1 of 69

Distilling Diffusion Models

Shreshth Saini · Sep 2026

shreshthsaini.github.io/blogs/diffusion-distillation.html

50 steps

4 steps

[1] Yin et al. 2024, CC BY 4.0

2 of 69

What we will cover today

2

What is the student allowed to forget?

1

Why diffusion

is slow

2

What distillation is

3

Three things a

student can copy

4

Copy the

teacher's path

5

Match the teacher's

distribution

6

Distilling

video models

7

What still

breaks

3 of 69

Part 1

Why diffusion is slow

Why does one image need a hundred network passes?

4 of 69

Sampling takes many small steps

4

[1] Yin et al. 2024, arXiv:2405.14867 · [49] Team Wan 2025, arXiv:2503.20314

noise, t = 1

image x₀, t = 0

every dot is one forward pass

the probability-flow ODE

cost per sample = NFE × network cost

5 of 69

From 100 network passes down to 4

5

[1] Yin et al. 2024, arXiv:2405.14867 · [19] Huang et al. 2025, arXiv:2506.08009 · [49] Team Wan 2025, arXiv:2503.20314

teacher

50 steps × 2 for guidance = 100 passes

student

1 to 4 passes

distillation

25 to 100 times fewer

Wan2.1 1.3B takes 103 s for one clip on an H100.

6 of 69

Why one big step gives a blur

6

noisy input

a cat

a dog

half cat, half dog

the one-step answer

This noise could become a cat or a dog.

Asked to answer in one step, the network returns the average, a blur.

7 of 69

One step lands on the average image

7

data

cat

dog

low noise

xₜ

average

mid noise

xₜ

high noise

xₜ

the blur, between the peaks

Squared-error training learns this average.

8 of 69

Part 2

What distillation is

How does a student learn from a teacher?

9 of 69

Distillation in one picture

9

[7] Hinton et al. 2015, arXiv:1503.02531

label

BMW

truck

carrot

teacher

BMW

truck

carrot

student

BMW

truck

carrot

match

The label says BMW. The teacher says BMW, a bit garbage truck, almost no carrot.

The student learns from that extra shape.

10 of 69

Language models went through this first

10

2015

Soft targets

Hinton

2016

Sequence-level KD

Kim and Rush

2024

Reverse KL

MiniLLM

2024

Student rollouts

GKD

Trajectory matching

on teacher states

DMD

on student samples

Self Forcing

on student rollouts

language models

diffusion

off-policy

on-policy

the student starts sampling here

11 of 69

In diffusion, the student learns to skip steps

11

[6] Salimans and Ho 2022, arXiv:2202.00512

Teacher

Student

copied weights

t = 1

t = 0

every noise level

just 1 to 4 levels run

12 of 69

Why diffusion is harder to distill

12

x₀

50 steps

?

No few-step answer key

ε(xₜ, c)

ε(xₜ, ∅)

γ

student, fixed γ

Guidance is in the target

encoder

teacher

student

loss

latent

decoder

The VAE sits outside

×

student

fake score

real score

Video needs more memory

13 of 69

Who sets the target, and on what samples

13

Off-policy

trained on data or teacher states

On-policy

trained on the student's own samples

Teacher distillation

target from a frozen teacher

Self-distillation

target from its own output

also called teacher-free

consistency training

rare

on-policy still

has a teacher

14 of 69

Learning to drive, three ways

14

off-policy

Watch the instructor drive.

on-policy

Drive while the instructor

corrects you.

self-distillation

Or check your answer against

your own later answer.

consistency training, MeanFlow

instructor

student watches

student

instructor

correction

student

you, later

own target

still has a teacher

15 of 69

Forward KL spreads, reverse KL picks a peak

15

forward KL, the teacher's loss

data p

fit q

reverse KL, the student's loss

data p

fit q

skipping this mode costs nothing

A student can beat its teacher on FID while covering less.

16 of 69

Part 3

Three things a student can copy

What exactly does the student try to match?

17 of 69

Three ways to learn to draw

17

trace

Trace the teacher's strokes.

trajectory

match the stack

Make your whole stack of

drawings look like theirs.

distribution

critic

Or let a critic call each

one real or fake.

discriminator

drawing

critic

real

fake

18 of 69

Every method copies one of three things

18

[27] Song et al. 2023, arXiv:2303.01469 · [40] Yin et al. 2024, arXiv:2311.18828 · [44] Sauer et al. 2023, arXiv:2311.17042

trajectory

a point

distribution

a density

discriminator

a verdict

teacher step

student

loss

noise

data

real

fake

score difference

student

real or

teacher-made

D

real

fake

19 of 69

Part 4

Copy the teacher's path

How did the first methods cut the steps?

