[diffusion] model: support LTX2.3 two stage by mickqian · Pull Request #22182 · sgl-project/sglang

mickqian · 2026-04-06T08:34:39Z

sglang generate \
  --model-path Lightricks/LTX-2.3 \
  --width 1536 --height 1024 \
  --prompt "A beautiful sunset over the ocean" \
  --pipeline-class-name LTX2TwoStagePipeline \
  --save-output

sglang:

sglang_two_stage.mp4

diffusers:

official_two_stage.mp4

difference:

mean_abs = 6.8681
PSNR = 26.7860
cosine = 0.992799

sglang (sp):

sglang_two_stage.mp4

Motivation

Modifications

Accuracy Tests

Speed Tests and Profiling

Checklist

Format your code according to the Format code with pre-commit.
Add unit tests according to the Run and add unit tests.
Update documentation according to Write documentations.
Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
Follow the SGLang code style guidance.

Review and Merge Process

Ping Merge Oncalls to start the process. See the PR Merge Process.
Get approvals from CODEOWNERS and other reviewers.
Trigger CI tests with comments or contact authorized users to do so.
- Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

gemini-code-assist

Code Review

This pull request implements support for the LTX-2.3 model, including the two-stage pipeline, updated sampling parameters for resolution scaling, and model overlay materialization. Changes also include refinements to the denoising and latent preparation stages, along with new alignment scripts and unit tests. Review feedback identifies opportunities to improve artifact resolution by prioritizing newer versions (22b and 1.1) and suggests adding a safety check for zero-division when calculating the cross-attention gate factor.

gemini-code-assist · 2026-04-06T08:39:00Z

+LTX23_DEV_CHECKPOINT_FILENAMES = (
+    "ltx-2.3-20b-dev.safetensors",
+    "ltx-2.3-22b-dev.safetensors",
+)
+LTX23_DISTILLED_LORA_FILENAMES = (
+    "ltx-2.3-20b-distilled-lora-384.safetensors",
+    "ltx-2.3-22b-distilled-lora-384.safetensors",
+)
+LTX23_SPATIAL_UPSAMPLER_FILENAMES = (
+    "ltx-2.3-spatial-upscaler-x2-1.0.safetensors",
+    "ltx-2.3-spatial-upscaler-x2-1.1.safetensors",
+)


The resolution order for LTX-2.3 artifacts currently prioritizes older or smaller versions (e.g., 20b over 22b, 1.0 over 1.1). This is inconsistent with the overlay_manifest.json which specifies the 22b and 1.1 versions. Prioritizing the newer/larger versions ensures that the materializer picks the intended official artifacts when multiple versions are present in the source directory.

gemini-code-assist · 2026-04-06T08:39:00Z

+        av_ca_gate_factor = (
+            self.av_ca_timestep_scale_multiplier / self.timestep_scale_multiplier
+        )


Calculating av_ca_gate_factor by dividing by self.timestep_scale_multiplier without a zero-check poses a risk of ZeroDivisionError. Although this multiplier is typically non-zero (e.g., 1000), it is safer to handle the zero case or pre-calculate this factor during initialization.

av_ca_gate_factor = ( self.av_ca_timestep_scale_multiplier / self.timestep_scale_multiplier if self.timestep_scale_multiplier != 0 else 1.0 )

gemini-code-assist · 2026-04-06T08:39:00Z

            os.path.join(model_path, "ltx-2.3-spatial-upscaler-x2-1.0.safetensors"),
+            os.path.join(model_path, "ltx-2.3-spatial-upscaler-x2-1.1.safetensors"),


The resolution order for spatial upsampler candidates should prioritize version 1.1 over 1.0, as 1.1 is the version listed in the overlay_manifest.json and is generally preferred for LTX-2.3.

Suggested change

os.path.join(model_path, "ltx-2.3-spatial-upscaler-x2-1.0.safetensors"),

os.path.join(model_path, "ltx-2.3-spatial-upscaler-x2-1.1.safetensors"),

os.path.join(model_path, "ltx-2.3-spatial-upscaler-x2-1.1.safetensors"),

os.path.join(model_path, "ltx-2.3-spatial-upscaler-x2-1.0.safetensors"),

gemini-code-assist · 2026-04-06T08:39:01Z

+            os.path.join(model_path, "ltx-2.3-20b-distilled-lora-384.safetensors"),
            os.path.join(model_path, "ltx-2.3-22b-distilled-lora-384.safetensors"),


The resolution order for distilled LoRA candidates should prioritize the 22b version over 20b to align with the official LTX-2.3 artifacts specified in the manifest.

Suggested change

os.path.join(model_path, "ltx-2.3-20b-distilled-lora-384.safetensors"),

os.path.join(model_path, "ltx-2.3-22b-distilled-lora-384.safetensors"),

os.path.join(model_path, "ltx-2.3-22b-distilled-lora-384.safetensors"),

os.path.join(model_path, "ltx-2.3-20b-distilled-lora-384.safetensors"),

mickqian · 2026-04-07T01:52:28Z

/tag-and-rerun-ci

This reverts commit c192a31.

