[mickqian]

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

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. 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
4. 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]

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]

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]

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]

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]

/tag-and-rerun-ci

[yhyang201]

/rerun-failed-ci

[yhyang201]

/rerun-failed-ci

[yhyang201]

/rerun-failed-ci

[yhyang201]

/rerun-failed-ci