Improved Distribution Matching Distillation for Fast Image Synthesis
1. Paper overview
In one sentence: DMD2 accelerates image diffusion through better critic tracking, real-data supervision and student-generated training inputs, while retaining guidance, diversity and implementation tradeoffs. e02e03e05e07e10e11e12e18
| At a glance | What to know |
|---|---|
| Research problem | Source description Original DMD matches generated and teacher distributions but stabilizes training with paired noise–image regression. Producing those teacher trajectories is expensive and constrains which mapping the student learns. DMD2 asks whether distribution-level supervision can retain quality without that large paired dataset, including at SDXL scale. e02e03 |
| Core mechanism | |
| A key reported result | SDXL zero-shot COCO 2014 text-to-image synthesis: Four-step DMD2: 19.32 / 20.86 / 0.332. FID ↓; patch FID ↓; CLIP ↑. 10K random validation prompts and corresponding real images; global FID at 512×512; patch FID on 299-pixel center crops; OpenCLIP-G alignment. Teacher cfg 6, 100 forward passes: 19.36 / 21.38 / 0.332. One-step DMD2: 19.01 / 26.98 / 0.336. Four steps improve local-detail statistics over one step without improving every metric. The small global-FID teacher gap has no reported uncertainty. e09e07e19 |
| Reading caution |
Core contributions
Figure 3. Two score estimates guide the generator; real images add a separate adversarial signal. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the red generator on the left. Its image is corrupted by forward diffusion and passed to two denoisers. The lock marks the frozen teacher, which estimates the target score; the blue network learns the score of current generated samples. The red return route denotes the score-difference training signal specified by Equation (2), rather than an inference-time feedback loop. Below, real images are also noised and enter the green-highlighted shared encoder and discriminator. The blue diffusion loss trains the fake denoiser on generated data. The caption and Algorithm 2 establish that generator updates combine distribution matching with the green adversarial objective. e02e03e04e06e10e16
What it supports. The design gives distinct jobs to the frozen teacher, online fake denoiser and real-data discriminator. The fake score must track a moving distribution; the discriminator can supply information absent from the teacher approximation. This explains why faster critic updates and real-image supervision are separate interventions in the ablation.
Where the evidence stops. Equation (4) and Algorithm 2 print different fake-data terms for discriminator training. This guide follows their shared architectural description and the algorithm’s separate generator/discriminator roles; the precise loss discrepancy remains unresolved.
2. Motivation
2.1 The problem and the proposed response
Original DMD matches generated and teacher distributions but stabilizes training with paired noise–image regression. Producing those teacher trajectories is expensive and constrains which mapping the student learns. DMD2 asks whether distribution-level supervision can retain quality without that large paired dataset, including at SDXL scale. e02e03
2.2 What this reading follows
A fast student need not reproduce the exact image its teacher would generate from each noise sample. DMD2 develops that distribution-matching idea into a practical training procedure: keep an online critic close to the student’s changing outputs, add a discriminator that sees real images, and train a multi-step student on inputs produced by its own sampling process. The resulting gains depend on what is measured. ImageNet FID, COCO text alignment, high-resolution patch statistics and human preferences answer different questions. Read the mechanism beside the ablations, and keep the one-step SDXL regression warm-start in view when interpreting the paper’s removal-of-regression claim. e02e03e05e07e10e11e12e18
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Foundational work |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
This table preserves the labels recorded at reading time. The current major category is Foundational work. View the current classification.