20 of 69

Learning to skip stairs

20

Learn to cover two of the teacher's steps in one.

Then the student becomes the teacher.

Repeat: 8, 4, 2, 1.

teacher

4

2

1

8

steps

one round of training each

21 of 69

Progressive distillation halves the steps each round

[6] Salimans and Ho 2022, arXiv:2202.00512

21

teacher distillation, off-policy

steps

CIFAR-10 FID

8

2.57

4

3.00

2

4.51

1

9.12

two teacher steps

one student step

each student becomes

the next teacher

one student step

two teacher DDIM steps

MSE

22 of 69

Guidance distillation folds two passes into one

[20] Meng et al. 2023, arXiv:2210.03142

22

teacher distillation, off-policy

teacher, two evaluations

student, one evaluation

guided target

MSE

evaluations per sample

2N

N

N/2, N/4, …

FID: guided teacher 9.84, one-step student 7.54

ImageNet 64, w = 1

the student beats its teacher

23 of 69

The idea behind consistency models

23

noise

the teacher's path

image

must agree

From any point on the trip, say where it ends.

Two nearby points must give the same answer.

24 of 69

A consistency model jumps straight to the image

[27] Song et al. 2023, arXiv:2303.01469

24

noise

data

ODE trajectory

identity here, by construction

same trajectory, same output

25 of 69

Consistency distillation makes neighbours agree

[27] Song et al. 2023, arXiv:2303.01469

25

teacher distillation, off-policy

consistency loss

noise

data

teacher ODE step

EMA copy, stop-grad

noised real data,

so off-policy

FID

1 step

2 steps

teacher

CD, CIFAR-10

3.55

2.93

2.04

CD, ImageNet 64

6.20

4.70

2.44

26 of 69

Consistency training drops the teacher

[27] Song et al. 2023, arXiv:2303.01469 · [28] Song and Dhariwal 2024, arXiv:2310.14189 · [30] Lu and Song 2025, arXiv:2410.11081

26

self-distillation (teacher-free), off-policy

Method

Teacher

Benchmark

1-step FID

2-step FID

CD, 2023

EDM

CIFAR-10

3.55

2.93

CT, 2023

none

CIFAR-10

8.70

5.83

iCT-deep, 2024

none

CIFAR-10

2.51

2.24

sCT-XXL, 2025

none

ImageNet 512

4.29

3.76

no teacher, and

it beats CD

27 of 69

LCM brought four steps to Stable Diffusion

[24] Luo et al. 2023, arXiv:2310.04378

27

teacher distillation, off-policy

LCM FID, 512px

35.36

1

13.31

2

11.10

4

11.84

8

steps

the four-step default

28 of 69

Reflow straightens the path

[31] Liu et al. 2022, arXiv:2209.03003 · [32] Liu et al. 2024, arXiv:2309.06380

28

teacher distillation, off-policy

1-rectified flow

noise

image

one Euler step misses

2-rectified flow

noise

image

one Euler step lands

On SD 1.5, InstaFlow gets FID 23.4 in one step, against the teacher's 20.1.

the price of a straight path

29 of 69

Flow maps learn whole jumps at once

[34] Geng et al. 2025, arXiv:2505.13447 · [33] Frans et al. 2025, arXiv:2410.12557

29

self-distillation, off-policy

instantaneous velocity

average velocity

noise

data

trajectory

one jump, no solver

MeanFlow identity

Shortcut models make one 2d step match two of their own d steps.

30 of 69

The trouble with copying the path

30

one step blurs

every teacher error is copied

the teacher is the ceiling

keeps the noise-to-image pairing and variety

Part 5 starts here

31 of 69

One step on a curved path blurs

31

[32] Liu et al. 2024, arXiv:2309.06380 (MIT) · [6] Salimans and Ho 2022, arXiv:2202.00512, Fig. 3 · cropped

SD 1.5, 25 steps

InstaFlow, 1 step

straightened by reflow

256 steps

1 step

the average image, drawn

32 of 69

One-step scores kept getting better

32

teacher distillation

self-distillation

CIFAR-10, one step

FID, lower is better

2022

2023

2024

PD 9.12

CT 8.70

CD 3.55

iCT-deep 2.51

ImageNet 256, no teacher

FID, lower is better

2025

2026

multi-step SiT, 2.06

DMF 2.16

iMF 1.72

no teacher anywhere

33 of 69

Part 5

Match the teacher's distribution

What if we judge the result and ignore the path?

