How to Implement Few-Step DMD Distillation for Faster Inference in LongLive

To implement few-step DMD distillation in LongLive, reduce the num_train_timestep configuration parameter from 1000 to a smaller value (e.g., 100) while using the score_distillation trainer, which enables the DMD class to sample from a compressed diffusion schedule during training.

LongLive is an open-source video generation framework by NVlabs that uses Distribution Matching Distillation (DMD) to train efficient student generators from teacher diffusion models. While standard DMD uses approximately 1000 diffusion timesteps for high-quality distillation, you can configure few-step DMD distillation to reduce inference time by 10× or more. This implementation requires modifying only the diffusion schedule parameters in model/dmd.py and utils/config.py, with no changes needed to the core loss computation.

Understanding the DMD Architecture and Schedule Parameters

The few-step distillation capability centers on the DMD class in model/dmd.py (lines 13-31, 132-166), which wraps the student generator, real-score network, and fake-score network. During training, this class samples timesteps from a configurable diffusion trajectory and computes the KL-gradient loss between teacher and student predictions.

Key methods governing the diffusion process include:

• _compute_kl_grad (lines 92-134): Implements the teacher-student KL gradient matching from the original DMD paper.
• compute_distribution_matching_loss (lines 166-226): Converts the raw KL gradient into the final optimization objective.
• generator_loss and critic_loss (lines 236-298, 300-378): Handle forward passes, timestep sampling via _get_timestep, and optional critic training.

The diffusion schedule itself is defined by four critical parameters read during DMD.__init__ (lines 42-45):

• num_train_timestep: Total number of diffusion steps in the trajectory (default 1000).
• min_step and max_step: Percentage bounds (typically 2 and 98) that constrain sampling to avoid boundary errors.
• ts_schedule and ts_schedule_max: Boolean flags that bias trajectory sampling toward specific timestep ranges.

Because the DMD class dynamically reads these values at initialization, reducing num_train_timestep automatically compresses the entire diffusion trajectory while preserving the mathematical correctness of the loss functions.

Configuring a Few-Step Schedule in utils/config.py

When your configuration specifies trainer: score_distillation, the loader in utils/config.py (lines 45-61) injects default DMD hyperparameters. To enable few-step distillation, override the num_train_timestep value in your YAML configuration file:

model_name: "wan_t2v_A14B"
trainer: score_distillation
num_train_timestep: 100      # Compress from 1000 to 100 steps

ts_schedule: true
ts_schedule_max: false
teacher_forcing: true        # Recommended for few-step convergence

min_step: 2                  # 2% of 100 = step 2

max_step: 98                 # 98% of 100 = step 98

Why this works: The DMD class reads num_train_timestep during initialization and uses it to bound the random timestep sampler in _get_timestep. By shrinking this value from 1000 to 100, you restrict the student to learn only a few-step diffusion trajectory, eliminating the need for full-width sampling during inference.

Optional Speed-Up Parameters

For optimal few-step convergence, consider adjusting these additional parameters defined in utils/config.py:

• teacher_forcing: Set to true to force the teacher network to generate latents matching the student’s trajectory, which stabilizes training when using fewer timesteps.
• backward_simulation: Keep as true (default) to maintain the backward-sampled generator input required for proper DMD gradient flow.
• real_guidance_scale and fake_guidance_scale: Maintain default values (3.0 and 0.0) unless experimenting with altered gradient magnitudes for specific few-step regimes.

Launching Few-Step Distillation Training

Execute the distillation process using the standard training entry point:

python train.py --config configs/few_step.yaml

The train.py script detects config.trainer == "score_distillation" and constructs a ScoreDistillationTrainer (imported from trainer/distillation.py), which instantiates the DMD class with your compressed schedule. During training, verify that timestep values logged under the timestep metric remain within your configured bounds (e.g., 2-98), confirming that the few-step schedule is active.

Running Fast Inference with Distilled Checkpoints

After training completes, deploy the distilled student using the standard inference scripts:

python inference.py \
  --checkpoint path/to/few_step_checkpoint.pt \
  --prompt "A futuristic cityscape at sunset" \
  --num_frames 16 \
  --output fast_generation.mp4

Because the student was trained specifically on a few-step diffusion schedule, the sampling loop executes in approximately 10× less time than the full 1000-step configuration while retaining the visual quality of the original DMD distillation.

Summary

• Reduce num_train_timestep in your configuration (e.g., from 1000 to 100) to enable few-step DMD distillation.
• The DMD class in model/dmd.py automatically adapts its _get_timestep sampler and loss computation to the compressed schedule without requiring code changes.
• Use trainer: score_distillation to trigger the ScoreDistillationTrainer, which initializes the DMD pipeline with your custom schedule parameters.
• Enable teacher_forcing: true for improved convergence when training with fewer than 200 timesteps.
• Inference uses the same scripts (inference.py or inference_sp.py) but completes in significantly fewer steps, delivering faster generation.

Frequently Asked Questions

What is the optimal number of timesteps for few-step DMD in LongLive?

Most implementations use between 50 and 200 timesteps when configuring few-step DMD distillation, with 100 steps serving as the optimal balance between speed and quality. This configuration typically reduces inference time by 90% compared to the default 1000-step schedule while maintaining visual fidelity comparable to the full teacher model.

Does reducing timesteps affect the DMD loss computation in model/dmd.py?

No, the core DMD mathematics remains unchanged; the _compute_kl_grad method (lines 92-134) and compute_distribution_matching_loss method (lines 166-226) function identically regardless of schedule length. The adaptation occurs purely in the timestep sampling logic within DMD.__init__ (lines 42-45) and _get_timestep, which respect the new num_train_timestep bounds automatically.

Where are the default DMD hyperparameters defined in the LongLive repository?

Default DMD parameters are injected in utils/config.py between lines 45 and 61 whenever the configuration specifies trainer: score_distillation. This block sets num_train_timestep: 1000, min_step: 2, max_step: 98, and guidance scale defaults, all of which you can override in your custom YAML configuration without modifying the source code.

Can few-step DMD distillation work with video models other than Wan-T2V?

Yes, the few-step schedule configuration is architecture-agnostic and functions with any diffusion backbone supported by LongLive, including various Wan-T2V sizes and other video generation architectures. The num_train_timestep parameter affects only the trajectory length during distillation, not the underlying student generator architecture or the inference pipeline in inference.py.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}