3.1 Evidence-based assessment
Supports the recorded classification
The recorded foundational training-optimization/distillation category fits. DMD2 generates images and learns training critics; it defines no action predictor, inverse-dynamics policy or environment transition model. World–action architecture and prediction quadrants are not applicable. Shared discriminator features do not establish a unified world–action model. e03e05e06
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Why matching distributions can free the student from teacher trajectories
Original DMD already offers an escape from pairwise imitation: its generator needs the gradient of a distribution discrepancy, estimated from the difference between real and fake scores. Yet its regression regularizer still asks a particular noise input to reproduce a particular teacher-generated image. Removing that regularizer changes what solutions are permitted, but it also removes a source of stability. DMD2 therefore cannot be understood as simply deleting a loss. Its critic-update change is what makes the deletion practical in the demonstrated settings. Reader interpretation: beating a particular teacher sampler is plausible because the student is not required to reproduce that sampler’s mapping, and real-data GAN supervision supplies another signal. This is not evidence of beating every teacher configuration. The SD v1.5 SDE comparison and the one-step SDXL regression warm-start sharply qualify the broad headline. e02e03e04e10e13e18
Figure 4. Backward simulation changes where the student’s intermediate training inputs come from. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the left strip from left to right: a real image becomes increasingly noisy, illustrating the usual construction of training inputs. Read the right strip in the opposite direction, following the blue arrows from noise toward the generated image. These intermediate examples come from the student’s sampling process. Section 4.4 makes the actual process explicit: denoise, then add noise at the next scheduled level before denoising again. Section 4.5 uses those student-produced inputs during training. The essential distinction is the origin of the intermediate image, not a new conditioning modality or a different final image target. e05e06e11
What it supports. Matching the origin of training and inference inputs improves the four-step SDXL ablation: patch FID is 20.86 with backward simulation versus 24.21 without it, while both report CLIP 0.332. The improvement therefore appears in image statistics without a change in the rounded alignment score.
Where the evidence stops. The strip is schematic and does not specify gradient flow through earlier simulated steps or how an intermediate timestep is selected. Algorithm 2 also leaves its multiStepSampling helper abstract; those details cannot be recovered from this figure.
5.2 Separate critic learning from the gradient used to train the generator
There are two different learning problems inside the training loop. The generator changes the distribution of images it produces. The fake denoiser must estimate the score of that changing distribution, using detached generated images as training data. If it falls behind, the score difference used to update the generator can be inaccurate; the authors connect this to brightness oscillations and unstable FID. Increasing critic-update frequency gives that estimator more opportunities to catch up. The GAN head then adds real-versus-generated classification on shared fake-denoiser features, while the frozen teacher continues to supply its target score. These critics are training machinery, not extra sampling networks at inference. Algorithm 2 separates the generator’s non-saturating GAN loss from the discriminator objective, although its latter expression differs from Equation (4). The source leaves that implementation discrepancy unresolved. e02e03e04e06e15e16
5.3 Make intermediate training inputs resemble the inputs the student will actually see
A four-step generator is repeatedly applied to partially denoised samples. After the first step, those samples depend on the student’s own earlier outputs. Training only on noised real images therefore teaches the denoiser under a different input distribution. DMD2 simulates the student’s short sampling chain during training and applies distribution-level supervision to the resulting outputs. The fixed schedule and alternating denoising/noise injection are shared with inference. Reader interpretation: this addresses an exposure mismatch without needing a paired teacher target for every simulated input. Table 4 provides the relevant test: removing backward simulation worsens global and patch FID, while rounded CLIP stays unchanged. This supports better image statistics under the tested protocol, not universal superiority at every sampling budget. The unspecified rollout gradient handling remains a necessary reproduction decision. e05e06e11e07
5.4 Training and inference
During training
Text-to-image experiments use 3 million LAION-Aesthetics prompts and 500,000 real images. Appendix J specifies prompts from 6.25+ and GAN images from 5.5+, excluding images below 1024×1024 and unsafe content. Removing paired regression does not make the complete method data-free. e17e18
ImageNet standard training uses AdamW, learning rate 2×10⁻⁶, batch 280, GAN weight 0.003 and 200K iterations on seven A100s, approximately two days. The 1.28-FID run pretrains without GAN for 400K iterations, resumes the best-FID checkpoint, and trains another 150K with GAN at learning rate 5×10⁻⁷. e16
SD v1.5 uses 40K iterations without GAN followed by 5K with GAN. Four-step SDXL uses AdamW at 5×10⁻⁷, batch 128, teacher guidance 8 and 20K iterations on 64 A100s; approximately 60 hours is reported. One-step SDXL additionally uses conditioning timestep 399 and brief regression pretraining on 10K pairs. e17e18
During inference
Only the trained generator performs sampling; critics belong to training. Starting from Gaussian noise, repeatedly denoise and inject fresh noise at the next scheduled level. The four-step schedule is 999, 749, 499, 249; return the last denoised image. Guidance is fixed during training, limiting later user adjustment. e05e06e19
5.5 Implementation flow
- Estimate two distributions
Initialize the generator and fake denoiser from the teacher. Freeze the real-score teacher. Add noise to generated outputs and evaluate both score models there; their difference supplies the distribution-matching gradient for the generator. e02e03e06
- Track the moving generator
Train the fake denoiser by score matching on detached generated samples. Update it more frequently than the generator so its estimate can follow the changing output distribution. Five updates per generator step is the ImageNet/SDXL setting; SD v1.5 uses ten. e03e06e16e17e18
- Introduce real-data supervision
Attach a discriminator head to the fake UNet bottleneck. Its shared encoder receives real-versus-generated classification supervision after noise injection. Generator training combines distribution matching and adversarial loss; real images need no corresponding input-noise pairs. e03e04e06e16
- Train on the student sampling distribution
For a multi-step student, construct intermediate noisy inputs by running the current generator, then supervise its denoised outputs. This backward simulation replaces noisy real-image inputs that differ from those encountered during inference. e05e06
6. Experiments & results
DMD2 turns pretrained image diffusion models into one- or few-step generators by stabilizing distribution matching, adding real-image adversarial supervision, and training multi-step models on their own intermediate samples. Critic accuracy and training-input choice matter alongside the loss itself. The one-step SDXL implementation retains a small regression warm-start.