34 of 69

Distribution matching in one picture

34

Push each student sample toward where the teacher's samples live, and away from where the student's own samples pile up.

teacher

student

real score

fake score

update = blue minus green

35 of 69

DMD moves samples by a score difference

[40] Yin et al. 2024, arXiv:2311.18828

35

fake score

trained online

real score

frozen teacher

z

G

x

gradient

on-policy, still

teacher distillation

36 of 69

DMD also needed teacher pairs

[40] Yin et al. 2024, arXiv:2311.18828

36

fake score

denoises G's samples

real score

frozen teacher

G(z)

teacher pair y

LPIPS

One step, FID

CIFAR-10

ImageNet 64

regression only

3.82

9.21

distribution only

5.58

5.61

both

2.66

2.62

mode collapse without pairs

37 of 69

DMD2 drops the pairs and adds a GAN

[1] Yin et al. 2024, arXiv:2405.14867

37

fake score

5 updates per step

real score

G

GAN head

real images

the only place real data enters

ImageNet 64, one step

FID

2.62

drop pairs

3.48

+ 5 critic updates

2.61

+ GAN head

1.51

GAN only

2.52

38 of 69

Four DMD2 steps next to fifty SDXL steps

[1] Yin et al. 2024, arXiv:2405.14867, Figs. 6 and 10 · images CC BY 4.0, cropped

38

SDXL, 50 steps, guidance

DMD2, 4 steps

50 steps

4 steps

FID 19.32 vs 19.36 · LPIPS diversity 0.61 vs 0.64

39 of 69

Moment matching in one picture

[75] Salimans et al. 2024, arXiv:2406.04103

39

Do not check each sample. At every noise level, make the student's average guess of the clean image match the teacher's.

one noise level t

teacher average

student average

match

40 of 69

Moment matching compares average predictions

[75] Salimans et al. 2024, arXiv:2406.04103, Fig. 1, cropped · CC BY-NC-SA 4.0

40

1

real image

+ noise → zₜ

2

student

x̃ = gη(zₜ)

3

sampler step

zₛ ~ q(zₛ | zₜ, x̃)

aux. denoiser

teacher

gradient

matched for every zₛ

starts from noised real data

ImageNet 64

8 steps 1.24

teacher 1.42

at 1024 steps

8-step text-to-image

41 of 69

A student can beat its teacher's score

41

teacher

student

recall drops

teacher

student

FID, lower is better

CIFAR-10

1.97

1.923

recall, higher is better

ImageNet 64

0.67

0.63

SiD, one step, against its EDM teacher

42 of 69

LADD judges images in the teacher's features

[44] Sauer et al. 2023, arXiv:2311.17042 · [45] Sauer et al. 2024, arXiv:2403.12015

42

noised image

Student, 4 steps

add noise

Teacher denoises

distillation loss

LADD drops this term

DINOv2 features, pixel space

real images

Teacher features, latent space

teacher samples

costs diversity

LADD

costs prompt alignment

Discriminator heads

adversarial loss

43 of 69

SDXL-Turbo at one, two and four steps

[44] Sauer et al. 2023, arXiv:2311.17042 · images CC BY 4.0, cropped

43

1 step

2 steps

4 steps

same seed

SDXL, 50 steps

SDXL-Turbo, 4 steps

44 of 69

What few-step images still get wrong

[74] Chen 2025, arXiv:2510.14974, Fig. 5 · [53] Zheng 2026, arXiv:2510.08431, Fig. 3 · [24] Luo 2023, arXiv:2310.04378, Fig. 5 · [6] Salimans 2022, arXiv:2202.00512, Fig. 3 · cropped

