PAPER REPORTENAll readings ↗

High-Resolution Image Synthesis With Latent Diffusion Models

English reading report: Method, equations, original figures, experiments and reproducibility.

Authors: Robin Rombach; Andreas Blattmann; Dominik Lorenz; Patrick Esser; Björn Ommer

Affiliations: Ludwig Maximilian University of Munich & IWR, Heidelberg University, Germany; Runway ML

Source: CVPR 2022 · ref-07638f74c59962f3e906 ↗ · Catalog record

Reading: 496 / 558 · 5 original figures & tables · ~18 min ·

1. Paper overview

In one sentence: Moving iterative diffusion into a mildly compressed image space reduces computation, while reconstruction quality and sampling guidance determine what fidelity and coverage can be retained. e-identitye-revisione-motivatione-autoencodere-compressione-imagenete-limitations

At a glanceWhat to know
Research problem
Source description

Pixel-space diffusion repeatedly evaluates a large network on details that may contribute little to perception. The paper asks whether a reusable, mildly compressed representation can reduce this cost while preserving spatial structure for high-quality synthesis. Its target is image generation and restoration, with no action-conditioned dynamics or executed control. e-motivatione-latente-conditioning

Core mechanism
Source description

Separate autoencoder training from diffusion training so reconstruction and prior learning do not require a joint compromise; reuse the representation across generative tasks. e-motivatione-autoencoder

A key reported resultClass-conditional ImageNet synthesis: FID 3.60; IS 247.67 ± 5.59; precision 0.87; recall 0.48.

FID ↓; Inception Score ↑; precision/recall ↑. 256×256; 50K generated samples versus training-set statistics; updated LDM-4, 250 DDIM steps, classifier-free guidance scale 1.5.

Unguided LDM-4: FID 10.56, recall 0.62. ADM-G: FID 4.59, recall 0.52. Table 3 reports 400M versus 608M parameters. Guidance improves FID while sacrificing coverage. Latent space, guidance, training and architecture differ across baselines. e-imagenete-evaluation

Reading caution
Source description

Sequential denoising remains slower than GAN sampling. Autoencoder reconstruction can limit fine pixel accuracy; the authors identify super-resolution as a likely example. e-limitations

Core contributions

  • Source description

    Separate autoencoder training from diffusion training so reconstruction and prior learning do not require a joint compromise; reuse the representation across generative tasks. e-motivatione-autoencoder

  • Source description

    Introduce cross-attention for flexible conditioning and retain concatenation for aligned image inputs. Moderate downsampling improves the tested quality–speed balance over pixel diffusion or severe compression. e-conditioninge-compression

Figure 3. Compression, iterative denoising and conditional information occupy separate parts of the generator. Original paper, p. 4 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Start at the left: the encoder E maps image x to latent z. The upper arrow runs toward the noisy endpoint z_T. Generation follows the lower path in the opposite direction, repeatedly applying the denoising UNet before D converts the final latent into an image. The loop explicitly marks repeated denoising; D is used once in ordinary sampling. Follow the condition encoder tau_theta along the bottom: its features supply K and V to the yellow attention blocks, while Q comes from UNet features. The switch also depicts the alternative concatenation route. Gray dashed arrows are UNet skip connections, not additional diffusion trajectories. e-autoencodere-latente-conditioninge-conditioner-config

What it supports. The computational saving comes from moving repeated denoising into a smaller spatial representation. Cross-attention lets the denoiser use a different kind of input, such as text, without making that input an image grid. The diagram describes a family of task-specific generators, rather than evidence that every condition is handled by one jointly trained universal model.

Where the evidence stops. The arrows agree with the Q/K/V assignments in Section 3.3. However, the printed Q and V projection-matrix dimensions across pp. 4–5 conflict with those assignments. This faithful crop does not resolve that notation error.