6.1 Read the original evidence
Table 1. One-step ImageNet quality improves further with a different, longer training schedule. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the middle column before the FID column: this table mixes one-pass generators with methods using hundreds of evaluations. The two bold DMD2 rows share one-pass inference but represent different training procedures. Appendix J.2 distinguishes standard training from a no-GAN pretraining stage followed by GAN finetuning for the extended result. At the bottom, keep the teacher’s ODE and SDE rows separate. Appendix K evaluates generated ImageNet images using EDM’s evaluation code and 50,000 samples. The RIN row is also relevant: it has lower FID than extended DMD2, but requires 1000 forward passes, so the table does not establish an unrestricted best-FID result. e08e07e16e19
What it supports. Standard DMD2 reaches FID 1.51 versus DMD’s 2.62. Extended training reaches 1.28, below the listed EDM ODE and SDE scores of 2.22 and 1.36. The improvement accompanies an altered training schedule; it should not be attributed to extra iterations as an isolated causal factor.
Where the evidence stops. Forward-pass counts are not measured wall-clock speedups. Training budgets differ across rows, and the paper reports no error bars. The small extended-DMD2 versus SDE-teacher gap is a reported point estimate.
Table 2. Four-step DMD2 improves patch statistics while global FID and text alignment tell a different story. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow each row across all three metrics instead of choosing the boldest single entry. Lower FID and patch FID are preferred; higher CLIP indicates stronger measured text alignment. The one- and four-step DMD2 rows are distinct trained models. The teacher rows vary classifier-free guidance and count 100 forward passes; Section 5.2 distinguishes those passes from its 50 sampling steps. Appendix K uses 10K random COCO validation prompts, corresponding reference images and 512-pixel outputs for global FID. Patch FID examines 299-pixel center crops, while CLIP uses OpenCLIP-G. These evaluations emphasize different properties of the same generator. e09e07e12e18e13
What it supports. Four-step DMD2 reports 19.32 FID, 20.86 patch FID and 0.332 CLIP, close to the cfg-6 teacher’s 19.36, 21.38 and 0.332. Its patch FID improves over one-step DMD2’s 26.98 even though the one-step model has slightly better global FID and CLIP.