44

garbled

signs

pi-Flow, 8 NFE

FLUX Turbo, 8 NFE

wrong

digits

teacher

sCM 4-step

burnt

colour

LCM, w=4

w=12

one-step

blur

256 steps

1 step

45 of 69

Why distilled models lose variety

[63] Jiang et al. 2025, arXiv:2511.13649 · [1] Yin et al. 2024, arXiv:2405.14867

45

the loss picks a peak

reverse KL

teacher

student

guidance sharpens the target

ImageNet 256, one-step DMD

guidance

FID

recall

none

2.13

0.64

1.5

5.20

0.48

4

15.70

0.11

only guidance changed

46 of 69

New seed, same picture

[80] Wu et al. 2026, arXiv:2602.03139, Fig. 1 · panels cropped and relabelled

46

SD3.5-M teacher, 60 NFE

DMD, 4 NFE

DP-DMD, 4 NFE

One prompt, 25 seeds per grid

one face, 25 seeds

47 of 69

Newer methods keep more of the variety

47

Qwen-Image teacher

50x2 NFE

0.200

Qwen-Image-Lightning

labelled DMD2

0.095

0.150

0.182

PDD-Euler

0.192

OneIG diversity, 4 NFE, one evaluator

Base model takes step 1

DreamSim 0.264 to 0.350, base 0.337

DP-DMD's step 1, SDXL

DINOv3 0.109 to 0.173, base 0.219

48 of 69

What the big labs ship for images

48

steps

family

NVIDIA

1 to 4

trajectory

discriminator

4 to 8

trajectory

ByteDance Seed

1 to 8

discriminator

1 to 8

trajectory

Nankai, Alibaba

CDM

4

distribution

Black Forest Labs

FLUX.2 [klein]

4

the card never says how

49 of 69

Part 6

Distilling video models

Why not run the image recipe on every frame?

50 of 69

Frame by frame, the video flickers

[49] Team Wan 2025, arXiv:2503.20314 · [19] Huang et al. 2025, arXiv:2506.08009 · [50] Yin et al. 2025, arXiv:2412.07772

50

Per-frame student

flicker

Chunk-wise student

steady

frames in a chunk share attention

time

Wan2.1-1.3B takes 100 passes over 21 latent frames, 103 s per clip on one H100.

51 of 69

A causal mask lets the student stream

[50] Yin et al. 2025, arXiv:2412.07772 · [19] Huang et al. 2025, arXiv:2506.08009

51

Bidirectional teacher

Block-causal student

Frame-causal student

keys

queries

sees later frames in its chunk

KV cache

one chunk per step

then the oldest drops

52 of 69

CausVid learns to stream from a bidirectional teacher

[50] Yin et al. 2025, arXiv:2412.07772

52

Bidirectional teacher

Causal student

same weights, block-causal mask

KV cache

Real score

Fake score

DMD loss

teacher

219.2 s

student

9.4 FPS

first frame at 1.3 s

53 of 69

Four streaming steps next to fifty teacher steps

[50] Yin et al. 2025, arXiv:2412.07772 · CC BY-NC-SA 4.0

53

bidirectional teacher, 50 steps

CausVid, 4 steps

0

1.3

2.6

3.9 s

Same prompt; the student streams from 1.3 s.

54 of 69

rCM keeps Wan2.1 sharp down to two steps

[53] Zheng et al. 2026, arXiv:2510.08431 · Apache-2.0

54

1 step

2 steps

4 steps

VBench

FPS

82.65

32.3

84.09

23.0

84.43

14.6

teacher, 50 steps, CFG

83.02

0.72

One step blurs the street.

55 of 69

Why small mistakes pile up in video

55

training

next

clean data frames

streaming

error

drift

In training the student continues from clean past frames. When streaming it continues from its own, and small errors add up.

56 of 69

Exposure bias returns once the student streams

[19] Huang et al. 2025, arXiv:2506.08009 · [50] Yin et al. 2025, arXiv:2412.07772

56

training

Off-policy context

On-policy context

Teacher forcing

ground truth

Gθ

≠

Diffusion forcing

noised ground truth

Gθ

≠

its own rollout

Gθ

=

inference

Gθ

Same Wan2.1-1.3B base, VBench total: Self Forcing 84.31, CausVid's recipe 81.20.

57 of 69

Self Forcing trains on its own rollouts

[19] Huang et al. 2025, arXiv:2506.08009

57

Causal student Gθ

rollout

it conditions on its own frames

+ noise

Real score

Wan2.1-14B, frozen

Fake score

trained on rollouts

DMD loss

on the whole clip

gradient through the rollout

On one H100, 17.0 FPS and a 0.69 s first frame. VBench 84.31.