yhyang201 · 2026-04-11T11:55:51Z

/rerun-failed-ci

yhyang201 · 2026-04-11T18:53:23Z

/rerun-failed-ci

yhyang201 · 2026-04-11T22:01:06Z

/rerun-failed-ci

yhyang201 · 2026-04-11T23:56:52Z

/rerun-failed-ci

yhyang201 · 2026-04-12T01:48:01Z

/rerun-failed-ci

mickqian requested review from BBuf, ping1jing2, yhyang201 and yingluosanqian as code owners April 6, 2026 08:34

github-actions Bot added the diffusion SGLang Diffusion label Apr 6, 2026

gemini-code-assist Bot reviewed Apr 6, 2026

View reviewed changes

github-actions Bot added the run-ci label Apr 7, 2026

mickqian added 17 commits April 7, 2026 09:53

Align LTX-2.3 two-stage artifacts and official helpers

f142e56

Align LTX-2.3 two-stage default resolution

c46a1e5

init

53d1a1f

Use bundled LTX-2.3 overlay metadata

629d32e

Align LTX-2.3 stage2 RNG with official

7297253

Align LTX-2.3 stage2 noise layout

9c3c1ca

Revert LTX-2.3 stage2 noise layout change

8e1252a

Align LTX-2.3 native latent dtype with official

89df1b0

Fix LTX2 AV gate timestep scaling

d3ff749

Keep LTX-2.3 overlay transformer config official

99bf97e

Revert "Keep LTX-2.3 overlay transformer config official"

eb5efab

This reverts commit c192a31.

Fuse LTX2 two-stage distilled LoRA in stage2

667f244

Quantize LTX2 video RoPE coords to hidden dtype

7ef914c

Pass hidden dtype into LTX2 timestep AdaLN

ae6f33e

Make LTX2 token timesteps explicit

d6f0281

upd

ac8d171

Use sigma for LTX2 prompt AdaLN

b5553e1

mickqian force-pushed the ltx-2.3-two-stage branch from 85a356b to b5553e1 Compare April 7, 2026 01:53

mickqian added 4 commits April 7, 2026 18:12

Align LTX-2.3 two-stage stage2 RNG semantics

00488d1

upd

1031d9e

upd

bb1483c

upd

e242e82

mickqian added 6 commits April 11, 2026 21:09

fix accuracy

261ffb5

Merge remote-tracking branch 'origin/main' into ltx-2.3-two-stage

7bb3f5a

fix accuracy

3dc3b39

upd

3f42d6f

upd

544c941

upd

bee61a1

mickqian added 6 commits April 12, 2026 14:57

fix ltx 2.3 one stage regression

f834eea

upd

67fef89

remove ut

738e056

upd

2eb6d14

upd

f41a7f9

refactor

8c5c6e8

github-actions Bot added the documentation Improvements or additions to documentation label Apr 12, 2026

mickqian added 8 commits April 12, 2026 18:58

Restore LTX-2.3 one-stage legacy semantics

54787e5

Avoid getattr in LTX denoising stage

2ec753f

Format LTX denoising kwargs for black

77a327d

Skip LTX-2.3 two-stage I2V consistency until GT lands

b1fa4b4

upd

a0637e6

upd

ffa63bc

upd

32dfe98

upd

a44be9a

mickqian merged commit 495ef8e into main Apr 12, 2026
78 of 80 checks passed

mickqian deleted the ltx-2.3-two-stage branch April 12, 2026 14:16

pyc96 pushed a commit to pyc96/sglang that referenced this pull request Apr 14, 2026

[diffusion] model: support LTX2.3 two stage (sgl-project#22182)

5f957a9

yhyang201 pushed a commit to yhyang201/sglang that referenced this pull request Apr 22, 2026

[diffusion] model: support LTX2.3 two stage (sgl-project#22182)

17a612b

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[diffusion] model: support LTX2.3 two stage#22182

[diffusion] model: support LTX2.3 two stage#22182
mickqian merged 125 commits intomainfrom
ltx-2.3-two-stage

mickqian commented Apr 6, 2026 •

edited

Loading

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

gemini-code-assist Bot Apr 6, 2026

Uh oh!

gemini-code-assist Bot Apr 6, 2026

Uh oh!

gemini-code-assist Bot Apr 6, 2026

Uh oh!

gemini-code-assist Bot Apr 6, 2026

Uh oh!

mickqian commented Apr 7, 2026

Uh oh!

yhyang201 commented Apr 11, 2026

Uh oh!

yhyang201 commented Apr 11, 2026

Uh oh!

yhyang201 commented Apr 11, 2026

Uh oh!

yhyang201 commented Apr 11, 2026

Uh oh!

yhyang201 commented Apr 12, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		os.path.join(model_path, "ltx-2.3-spatial-upscaler-x2-1.0.safetensors"),
		os.path.join(model_path, "ltx-2.3-spatial-upscaler-x2-1.1.safetensors"),

		os.path.join(model_path, "ltx-2.3-20b-distilled-lora-384.safetensors"),
		os.path.join(model_path, "ltx-2.3-22b-distilled-lora-384.safetensors"),

Conversation

mickqian commented Apr 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Modifications

Accuracy Tests

Speed Tests and Profiling

Checklist

Review and Merge Process

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Apr 6, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 6, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 6, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 6, 2026

Choose a reason for hiding this comment

Uh oh!

mickqian commented Apr 7, 2026

Uh oh!

yhyang201 commented Apr 11, 2026

Uh oh!

yhyang201 commented Apr 11, 2026

Uh oh!

yhyang201 commented Apr 11, 2026

Uh oh!

yhyang201 commented Apr 11, 2026

Uh oh!

yhyang201 commented Apr 12, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

mickqian commented Apr 6, 2026 •

edited

Loading