Where the evidence stops. The one-step SDXL model uses a 10K-pair regression warm-start and timestep shift. Its comparison with the four-step model changes training as well as sampling. COCO numbers from the separate SD v1.5 protocol are not directly comparable.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| ImageNet-64×64 class-conditional image synthesis 50,000 generated images; EDM evaluation code; standard and extended training reported separately. | Standard DMD2: 1.51; extended training: 1.28; both one forward pass. FID ↓; forward passes ↓ | DMD: 2.62. EDM teacher: 2.22 with ODE or 1.36 with SDE, each 511 forward passes. Extended training beats both listed teacher samplers by FID; standard training does not beat the SDE teacher. Training budgets differ. e08e16e07 |
| SD v1.5 zero-shot COCO 2014 text-to-image synthesis 30K validation prompts; generated images downsampled to 256×256; clean-FID against 40,504 real validation images. | One-step DMD2: 8.35. FID ↓ | DMD: 11.49; SD v1.5 ODE teacher, 50 steps/cfg 3: 8.59; SDE teacher, 200 steps/cfg 2: 7.21. A 3.14-point improvement over DMD, but teacher superiority depends on sampler and guidance configuration. e13e07 |
| SDXL zero-shot COCO 2014 text-to-image synthesis 10K random validation prompts and corresponding real images; global FID at 512×512; patch FID on 299-pixel center crops; OpenCLIP-G alignment. | Four-step DMD2: 19.32 / 20.86 / 0.332. FID ↓; patch FID ↓; CLIP ↑ | Teacher cfg 6, 100 forward passes: 19.36 / 21.38 / 0.332. One-step DMD2: 19.01 / 26.98 / 0.336. Four steps improve local-detail statistics over one step without improving every metric. The small global-FID teacher gap has no reported uncertainty. e09e07e19 |
| ImageNet component ablation Table 3, one-step ImageNet-64×64 models. | DMD 2.62 → remove regression 3.48 → add TTUR 2.61 → add GAN 1.51. FID ↓ | GAN alone: 2.56; GAN plus TTUR: 2.52. TTUR restores quality lost by removing regression; real-data GAN supervision contributes the next gain. GAN-only TTUR has a small reported change. e10 |
| SDXL four-step component ablation Table 4, 10K COCO 2014 prompts. | Full model: 19.32 / 20.86 / 0.332. FID ↓ / patch FID ↓ / CLIP ↑ | Without GAN: 26.90 / 27.66 / 0.328; without distribution matching: 13.77 / 27.96 / 0.307; without backward simulation: 20.66 / 24.21 / 0.332. Global FID alone favors the GAN-only model despite worse alignment and local detail. Backward simulation improves both FIDs at unchanged rounded CLIP. e11e07 |
| Human preference against the SDXL teacher 128 PartiPrompts from the LADD subset; five evaluators per comparison; four student versus 50 teacher sampling steps. | DMD2: 62.0% image quality; 50.5% prompt alignment. Pairwise preference share | Teacher: 38.0% image quality; 49.5% prompt alignment. The quality margin is 24 percentage points, not a 24% win rate. The figure supports nearly tied alignment; this protocol differs from COCO FID. e12 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Tables 3 and 4. Removing components reveals both optimization gains and metric tradeoffs. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The left table is ImageNet. Read its first four rows cumulatively: DMD, removal of regression, addition of the two-time-scale update rule, then addition of GAN supervision. The bottom two rows are separate GAN-only controls, with and without TTUR. The right table concerns four-step SDXL on 10K COCO prompts; compare each component-removal row with the full model at the bottom. FID and patch FID point downward, while CLIP points upward. The bold 13.77 FID belongs to the model without distribution matching. Its neighboring patch FID and CLIP values prevent interpreting that number as an across-the-board improvement. e10e11e07e19
What it supports. On ImageNet, TTUR changes FID from 3.48 to 2.61 after regression removal; adding GAN yields 1.51. On SDXL, removing distribution matching produces the best global FID here but worsens patch FID to 27.96 and CLIP to 0.307. The combined objective balances properties that one metric alone misses.
Where the evidence stops. The left and right tables use different backbones and evaluation settings. Neither reports uncertainty. GAN-only TTUR changes ImageNet FID from 2.56 to 2.52, so the prose claim of no improvement should be read as no substantial demonstrated gain.
Figure 10. Extra critic updates can improve convergence per hour, but more updates are not always better. Original paper, p. 20 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The horizontal axis is training hours, not generator steps; lower curves indicate better FID. Dark blue is original DMD with regression. Red removes regression while retaining one fake update per generator update, and shows conspicuous instability. Green uses five fake updates and magenta ten. Cyan keeps the asynchronous-learning-rate alternative, described in the caption as a fivefold larger fake-model learning rate. The legend’s iteration value denotes the critic-update ratio, not the number of inference sampling steps. Appendix G connects these trajectories to the separate brightness diagnostic in Figure 9 and recommends tuning the ratio for new models or datasets. e15e03e17e19
What it supports. The five-update setting reaches lower FID sooner in the displayed run than the ten-update and asynchronous-learning-rate settings. This supports the paper’s practical choice of giving the critic time to track the generator, while showing that allocating still more updates can slow progress.