58 of 69

CausVid oversaturates, Self Forcing does not

[19] Huang et al. 2025, arXiv:2506.08009 · CC BY-NC-SA 4.0 (inferred)

58

Wan2.1-1.3B

teacher

CausVid,

4 steps

Self Forcing,

4 steps

saturated,

smoothed

One prompt, four seconds

59 of 69

Self Forcing drifts after five seconds

[61] Liu et al. 2025, arXiv:2509.25161, Fig. 4 · CC BY 4.0, cropped

59

0 s

120 s

Rolling

Forcing

Self

Forcing

drift starts here

Imaging drift over 30 s: Self Forcing 1.66, Rolling Forcing 0.01

60 of 69

Video students lose variety too

60

82

83

84

0.08

0.12

VBench total

Diversity across 5 seeds

teacher

same score, 57% of the spread

61 of 69

What the big labs ship for video

61

steps

family

NVIDIA

1 to 4

trajectory

distribution

Cosmos-Predict2.5

4

trajectory

distribution

2 to 4

trajectory

distribution

streams

ByteDance Seed

1

discriminator

1

discriminator

streams

others

4

distribution

streams

4

distribution

one step per latent frame

62 of 69

Distillation from 2015 to 2026

62

the old way

the problem

now

knowledge

distillation

Google

2015

DDPM

UC Berkeley

2020

progressive

distillation

Google

2022

guidance

distillation

Google, Stanford

consistency

models

OpenAI

2023

Stability AI

MIT, Adobe

MIT, Adobe

2024

OpenAI

Adobe, UT Austin

2025

NVIDIA, Tsinghua

Tsinghua, NVIDIA

2026

one step blurs

63 of 69

Part 7

What still breaks

What do the scores not show?

64 of 69

Problems nobody has solved yet

64

Long videos drift

Variety drops and is hard to measure

FID and GenEval disagree

One-step video still needs a discriminator

Latency numbers skip the VAE

SDXS, APT

one DMD2, three LPIPS values

65 of 69

Where I would start

65

images

1

SDXL teacher

2

3

5 fake-score

updates

4

4 steps

5

check variety

about 60 h on 64 A100s

video

1

Wan teacher

2

ODE init

3

4

rolling cache

5

check drift

about 600 iterations on 64 H100s

Give the teacher the same wall-clock before you call it a win.

66 of 69

Using a distilled model takes a few lines

[1] Yin 2024 · model card: huggingface.co/tianweiy/DMD2 · image CC BY 4.0

66

pipe = DiffusionPipeline.from_pretrained(

"stabilityai/stable-diffusion-xl-base-1.0",

torch_dtype=torch.float16, variant="fp16").to("cuda")

pipe.load_lora_weights(hf_hub_download(

"tianweiy/DMD2", "dmd2_sdxl_4step_lora_fp16.safetensors"))

pipe.fuse_lora(lora_scale=1.0)

pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)

image = pipe(prompt="a photo of a cat",

num_inference_steps=4, guidance_scale=0,

timesteps=[999, 749, 499, 249]).images[0]