2. Motivation

2.1 The problem and the proposed response

Source description

Pixel-space diffusion repeatedly evaluates a large network on details that may contribute little to perception. The paper asks whether a reusable, mildly compressed representation can reduce this cost while preserving spatial structure for high-quality synthesis. Its target is image generation and restoration, with no action-conditioned dynamics or executed control. e-motivatione-latente-conditioning

2.2 What this reading follows

A diffusion model can spend substantial effort modeling pixel detail before it learns the broader composition of an image. Rombach and colleagues separate those jobs: train an autoencoder to preserve perceptually useful content, then learn a diffusion prior over its spatial latents. This reading follows the resulting information flow and tests the tradeoff against the paper’s own reconstruction table, compression sweep and conditional generation results. The supplied source is arXiv v2, whose changelog updates the text and ImageNet models. Its evidence supports a versatile image generator, while leaving important limits on exact reconstruction, benchmark comparability and reproduction of some printed configurations. e-identitye-revisione-motivatione-autoencodere-compressione-imagenete-limitations

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 dimensionRecorded classification
Major categoryFoundational work
ArchitectureNot applicable
Prediction paradigmNot applicable
QuadrantNot applicable

This table preserves the labels recorded at reading time. The current major category is Components of WAMs. View the current classification.

3.1 Evidence-based assessment

Supports the recorded classification

Reader analysis

The foundational-work label and Not applicable architecture, prediction-paradigm and quadrant entries fit this image-generation method. It predicts image latent noise, with no action head, inverse dynamics or execution loop. Video-backbone relevance is a catalog-level judgment: this PDF demonstrates still images. Joint denoiser/conditioner training does not establish a One Model world-action architecture. e-latente-conditioninge-spatial

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

InputsOutputs
  • RGB training images; optionally paired text, classes, layouts, semantic maps, low-resolution or masked images
  • Gaussian latent noise and optional task conditioning at generation time
  • Generated RGB images, including super-resolved and inpainted images; no predicted actions

4.2 Equations and their role

z=E(x),x~=D(z),f=H/h=W/wz=\mathcal{E}(x),\quad \tilde{x}=\mathcal{D}(z),\quad f=H/h=W/w
The RGB image x has height H and width W; z is an h-by-w grid with c channels. E and D denote encoder and decoder; f is spatial downsampling. Larger f reduces the grid but can discard detail. e-autoencoder
LLDM=EE(x),y,ϵN(0,1),t ⁣[ϵϵθ(zt,t,τθ(y))22]L_{\mathrm{LDM}}=\mathbb{E}_{\mathcal{E}(x),y,\epsilon\sim\mathcal{N}(0,1),t}\!\left[\left\|\epsilon-\epsilon_\theta(z_t,t,\tau_\theta(y))\right\|_2^2\right]
Equation (3) predicts noise: z_t is a noisy encoded image, epsilon is sampled noise, t is uniformly sampled over diffusion steps, y is the condition and tau_theta encodes it. The denoiser epsilon_theta and conditioner are optimized; the image autoencoder stays fixed. e-latente-conditioninge-motivation
Attention(Q,K,V)=softmax ⁣(QKTd)V\operatorname{Attention}(Q,K,V)=\operatorname{softmax}\!\left(\frac{QK^{\mathsf T}}{\sqrt d}\right)V
Q projects UNet features; K and V project the condition. Attention dimension d scales the dot products. The printed Q/K/V assignments are clear, but adjacent Q and V projection dimensions on pp. 4–5 conflict with those assignments; corrected shapes are not asserted here. e-conditioning

5. Method in detail

5.1 Choose the latent space before learning its distribution

Source description

The two training stages solve different problems. First, the autoencoder must reconstruct an observed image well enough that useful visual information survives compression. Its perceptual and adversarial losses encourage realistic reconstructions, while weak regularization keeps the representation usable. Only afterward does diffusion learn the distribution of those representations, with the autoencoder fixed. For VQ regularization, the diffusion model still sees continuous latents before quantization; the decoder performs quantization when returning to pixels. This matters because the method is not predicting a sequence of discrete codebook indices. Table 8 assesses the first stage, and Figure 7 assesses the combined generator. Reading them together explains why aggressive compression may speed individual steps yet leave the generative model with an inadequate representation. e-autoencodere-latente-motivatione-reconstructione-compression

