[MAX] Add Wan video generation examples and comparison benchmark

[jglee-sqbits]

Stacked PRs:

• ->[MAX] Add Wan video generation examples and comparison benchmark #19
• [MAX] Add Wan I2V diffusion pipeline #18
• [MAX] Add Wan T2V diffusion pipeline with MoE support #17
• [MAX] Add Wan transformer model with block-level compilation #16
• [MAX] Add Wan VAE and refactor autoencoder module #15
• [MAX] Add UMT5 text encoder for Wan diffusion #14
• [MAX] Add UniPC multistep scheduler for Wan diffusion #13

---

[MAX] Add Wan video generation examples and comparison benchmark

Summary

Add a standalone video generation example script for Wan T2V and I2V pipelines.

Description

• simple_offline_video_generation.py: end-to-end script for generating videos from text or image prompts
• Supports all Wan model variants: 2.2-A14B (MoE), 2.1-14B, T2V and I2V
• LoRA turbo support (e.g. Lightning 4-step)
• Built-in profiling with component-level timing breakdown
• Input images can be local files or URLs (downloaded at runtime)
• Outputs MP4 video via av (PyAV)

Validation (H200 140GB, 720p 81 frames)

# T2V base (Wan2.2-A14B MoE, 720p, 40 steps)
MODULAR_DEVICE_CONTEXT_MEMORY_MANAGER_CHUNK_PERCENT=100 \
./bazelw run //max/examples/diffusion:simple_offline_video_generation -- \
    --model Wan-AI/Wan2.2-T2V-A14B-Diffusers \
    --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage." \
    --negative-prompt "low quality" \
    --height 720 --width 1280 --num-frames 81 \
    --num-inference-steps 40 --guidance-scale 4.0 \
    --guidance-scale-2 3.0 \
    --output t2v_base.mp4

# T2V LoRA turbo (4 steps)
MODULAR_DEVICE_CONTEXT_MEMORY_MANAGER_CHUNK_PERCENT=100 \
./bazelw run //max/examples/diffusion:simple_offline_video_generation -- \
    --model Wan-AI/Wan2.2-T2V-A14B-Diffusers \
    --prompt "A cat playing piano" \
    --height 720 --width 1280 --num-frames 81 \
    --num-inference-steps 4 --guidance-scale 1.0 \
    --lora-repo-id lightx2v/Wan2.2-Lightning \
    --lora-subfolder Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V2.0 \
    --output t2v_lora.mp4

# I2V base (720p, 40 steps)
MODULAR_DEVICE_CONTEXT_MEMORY_MANAGER_CHUNK_PERCENT=100 \
./bazelw run //max/examples/diffusion:simple_offline_video_generation -- \
    --model Wan-AI/Wan2.2-I2V-A14B-Diffusers \
    --prompt "A cat surfing on a wave" \
    --negative-prompt "low quality" \
    --height 720 --width 1280 --num-frames 81 \
    --num-inference-steps 40 --guidance-scale 4.0 \
    --guidance-scale-2 3.0 \
    --input-image https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B/resolve/main/examples/i2v_input.JPG \
    --output i2v_base.mp4

# I2V LoRA turbo (4 steps)
MODULAR_DEVICE_CONTEXT_MEMORY_MANAGER_CHUNK_PERCENT=100 \
./bazelw run //max/examples/diffusion:simple_offline_video_generation -- \
    --model Wan-AI/Wan2.2-I2V-A14B-Diffusers \
    --prompt "A cat surfing on a wave" \
    --height 720 --width 1280 --num-frames 81 \
    --num-inference-steps 4 --guidance-scale 1.0 \
    --lora-repo-id lightx2v/Wan2.2-Lightning \
    --lora-subfolder Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1 \
    --input-image https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B/resolve/main/examples/i2v_input.JPG \
    --output i2v_lora.mp4

Dependencies

Depends on all previous PRs: modular#6298–modular#6303.

Checklist

• PR is small and focused
• I ran ./bazelw run format to format my changes

Assisted-by: Claude Code

Assisted-by: Claude Code

[gemini-code-assist]

Code Review

This pull request introduces two new Python utilities for Wan video generation: a benchmarking script for model speed metrics and an offline video generation example. The feedback identifies a critical deadlock risk when executing nested Bazel commands, highlights a mismatch in target names within the documentation, suggests a more robust method for resolving the workspace directory, and points out an inefficiency where input images are downloaded multiple times.

[gemini-code-assist]

Running bazel run inside a script that is itself executed via bazel run will cause a deadlock because the first Bazel process holds the workspace lock. Consider running the underlying Python script directly or building the target first and executing the resulting binary.

[gemini-code-assist]

The Bazel target name in the usage examples (full_metric) does not match the actual target name defined in BUILD.bazel (all_wan_model_speed_metric).

Suggested change

	./bazelw run //max/examples/diffusion:full_metric
	# Specific model only
	MODULAR_DEVICE_CONTEXT_MEMORY_MANAGER_CHUNK_PERCENT=100 \
	./bazelw run //max/examples/diffusion:full_metric -- \
	--model wan2.2-t2v-a14b
	./bazelw run //max/examples/diffusion:all_wan_model_speed_metric
	# Specific model only
	MODULAR_DEVICE_CONTEXT_MEMORY_MANAGER_CHUNK_PERCENT=100 \
	./bazelw run //max/examples/diffusion:all_wan_model_speed_metric -- \
	--model wan2.2-t2v-a14b

[gemini-code-assist]

Resolving the workspace root using __file__ and parents[3] is brittle and may fail when the script is executed via Bazel (where __file__ points into the runfiles directory). It is safer to use the BUILD_WORKSPACE_DIRECTORY environment variable provided by bazel run.

Suggested change

	cwd=str(Path(__file__).resolve().parents[3]),
	cwd=os.environ.get("BUILD_WORKSPACE_DIRECTORY", str(Path(__file__).resolve().parents[3])),

[gemini-code-assist]

The input image is downloaded here to compute dimensions and then downloaded again in generate_video. This is inefficient. Consider downloading the image once and passing the loaded PIL.Image object or the raw bytes to the rest of the pipeline.