[Question] Rationale for timestep-pair sampling vs. LCM-style DDIM anchors

[Aircraft-1111]

Hi, thank you for releasing the Flash-WAM code and paper!

I am trying to understand the rationale behind the timestep-pair sampling strategy used for consistency distillation.

From the current implementation, my understanding is:

1. The starting timestep index is sampled uniformly from all 1,000 training timesteps for each frame:
   distillation/data.py

2. The stride is computed as:

   k = num_train_timesteps // num_ddim_timesteps

   distillation/trainer.py

3. The endpoint is then selected by advancing k scheduler indices and clamping to the final index:

   end_ids = (start_ids + k).clamp(max=num_train_timesteps - 1)

   This is done for both video and action:
   distillation/step.py

With the released configuration,

num_train_timesteps = 1000
num_ddim_timesteps = 2

we have k = 500:

distillation/config.py

This produces pairs such as:

0   -> 500
100 -> 600
499 -> 999
500 -> 999
700 -> 999
900 -> 999
999 -> 999

Therefore, 501 out of the 1,000 possible starting indices map to the same terminal index 999. In addition, start_id = 999 produces a zero-length pair.

This seems different from the standard LCM training implementation. LCM first constructs a reduced DDIM solver grid—commonly 50 DDIM timesteps—and samples adjacent pairs from that grid. The number of teacher/DDIM discretization steps is therefore not necessarily the same as the final number of student inference steps:

• LCM DDIMSolver implementation
• LCM timestep-pair sampling

I am particularly curious about the following points:

1. Why are starting timesteps sampled from all 1,000 scheduler indices instead of from an LCM-style reduced grid, such as 50 DDIM anchors?

2. Is num_ddim_timesteps=2 intentionally tied to the target inference NFE?
   In LCM, the DDIM solver resolution used to construct training pairs can be much larger than the final student sampling NFE.

3. Is the terminal concentration intentional?
   With k=500, approximately 50.1% of all starting indices are clamped to end_id=999. Does this act as an intentional clean-boundary weighting, or is it mainly an implementation simplification?

4. Why is the interval defined in scheduler-index space rather than sigma space?
   Flash-WAM uses different SNR shifts for video and action (shift=5 vs. shift=1), so the same 500-index stride corresponds to very different and highly non-uniform Δσ values.

   For example, approximately:

   Video, shift=5:
   index 0   -> 500: sigma 1.000 -> 0.833
   index 500 -> 999: sigma 0.833 -> 0.005
   
   Action, shift=1:
   index 0   -> 500: sigma 1.000 -> 0.500
   index 500 -> 999: sigma 0.500 -> 0.001
   

5. Does a single teacher Euler step remain sufficiently accurate for the largest sigma intervals?
   Since the Euler target error generally grows with the interval size, I wonder whether a finer teacher grid, multiple teacher substeps, or a higher-order solver was considered.

6. Did you compare this strategy with any of the following alternatives?

   • LCM-style 50-step DDIM anchor pairs;
   • sampling exact pairs used by the final 1-step/2-step inference schedule;
   • phase-aligned endpoints, as in Phased Consistency Models;
   • continuous or arbitrary timestep pairs, as in ECT or CTM;
   • sampling based on a fixed Δσ instead of a fixed scheduler-index stride.

7. Could you share the theoretical or empirical basis for this timestep sampler?
   In particular, was it inspired by a previous consistency-distillation or flow-matching paper? If so, could you point to the relevant paper, section, or implementation?

Some potentially related works I have been comparing against are:

• Latent Consistency Models
• Phased Consistency Models
• Consistency Models Made Easy / ECT
• Consistency Trajectory Models
• Target-Driven Distillation
• Simplifying, Stabilizing and Scaling Continuous-Time Consistency Models

I may be misunderstanding the scheduler indexing direction or the intended relationship between the training stride and inference NFE, so please correct me if that is the case.

Thank you for your time and for open-sourcing the project!

[armanakbari]

Hi, thank you for releasing the Flash-WAM code and paper!

I am trying to understand the rationale behind the timestep-pair sampling strategy used for consistency distillation.

From the current implementation, my understanding is:

1. The starting timestep index is sampled uniformly from all 1,000 training timesteps for each frame:
   distillation/data.py

2. The stride is computed as:
   k = num_train_timesteps // num_ddim_timesteps

   distillation/trainer.py

3. The endpoint is then selected by advancing k scheduler indices and clamping to the final index:
   end_ids = (start_ids + k).clamp(max=num_train_timesteps - 1)

   This is done for both video and action:
   distillation/step.py

With the released configuration,

num_train_timesteps = 1000
num_ddim_timesteps = 2
we have k = 500:

distillation/config.py

This produces pairs such as:

0   -> 500
100 -> 600
499 -> 999
500 -> 999
700 -> 999
900 -> 999
999 -> 999

Therefore, 501 out of the 1,000 possible starting indices map to the same terminal index 999. In addition, start_id = 999 produces a zero-length pair.

This seems different from the standard LCM training implementation. LCM first constructs a reduced DDIM solver grid—commonly 50 DDIM timesteps—and samples adjacent pairs from that grid. The number of teacher/DDIM discretization steps is therefore not necessarily the same as the final number of student inference steps:

• LCM DDIMSolver implementation
• LCM timestep-pair sampling

I am particularly curious about the following points:

1. Why are starting timesteps sampled from all 1,000 scheduler indices instead of from an LCM-style reduced grid, such as 50 DDIM anchors?

