Skip to content

New Model: LTX 2.5 - #1694

Open
dxqb wants to merge 3 commits into
Nerogar:masterfrom
dxqb:ltx-noadapter-squashed
Open

New Model: LTX 2.5#1694
dxqb wants to merge 3 commits into
Nerogar:masterfrom
dxqb:ltx-noadapter-squashed

Conversation

@dxqb

@dxqb dxqb commented Aug 15, 2026

Copy link
Copy Markdown
Collaborator

Summary

Contains many other PRs as base
Actual diff of this PR is e4423b5

Test plan

  • pre-commit run --all-files passes
  • Launched the affected UI or script and exercised the change
  • Tested with at least one real preset / config when relevant (note which: ____)

AI assistance

  • AI-assisted — I have read every line in this diff and can defend each change

dxqb and others added 2 commits August 15, 2026 19:09
@dxqb @claude
Dependency snapshot for ltx-noadapter
63fb576 
Carries the branches this feature is stacked on, each already under review as
its own PR off master: streaming, streaming-simplex, nonreentrant-offload,
kernel-lora-fusion, staged-sampling, chunked-weight-quantization,
quantization-fallback-dtype, profile-steps-env, fix-sample-window-compile-init,
sample-shift-override, and train-remote-traceback. The only change unique to
this commit reconciles streaming's is_quantized rename against
kernel-lora-fusion's forward_with_lora, a semantic overlap git merged cleanly.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@dxqb @claude
LTX 2 / LTX 2.5 without linear-adapter
e4423b5 
Adds ModelType.LTX_2 with its sampler, fine-tune and LoRA setups, model savers,
and an LTX 2.5 LoRA preset. Its model parts are transformer, text_encoder,
connectors, low_noise_transformer and vae; both transformers load from a
diffusers repo, while a single-file low-noise expert loads whole via
from_single_file. LTX_2 joins the dynamic-timestep-shift set and is registered
as a video model (its audio branch is frozen and never decoded, so it stays out
of the audio path for now).

Pin bumps ride along: transformers 5.14.1, huggingface-hub 1.27.0, and diffusers
175fe6b. mgds is pointed at dxqb/mgds-private@ltx2 for EncodeGemma3Text, which is
not yet upstream. The transformers pin also leaves the CLIP-based models (SD1.x,
SDXL, Flux, HunyuanVideo, Wuerstchen) broken until the flattening migration of
Nerogar#1506, so this is not upstreamable as it stands.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@dxqb

dxqb commented Aug 15, 2026

Copy link
Copy Markdown
Collaborator Author

Contains:

  ┌─────────────────┬─────────────────────────────────────────────────────────────────────────┐
  │       PR        │                                  Title                                  │
  ├─────────────────┼─────────────────────────────────────────────────────────────────────────┤
  │ #1642           │ Streaming model loader                                                  │
  ├─────────────────┼─────────────────────────────────────────────────────────────────────────┤
  │ #1644           │ Simplex offloading (2x speed)                                           │
  ├─────────────────┼─────────────────────────────────────────────────────────────────────────┤
  │ #1649           │ Non-reentrant offloading, fix race condition, fix activation offloading │
  ├─────────────────┼─────────────────────────────────────────────────────────────────────────┤
  │ #1659           │ Chunk weight quantization to bound fp32 VRAM spikes                     │
  ├─────────────────┼─────────────────────────────────────────────────────────────────────────┤
  │ #1680           │ nunchaku-style kernel fusion — 20-30% faster training                   │
  ├─────────────────┼─────────────────────────────────────────────────────────────────────────┤
  │ #1688           │ Staged sampling                                                         │
  ├─────────────────┼─────────────────────────────────────────────────────────────────────────┤
  │ #1629           │ fallback dtype for quantization layer filter                            │
  ├─────────────────┼─────────────────────────────────────────────────────────────────────────┤
  │ #1687           │ Raise the recompile limit in the sample window's thread                 │
  ├─────────────────┼─────────────────────────────────────────────────────────────────────────┤
  │ #1692           │ field to override the timestep shift for sampling                       │
  ├─────────────────┼─────────────────────────────────────────────────────────────────────────┤
  │ #1691           │ Cloud: Print the traceback before the final save in case of error       │
  ├─────────────────┼─────────────────────────────────────────────────────────────────────────┤
  │ #1628           │ Quiet non-actionable startup warnings & status in progress bar          │
  ├─────────────────┼─────────────────────────────────────────────────────────────────────────┤
  │ #1682 ✅ merged │ OT_DEBUG_PROFILES to enable the torch profiler / memory snapshots       │
  └─────────────────┴─────────────────────────────────────────────────────────────────────────┘

@dxqb dxqb mentioned this pull request Aug 15, 2026
Open
@dxqb dxqb linked an issue Aug 15, 2026 that may be closed by this pull request
[Feat]: LTX 2.3 #1636
Open
@dxqb @claude
Pin MGDS to the published ltx2-squashed commit
79ee1a8 
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@dxqb dxqb mentioned this pull request Aug 16, 2026
Closed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

[Feat]: LTX 2.3

1 participant

@dxqb