Where the evidence stops. The plot does not directly measure fake-score estimation error or establish a universal optimal ratio. Hardware-normalized comparisons and repeated runs would be needed for broader claims; SD v1.5 uses ten updates in the paper’s own recipe.
7. Analysis & limitations
7.1 What the evidence leaves open
Diversity is reduced: four-image-per-prompt LPIPS diversity is 0.61 versus the SDXL teacher’s 0.64. The authors also identify fixed guidance, four SDXL steps for their quality target, and substantial training compute. e14e19
The checklist explicitly reports no error bars. Small differences should not be treated as statistically established improvements, and selected qualitative comparisons cannot measure failure prevalence. e19e20
Equation (4) uses −log D(fake) in the stated discriminator-maximization objective, whereas Algorithm 2 line 32 uses log(1−D(fake)). Algorithm 2 separately gives the generator’s non-saturating loss. This discrepancy remains unresolved in the supplied PDF. e04
7.2 Questions for discussion
- How much of TTUR’s benefit remains when critic compute and wall-clock budget are controlled?
- Does backward simulation still help when rollout gradient handling and training-step sampling are independently controlled?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs the specified EDM/SD checkpoints, backbone-specific update ratios, training stages and evaluation image sets. Appendix J omits the one-step SDXL warm-start duration and SDXL GAN coefficient; the unsafe-content filter implementation and exact sampled data identities are also absent. e16e17e18
Algorithm 2 detaches fake images for critic updates, but multiStepSampling does not specify gradient flow through earlier rollout steps or intermediate-step selection. Resolve these choices and the GAN-loss discrepancy before interpreting implementation differences as replication failures. e06e04
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Does critic tracking improve stability at a matched compute budget?
Reader-proposed, not performed: start from the same EDM checkpoint on ImageNet-64×64 and remove regression while holding data, optimizer, batch size and GAN setting fixed. Compare one, five and ten fake updates per generator update, plus the paper’s fivefold fake-learning-rate control. Repeat with matched random seeds and record pixel-brightness variability, held-out fake-denoising loss and FID against both elapsed time and generator-update count. Treat denoising loss as a tracking proxy, not a measured true-score error. The tracking explanation predicts that improved critic fit accompanies reduced oscillation. If only equal-generator-update comparisons improve, the claimed efficiency benefit would need qualification. e03e10e15e16e07
Check 2: Does backward simulation help after other training choices are controlled?
Reader-proposed, not performed: train two four-step SDXL students from identical initial weights using the same prompts, GAN images, guidance, losses, update ratio and timestep schedule. Change only intermediate-input construction: noised real images versus current-student sampling. Declare rollout gradient handling and step-selection probabilities explicitly, and add a matched-wall-clock comparison to expose the simulation overhead. Evaluate both on the same 10K COCO prompts with global FID, patch FID and OpenCLIP-G; assess diversity separately using four samples per PartiPrompt and pairwise LPIPS. Repeated improvement in patch FID without degraded alignment or diversity would strengthen the input-mismatch explanation; disappearance of the effect under these controls would weaken it. e05e06e11e18e07e14
8.3 Reading coverage
Visual audit: Visually inspected the title/authors/affiliations on page 1; equations and mechanism on pages 5–7; all Tables 1–7; Figures 3–10 and 14; Algorithm 2; limitations, diversity, implementation, evaluation and uncertainty disclosures on the declared pages. Six final original crops were inspected individually with labels, axes and legends retained. Figure 3’s frozen-teacher marker and score route were checked against Equations (1)–(2) and Algorithm 2; Figure 4’s opposing arrow directions were checked against Sections 4.4–4.5. Equation (4) and Algorithm 2 contain different discriminator fake-data terms, disclosed in the report and architecture caution. Text reading covers all 33 pages; gallery images in Figures 1–2 and 11–13 were not visually inspected. Separate supplements, code and later editions were not inspected.