LCM defaults differ from training

1

load teacher pipeline

2

swap in student weights

3

4 steps, guidance off

67 of 69

Which kind of distillation to pick

67

Family

Matches

Target from

Trained on

Weak spot

Trajectory

PD, CD, LCM, reflow

a teacher ODE point

teacher

off-policy

few-step blur, teacher ceiling

Self-distillation

CT, shortcut, MeanFlow

its own stop-gradient output

self

off-policy

from scratch, ImageNet only

Distribution

DMD2, SiD, Self Forcing

teacher score minus fake score

teacher

on-policy

mode dropping, frozen guidance

Discriminator

ADD, LADD

a discriminator's verdict

teacher

on-policy

collapse, lower diversity

on-policy still has a teacher

68 of 69

References

68

[1] Yin 2024, arXiv:2405.14867

[2] Blau 2018, arXiv:1711.06077

[3] Robbins 1956, projecteuclid.org

[4] Efron 2011, doi.org

[5] Dieleman 2024, sander.ai

[6] Salimans 2022,�arXiv:2202.00512

[7] Hinton 2015, arXiv:1503.02531

[8] Kim 2016, arXiv:1606.07947

[9] Gu 2024, arXiv:2306.08543

[10] Agarwal 2024,�arXiv:2306.13649

[11] Lu 2025, thinkingmachines.ai

[12] Lightricks 2026, huggingface.co

[13] Black Forest Labs 2024,�huggingface.co

[14] Stability AI 2024,�huggingface.co

[15] Bandyopadhyay 2025,�arXiv:2509.21318

[16] Chen 2025, arXiv:2503.09641

[17] ModelTC 2025, github.com

[18] Huang 2025, github.com

[19] Huang 2025, arXiv:2506.08009

[20] Meng 2023, arXiv:2210.03142

[21] Dieleman 2022, sander.ai

[22] Bradley 2024, arXiv:2408.09000

[23] Karras 2024, arXiv:2406.02507

[24] Luo 2023, arXiv:2310.04378

[25] Luo 2023, arXiv:2311.05556

[26] Ren 2024, arXiv:2404.13686

[27] Song 2023, arXiv:2303.01469

[28] Song 2024, arXiv:2310.14189

[29] Kim 2024, arXiv:2310.02279

[30] Lu 2025, arXiv:2410.11081

[31] Liu 2022, arXiv:2209.03003

[32] Liu 2024, arXiv:2309.06380

[33] Frans 2025, arXiv:2410.12557

[34] Geng 2025, arXiv:2505.13447

[35] Boffi 2024, arXiv:2406.07507

[36] Sabour 2025, arXiv:2506.14603

[37] Zhou 2025, arXiv:2503.07565

[38] Wang 2023, arXiv:2305.16213

[39] Luo 2023, arXiv:2305.18455

[40] Yin 2024, arXiv:2311.18828

[41] Zhou 2024, arXiv:2404.04057

[42] Zhou 2025, arXiv:2406.01561

[43] Xie 2024, arXiv:2405.16852

[44] Sauer 2023, arXiv:2311.17042

[45] Sauer 2024, arXiv:2403.12015

[46] Lin 2024, arXiv:2402.13929

[47] Lin 2025, arXiv:2501.08316

[48] Xu 2023, arXiv:2311.09257

[49] Team Wan 2025,�arXiv:2503.20314

[50] Yin 2025, arXiv:2412.07772

[51] Li 2024, arXiv:2405.18750

[52] Li 2025, arXiv:2410.05677

[53] Zheng 2026, arXiv:2510.08431

[54] HaCohen 2025,�arXiv:2501.00103

[55] Kynkäänniemi 2019,�arXiv:1904.06991

[56] Naeem 2020, arXiv:2002.09797

[57] Ghosh 2023, arXiv:2310.11513

[58] Song 2024, arXiv:2403.16627

[59] Poole 2022, arXiv:2209.14988

[60] Yang 2025, arXiv:2509.22622

[61] Liu 2025, arXiv:2509.25161

[62] Cui 2025, arXiv:2510.02283

[63] Jiang 2025, arXiv:2511.13649

[64] Lin 2025, arXiv:2506.09350

[65] Gandikota 2025,�arXiv:2503.10637

[66] Huang 2024, arXiv:2311.17982

[67] Zhang 2025, arXiv:2510.20771

[68] Lee 2025, arXiv:2510.24474

[69] Geng 2025, arXiv:2512.02012

[70] Lu 2025, arXiv:2507.18569

[71] Ge 2025, arXiv:2506.00523

[72] Chang 2025, arXiv:2506.07977

[73] Shaul 2026, arXiv:2607.26004

[74] Chen 2025, arXiv:2510.14974

[75] Salimans 2024,�arXiv:2406.04103

[76] Zhou 2024, arXiv:2410.14919

[77] Xu 2025, arXiv:2502.15681

[78] Wang 2025, arXiv:2505.20755

[79] Cheng 2025, arXiv:2512.05150

[80] Wu 2026, arXiv:2602.03139

[81] Zheng 2026, arXiv:2606.25473

[82] Zhu 2026, arXiv:2602.02214

[83] Liu 2026, arXiv:2609.14725

[84] Wang 2024, arXiv:2405.18407

[85] Duan 2026, arXiv:2609.11265

[86] Fan 2025, arXiv:2510.27684

[107] Liu 2026, arXiv:2605.06376

[108] Black Forest Labs 2026,�huggingface.co

[113] NVIDIA 2025,�arXiv:2511.00062

Example images: [1], [44] CC BY 4.0; [19], [50] CC BY-NC-SA 4.0; [32] MIT; [53] Apache-2.0; web figures as cited. Cropped and combined.

69 of 69

Thanks

shreshthsaini.github.io/blogs/diffusion-distillation.html

Questions