claude
committed
2b6106e
Enable LTX-2 prompt enhancement by default when the enhancer is loaded

dg845
and
claude
committed
661a6e4

Commits on Aug 6, 2026

Commits on Aug 7, 2026

Merge branch 'pr-2-ltx-2-4-integration' into ltx-2-4-modular
dg845
committed
b24e69f
Align LTX-2 modular blocks with the merged standard-pipeline changes

dg845
and
claude
committed
5d18b5d
Evaluate the decoder's SwiGLU in token tiles
linoytsaban
committed
8e5df58
Declare _no_split_modules on the diffusion decoder
linoytsaban
committed
031b562
Match the encoder defaults to the published LTX-2.4 config
linoytsaban
committed
a68d6d2
Cover the token-tiled SwiGLU path in the tests
linoytsaban
committed
68755f0
Let the parity harness run against the LTX-2.5 reference too
linoytsaban
committed
c984a00

Commits on Aug 9, 2026

Add LTX-2.X modular condition blocks

dg845
and
claude
committed
e117b22
Add LTX-2.X modular in-context (IC-LoRA) blocks

dg845
and
claude
committed
58052bf
Make the diffusion decoder a decode pipeline instead of a VAE
linoytsaban
committed
d71af8c
Add a modular decode block for the LTX-2 diffusion decoder
linoytsaban
committed
42f4510
Test the diffusion decode pipeline, and give it a video-only output
linoytsaban
committed
474452d
Fix the decode example: output_type="latent" is already denormalized
linoytsaban
committed
8e978fd

Commits on Aug 10, 2026

Fix IC-LoRA video self-attention mask grouping

dg845
and
claude
committed
dd92e46
Add LTX25AutoBlocks and LTX25ModularPipeline
linoytsaban
committed
0a14715
Drop the encoder half in the diffusion decoder converter
linoytsaban
committed
1d6fdce
Share the neighborhood attention mask and parameterise the AdaLN chunk count
linoytsaban
committed
d97dcac
Export the LTX-2.5 modular classes from the top-level namespace
linoytsaban
committed
d354288
Name the diffusion decoder LTX-2.5 and save it to diffusion_decoder/
linoytsaban
committed
31414b8

Commits on Aug 11, 2026

Move the LTX-2 text-input expansion into the core denoise step

dg845
and
claude
committed
24404d1
Describe the connector inputs on LTX2TextInputStep

dg845
and
claude
committed
1f0e2e2
Merge branch 'ltx-2-4-modular' into ltx-2-4-modular-condition-ic
dg845
committed
c6bd5a6
Build the condition and in-context workflows from LTX2AutoBlocks

dg845
and
claude
committed
5bb24c6
Run the text-input step in the condition and in-context denoise steps

dg845
and
claude
committed
47cf758
Broadcast keyframe and reference conditioning across the generation batch

dg845
and
claude
committed
4f7a060
Merge pull request #12 from huggingface/ltx-2-4-modular-condition-ic
dg845
authored
63fdd15
Merge remote-tracking branch 'origin/ltx-2-4-modular' into ltx-2.5
DN6
committed
33b3c06
Merge branch 'ltx-2-4-modular' into ltx-2-4-diffusion-vae-modular
linoytsaban
committed
7a73ce6
feat: use natten for processing. (#15)
sayakpaul
authored
6593e3b
Merge PR #8 (LTX-2.4 diffusion VAE decoder) into ltx-2.5
DN6
committed
64839af
Merge PR #8 update: fetch NATTEN as a Hub kernel
DN6
committed
5b6f9de
make vae properties accessible without loading it. (#17)
sayakpaul
authored
4d32963
Address review: replace LTX2 mixins with Copied-from pipeline methods.

alexanderar
authored and
DN6
committed
50bc9f8
Point the modular enhance-prompt comment at LTX2Pipeline
DN6
committed
6e92e15
Merge PR #8 update: read latent statistics without loading the vae
DN6
committed
f512cbd
Remove the integrations/ dev scripts
DN6
committed
bded369
moular tests (#19)
sayakpaul
authored
7ee1831
Keep LTX-2.5.1 keyframes_abs_pos_embedding through convert. (#20)
alexanderar
authored
f20c000
remove test_ltx2_keyframes_abs_pos_embedding_load_and_forward
sayakpaul
committed
a6803d6
Update the LTX-2.5 docs: two stage generation, distilled default, repo id (#21)
linoytsaban
authored
a1c06c9
feat: implement tiling for the diffusion decoder. (#22)
sayakpaul
authored
c8e16b9
add a note on tiling (#23)
sayakpaul
authored
68c743b
Merge branch 'main' into ltx-2.5
sayakpaul
authored
fb8f932
fix documentation.
sayakpaul
committed
83cca54