PDF pages inspected for this edition: 1, 5, 6, 7, 8, 9, 10, 15, 16, 17, 18, 19, 20, 24, 25, 30. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract; Sections 1–2: Introduction and Related Work
- Section 3: Background, Equations (1)–(3)
- Sections 4.1–4.6: Improved Distribution Matching Distillation
- Sections 5.1–5.3: Experiments and ablations
- Section 6: Acknowledgements; References
- Appendices A–I: limitations, release statement, broader impact, additional results, algorithms and diagnostics
- Appendices J.1–J.4, K and L: implementation, evaluation and human study
- Appendix M: gallery prompts
- NeurIPS checklist, items 1–15; all eight supplied text chunks spanning PDF pages 1–33
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- The extraction limitation was addressed for the declared inspected pages and six original crops. Figures 1–2 and 11–13 were read through their captions and associated text, but their gallery images were not visually inspected.
- Separate supplemental material availability has not been fully verified.
- Code, project website and external assets were not inspected; no experiments were reproduced.
- Version scope: the supplied 33-page NeurIPS 2024 proceedings PDF matches the catalog title and all seven authors. No numbered revision is stated. Appendix B mentions continuing updates, but no other edition or revision chain was supplied or compared.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block and venue footer; p. 15, Appendix B
The title and seven authors match the supplied catalog: Tianwei Yin; Michaël Gharbi; Taesung Park; Richard Zhang; Eli Shechtman; Frédo Durand; William T. Freeman. Affiliations are Massachusetts Institute of Technology and Adobe Research. The footer identifies NeurIPS 2024; Appendix B mentions continuing updates without identifying a numbered revision.
Go to primary source ↓e02PDF p. 5, Section 3, Equations (1)–(3) and accompanying definitions
DMD estimates target and generated-data scores with a frozen teacher and online fake denoiser. The expected diffused KL gradient uses their score difference. Original DMD adds paired noise–image regression using teacher ODE outputs.
Go to primary source ↓e03PDF p. 6, Figure 3 and Sections 4.1–4.3
The diagram shows a few-step generator, frozen real-score network, trainable fake-score network and shared GAN branch. The text proposes removing regression and using five fake-score updates per generator update, attributing instability to inaccurate fake scores.
Go to primary source ↓e04PDF p. 7, Section 4.3, Equation (4); p. 18, Algorithm 2, lines 19–23 and 31–33
The GAN branch classifies noised real and generated images. Equation (4) prints a negative log D(fake) term in the stated discriminator objective. Algorithm 2 line 32 instead prints log(1−D(fake)); line 19 separately gives negative log D(fake) for the generator.
Go to primary source ↓e05PDF p. 7, Sections 4.4–4.5 and Figure 4
Multi-step inference alternates denoising and scheduled noise injection, using timesteps 999, 749, 499, 249 for four steps. Training simulates current-student intermediate inputs. Figure 4 contrasts rightward forward diffusion from real images with leftward student sampling from noise.
Go to primary source ↓e06PDF p. 18, Figure 7, Algorithms 1–2; Algorithm 2 lines 2–4, 8–9, 15–33
Generator and fake model copy teacher weights. Generator updates are less frequent; critic denoising and classification use detached generated outputs. The multi-step sampling helper is abstract and does not specify rollout gradient handling or step-selection details.
Go to primary source ↓e07PDF p. 8, Section 5 metric definitions; p. 25, Appendix K
ImageNet evaluation uses 50,000 generated images and EDM code. SD v1.5 uses 30K COCO validation prompts, 256-pixel outputs and 40,504 real images with clean-FID. SDXL uses 10K prompts, corresponding references and 512-pixel outputs; CLIP uses OpenCLIP-G. Patch FID uses 299-pixel center crops.
Go to primary source ↓e08PDF p. 8, Table 1, DMD/DMD2/longer-training/EDM/RIN rows
One-pass DMD is 2.62 FID; DMD2 is 1.51 and extended DMD2 is 1.28. EDM ODE/SDE teachers have FID 2.22/1.36 at 511 forward passes. The 1000-pass RIN row is 1.23.
Go to primary source ↓e09PDF p. 8, Table 2, DMD2 and SDXL teacher rows
One-step DMD2 reports FID/patch FID/CLIP 19.01/26.98/0.336; four-step reports 19.32/20.86/0.332. Teacher cfg 6 reports 19.36/21.38/0.332 at 100 forward passes; cfg 8 reports 20.39/23.21/0.335.
Go to primary source ↓e10PDF p. 9, Table 3, all rows and Section 5.3 discussion
ImageNet FIDs progress from DMD 2.62 to 3.48 without regression, 2.61 with TTUR, and 1.51 with GAN. GAN-only rows are 2.56 without and 2.52 with TTUR.