2. Is num_ddim_timesteps=2 intentionally tied to the target inference NFE?
   In LCM, the DDIM solver resolution used to construct training pairs can be much larger than the final student sampling NFE.

3. Is the terminal concentration intentional?
   With k=500, approximately 50.1% of all starting indices are clamped to end_id=999. Does this act as an intentional clean-boundary weighting, or is it mainly an implementation simplification?

4. Why is the interval defined in scheduler-index space rather than sigma space?
   Flash-WAM uses different SNR shifts for video and action (shift=5 vs. shift=1), so the same 500-index stride corresponds to very different and highly non-uniform Δσ values.
   For example, approximately:

   Video, shift=5:
   index 0   -> 500: sigma 1.000 -> 0.833
   index 500 -> 999: sigma 0.833 -> 0.005
   
   Action, shift=1:
   index 0   -> 500: sigma 1.000 -> 0.500
   index 500 -> 999: sigma 0.500 -> 0.001
   

5. Does a single teacher Euler step remain sufficiently accurate for the largest sigma intervals?
   Since the Euler target error generally grows with the interval size, I wonder whether a finer teacher grid, multiple teacher substeps, or a higher-order solver was considered.

6. Did you compare this strategy with any of the following alternatives?

   • LCM-style 50-step DDIM anchor pairs;
   • sampling exact pairs used by the final 1-step/2-step inference schedule;
   • phase-aligned endpoints, as in Phased Consistency Models;
   • continuous or arbitrary timestep pairs, as in ECT or CTM;
   • sampling based on a fixed Δσ instead of a fixed scheduler-index stride.

7. Could you share the theoretical or empirical basis for this timestep sampler?
   In particular, was it inspired by a previous consistency-distillation or flow-matching paper? If so, could you point to the relevant paper, section, or implementation?

Some potentially related works I have been comparing against are:

• Latent Consistency Models
• Phased Consistency Models
• Consistency Models Made Easy / ECT
• Consistency Trajectory Models
• Target-Driven Distillation
• Simplifying, Stabilizing and Scaling Continuous-Time Consistency Models

I may be misunderstanding the scheduler indexing direction or the intended relationship between the training stride and inference NFE, so please correct me if that is the case.

Thank you for your time and for open-sourcing the project!

Thanks for the very careful read, and sorry for my late response. Your understanding of the sampler is correct on all counts. Most of what you flagged (terminal concentration, coarse stride, index-space interval) is a fair reading of an implementation choice rather than a deeply theorized one, and several of your suggested alternatives are things we'd genuinely like to add.

1. Why sample start from all 1000 indices instead of an LCM-style reduced grid?
The start is the noise level of the flow-matching training sample, drawn uniformly over the whole schedule (per frame), so the student is supervised to be consistent everywhere along the trajectory rather than only at a handful of anchors. The endpoint is then start + k. This does differ from LCM restricting starts to the reduced grid; with a coarse k it also produces the boundary concentration you noticed.

2. Is num_ddim_timesteps = 2 intentionally tied to the target inference NFE?
Yes — the released grid is set coarse to target the 1–2-step inference regime, so here the teacher grid resolution and the student NFE are not decoupled the way LCM decouples a 50-step solver from a few-step student. Candidly, this is an aggressive-distillation simplification. Decoupling them (a finer teacher grid, e.g. 50 anchors, independent of the inference budget) is a reasonable refinement we'd expect to help.

3. Is the terminal concentration (~50.1% of starts → end = 999) intentional?
It is a direct consequence of k = 500 with a uniform start and the clamp, not a separately motivated clean-boundary weighting. It does up-weight the clean boundary, and start = 999 gives a degenerate zero-length pair (~0.1% of draws). A finer grid removes both effects.

4. Why an index-space interval rather than a σ-space interval, given the different SNR shifts?
You're right that a fixed index stride under shift = 5 (video) vs shift = 1 (action) maps to very different, non-uniform $\Delta \sigma$ for the two streams — your numbers are correct. The modality-awareness in our method is deliberately placed in the consistency function f = a(σ)x + b(σ)v — a variance-preserving/Karras b(σ) for the high-σ video stream and the linear b(σ) = −σ (f = x − σv) for the low-σ action stream — rather than in the sampler stride. A σ-space stride, or a per-modality k, is a sensible and fully compatible extension; it just wasn't necessary for the claims in the paper.

5. Is a single Euler step accurate enough for the largest-σ intervals?
A fair concern — Euler target error does grow with interval size. In our setting the inference budget is itself 1–2 steps, so we tolerate a coarse teacher step, and empirically it is the modality-aware consistency function that recovers task success. A finer teacher grid, multiple teacher substeps, or a higher-order solver would tighten the teacher target and is a reasonable improvement.

6. Comparisons to LCM-50-anchor / exact-inference pairs / PCM phase-aligned / ECT–CTM continuous pairs / fixed-Δσ?
Our sampler follows the LCM/flow-matching recipe and we did not ablate it against these alternatives. Given your analysis, these are exactly the right axes to study — phase-aligned endpoints (PCM), continuous/arbitrary pairs (ECT/CTM), and a fixed-$\Delta \sigma$ or per-modality stride in particular — and we'd be glad to add such a comparison.

7. Theoretical/empirical basis and inspiration.
The sampler is adapted from the LCM / latent-consistency line (reduced-DDIM grid + one guided Euler step), not from a novel timestep-scheduling result.

---

You've correctly identified real simplifications in the timestep sampler. I am thinking of a follow-up work on this work to enhance the method as Flash-WAM is only the first work directly targeting WAM for distillation. Thanks again for the thorough and precise feedback.