ICANN 2026 · Poster

Training-Free Adaptive Sampling Scheduler for Discrete-Time Consistency Distillation

Qi Wang · Jinjia Zhou

Graduate School of Science and Engineering, Hosei University, Tokyo, Japan

In one sentence. Few-step consistency models fail not because the student is weak but because the schedule is rigid: we score every timestep by an SNR-derived Importance, spend the same number of steps where the trajectory actually moves, and add a matching sampler and colour-balancing pass — no retraining, drop-in for LCM, PCM, TCD and TDD. At 2 steps FID falls by up to 86%.

1. Why few-step sampling still breaks

Scheduler rigidity. Every discrete-time distillation method ships a timestep schedule welded to its own training framework and step budget. Deterministic-target methods (LCM, PCM) and stochastic-target methods (TCD, TDD) alike degrade as soon as the scheduler or the step count changes at inference time — the schedule was never chosen on the merits of the timesteps themselves.

CFG compounding. Classifier-free guidance multiplies with the inner distillation scale, pushing pixels into saturation and drifting the colour balance. The fewer steps you take, the less opportunity the model has to correct it.

2. Timesteps are not equally informative

Forward diffusion at a sequence of timesteps: the image stays legible early, becomes unrecognisable late, and changes fastest in between.
Forward diffusion (DDPM). Up to the content is still fully discernible; past it is essentially unrecognisable. Almost all structural change is packed into the middle of the trajectory.

We quantify that observation with an Importance measure : the rate of change of log-SNR, inverted and normalised to . A large marks a timestep where the SNR moves fast — where the ODE trajectory turns. The complementary skip rate says how much of that step can safely be skipped.

Importance & skip rate (Eq. 8)  |  ,
Importance per timestep: a curve peaking near t = 350. Skip ratio per timestep: the mirror image of the Importance curve.
Fig. 3. (a) peaks near ; the dashed line is the threshold , whose two crossings bracket the band worth sampling densely. (b) The skip rate mirrors it.

3. Adaptive schedule

Let be the original uniformly spaced timesteps and the timesteps that attain the highest Importance within predefined intervals. We keep the uniform-grid steps where change is slow and substitute high-Importance candidates where it is fast. The step count never changes — only its spacing. That is what makes the method free: no retraining, no extra parameters, no change to the sampler loop's cost.

Adaptive schedule (Eq. 9),

4. The -I sampler

CTM re-injects noise with a fixed at every step. We instead scale the re-injection by the local Importance, so exploration widens exactly where the trajectory turns fastest and narrows where it is nearly straight.

Uniform step (Eq. 10)  /  high-Importance step (Eq. 11)
Diagram of adaptive timestep selection: dense sampling in the high-Importance region, sparse elsewhere. Diagram of the Importance-modulated forward and backward transitions in the gamma-I sampler.
Fig. 4. (a) Dense sampling in high-Importance regions, sparse elsewhere, at a preserved budget. (b) Importance-modulated forward and backward transitions lift sampling quality precisely where the SNR changes most rapidly.

5. Smoothing & colour balancing

To undo CFG compounding we squash outliers with , then re-centre the channel-wise and global means so values return to the function's linear region instead of pinning at . Unlike prior mapping methods this needs no prior deformation, and more information survives under strong guidance.

Soft clamp & re-centring (Eq. 12–13),

6. Quantitative results

MS-COCO 2014 validation set, Karpathy 30k captions, CFG . Columns are 2 / 4 / 8 sampling steps. MPI = mean improvement over the paired baseline.

Method FID ↓ CLIP Score ↑ Inception Score ↑
248 248 248
SDXL — 1024×1024
PCM372.82112.6531.7318.4424.1830.441.7111.6725.78
PCM + Ours65.8329.4023.2130.2230.4031.5216.5424.5931.80
TCD363.50103.6653.7218.7326.0530.441.8212.4717.83
TCD + Ours62.8928.5127.4428.8831.7132.0616.3328.0232.17
TDD58.4529.8027.7229.2731.1231.4717.1827.0229.72
TDD + Ours55.7127.8826.6029.4931.4731.7417.6229.3632.88
△ MPI203.4553.4411.977.384.080.999.9310.277.84
SD v1.5 — 512×512
LCM86.3388.20109.8428.0226.4825.1914.2811.738.71
LCM + Ours58.0130.0447.6730.1830.7130.7917.1828.8119.84
PCM424.0489.9938.8218.9926.5130.021.7612.4921.07
PCM + Ours60.6623.1122.4729.9330.3731.0316.8628.9331.65
△ MPI195.8562.5239.266.554.053.319.0016.7610.86

Scroll the table sideways to see all nine columns.

Reading the table

All five baselines improve on every metric at every budget, and the gains land where few-step sampling is weakest — at 2 steps FID falls 82–86% for PCM and TCD. The most telling case is LCM: on its own it gets worse with more steps (86.33 → 88.20 → 109.84 FID), the signature of CFG overexposure compounding. Our scheduler reverses that trend. The pattern holds across both resolutions and all four distilled students, not just on the metric we optimised for.

7. Qualitative comparison

Grid comparing LCM, PCM, TCD and TDD against the same methods with our scheduler, at 2, 4 and 8 sampling steps, on SD v1.5 and SDXL.
Fig. 5. At 2 and 4 steps the baselines lose global structure and wash the colour out; ours keeps the layout coherent and the detail sharper on both backbones. By 8 steps the gap narrows, as expected once the student has budget enough to correct itself.

8. Ablations

Sample generated without the gamma sampler.
no
Sample generated with CTM's original fixed gamma sampler.
CTM's
Sample generated with our Importance-scaled gamma-I sampler.
ours (-I)
Fig. 7 — the sampler. Scaling noise re-injection by the local Importance, instead of CTM's fixed , gives more stable shapes and fewer structural errors at the same step budget. Threshold sweep: best .
Two rows of samples: the top row uses the clipping of prior work, the bottom row uses our smoothing and colour balancing.
Fig. 8 — smoothing & colour balance. Top row: the clipping used by prior work. Bottom row: ours — natural colour, fewer saturated regions and more surviving detail at high CFG.

9. What to take away