Go to primary source ↓e11PDF p. 9, Table 4, all rows; p. 10, Section 5.3 continuation
SDXL full model reports 19.32/20.86/0.332. No-GAN gives 26.90/27.66/0.328, no-distribution-matching 13.77/27.96/0.307, and no-backward-simulation 20.66/24.21/0.332. The discussion links pure GAN training to worse prompt alignment and no backward simulation to worse local detail.
Go to primary source ↓e12PDF p. 9, Section 5.2 and Figure 5, SDXL-Teacher comparison bars; p. 25, Appendix L
The study uses 128 prompts, five evaluators per comparison, and randomized left/right presentation. Figure 5 shows 62.0% versus 38.0% quality preference and 50.5% versus 49.5% alignment. Its caption specifies four distilled versus 50 teacher sampling steps; the text counts four versus 100 forward passes.
Go to primary source ↓e13PDF p. 16, Table 5, DMD/DMD2 and SD v1.5 teacher rows
DMD2 reports 8.35 FID on 30K COCO prompts versus DMD 11.49, 50-step ODE/cfg 3 teacher 8.59, and 200-step SDE/cfg 2 teacher 7.21.
Go to primary source ↓e14PDF p. 17, Table 6, DMD2 and teacher rows; Appendix F protocol paragraph
Diversity is mean pairwise LPIPS for four generated images per prompt on the LADD PartiPrompts subset. DMD2 scores 0.61 and both teacher guidance settings score 0.64; the adjoining FID/CLIP values use the separate COCO protocol.
Go to primary source ↓e15PDF p. 17, Appendix G; p. 20, Figures 9–10 and captions
One fake update without regression produces brightness oscillations and unstable FID. Five updates stabilize training and improve convergence versus the displayed one-update, ten-update and asynchronous-learning-rate variants. Figure 10 plots training hours; the learning-rate control uses a fivefold fake-model rate.
Go to primary source ↓e16PDF p. 24, Appendices J.1–J.2
The discriminator uses convolutional downsampling, group normalization, SiLU and final linear projection from fake-model middle-block features. Standard ImageNet training uses AdamW at 2×10⁻⁶, batch 280, seven A100s, 200K iterations and GAN weight 0.003. Extended training pretrains without GAN for 400K iterations, resumes the best-FID checkpoint and adds 150K iterations at 5×10⁻⁷.
Go to primary source ↓e17PDF p. 9, Section 5.2 training-data paragraph; p. 24, Appendix J.3
Text-to-image training uses 3M prompts and 500K real images. SD v1.5 uses LAION-Aesthetic 6.25+ prompts and filtered 5.5+ images, teacher guidance 1.75, ten fake updates per generator update, batch 2048 and 64 A100s. Training has 40K no-GAN iterations then 5K with GAN weight 0.001 and reduced learning rate.
Go to primary source ↓e18PDF p. 24, Appendix J.4; p. 25, Appendix J.4 continuation
SDXL uses filtered LAION prompts/images, AdamW at 5×10⁻⁷, five fake updates, guidance 8, batch 128 and 64 A100s. Four-/one-step training uses 20K/25K iterations; approximately 60 hours is reported. One-step SDXL additionally uses conditioning timestep 399 and a brief regression warm-start on 10K pairs, whose duration is not quantified. The SDXL GAN coefficient and exact sampled data identities are not specified here.
Go to primary source ↓e19PDF p. 15, Appendix A; p. 30, checklist item 7
The authors acknowledge reduced diversity, fixed guidance, four steps for their SDXL quality target and high compute costs. The checklist explicitly answers No to reporting error bars, attributing the omission to training cost.
Go to primary source ↓e20PDF p. 10, Figure 6 and caption; p. 19, Figure 8 and caption
The selected qualitative comparisons use identical noise and prompts within comparisons. Figure 8 shows full DMD2 beside no-distribution-matching, no-GAN and no-backward-simulation variants; the authors describe oversaturation/smoothing without GAN and reduced alignment without distribution matching.
Go to primary source ↓8.5 Primary sources
Improved Distribution Matching Distillation for Fast Image Synthesis ↗
PDF · 14,433 extracted words
Source fingerprint
5992974f18e97081934680f40f9dce2c747bad605f0b63c2e698ddddc4bc9368