5.2 Condition each denoising step without confusing training and sampling

Source description

For text conditioning, the tokenizer and transformer turn a prompt into features that the UNet can consult through cross-attention. At an intermediate UNet layer, visual features produce queries and the condition features produce keys and values. The denoising objective trains the condition encoder together with the noise predictor, while the image autoencoder remains fixed. Spatial inputs such as low-resolution images can instead be concatenated because they already align with the output grid. During ordinary generation, repeated UNet evaluations refine a noisy latent before one final decoding pass. Appendix C describes a different, optional mechanism: post-hoc image guidance differentiates an image comparison through the decoder during sampling. That inference-time computation should not be confused with the perceptual loss used to train the autoencoder. e-conditioninge-conditioner-confige-motivatione-spatiale-latente-image-guidance

5.3 Read the operating point, not just the headline score

Reader analysis

Reader analysis: the strongest interpretation separates representation, sampling and evaluation choices. Figure 7 varies compression and denoising steps, so a point combines architectural and runtime decisions. Table 3 then shows that guidance can improve FID without changing the reported LDM parameter count, while recall declines. Table 7 adds another variable: extra training at the final inpainting resolution substantially improves FID. None of these observations means that every perceptual metric or task benefits equally. Super-resolution illustrates the boundary directly: LDM’s reported FID is favorable while SR3 has higher PSNR. A reproduction should therefore select a checkpoint, guidance scale, sample count, reference split and resolution protocol before comparing scores. Otherwise a change in operating point can be mistaken for evidence about the latent representation itself. e-compressione-sweep-protocole-imagenete-inpaintinge-inpainting-finetunee-superresolutione-evaluation

5.4 Training and inference

During training

Source description

Train autoencoders on OpenImages, then jointly optimize the denoiser and condition encoder on task data. The KL penalty is approximately 10^{-6}; Appendix G specifies sampling KL latents and optionally normalizing them using a first-batch standard-deviation estimate. e-autoencodere-reconstructione-conditioninge-scale

Source description

Text training uses LAION-400M; image generation uses ImageNet, face datasets and LSUN; inpainting uses Places. Conditional configurations specify 1,000 diffusion steps and a linear schedule. The updated ImageNet run uses 178K iterations and batch size 1,200, distinct from the compression sweep. e-texte-unconditionale-inpaintinge-hyperparameterse-imagenet

During inference

Source description

Start with Gaussian latent noise, repeatedly denoise using a sampler such as DDIM, then decode the final latent once. One decoder pass does not make the generator a one-step model. Classifier-free guidance changes the sampling operating point. e-latente-imagenet

Source description

Convolutional sampling can exceed training resolution, but latent scale affects coherence and the larger inpainting model needs resolution-specific fine-tuning. Optional post-hoc image guidance differentiates image similarity through the decoder during sampling, unlike the training-only reconstruction loss. e-spatiale-scalee-inpainting-finetunee-image-guidance

5.5 Implementation flow

  1. Learn a perceptual representation

    An encoder maps RGB images to a spatial grid. Perceptual reconstruction and a patch discriminator encourage local realism. Weak KL regularization or vector quantization constrains the latent distribution while permitting mild spatial compression. e-autoencoder

  2. Learn a latent noise predictor

    With the autoencoder fixed, corrupt encoded images using the forward diffusion process. A time-conditioned convolutional UNet predicts the added Gaussian noise. VQ models diffuse continuous pre-quantization latents; quantization belongs to the decoder. e-latente-autoencoder

  3. Route the condition into denoising

    Cross-attention takes queries from UNet features and keys and values from the condition encoder. Text uses a BERT tokenizer and learned unmasked transformer; classes use an embedding. Aligned image conditions can instead be concatenated to the UNet input. e-conditioninge-conditioner-confige-texte-spatial

6. Experiments & results

Latent diffusion separates perceptual compression from generative modeling: a pretrained autoencoder supplies a spatial latent grid, a diffusion UNet learns to denoise it, and a decoder returns an image. Cross-attention adds text, class or layout conditioning. Experiments favor moderate compression and demonstrate image synthesis and editing, while exposing a reconstruction bottleneck and tradeoffs between distributional quality, coverage and pixel fidelity.

6.1 Read the original evidence

Table 3. Guidance improves ImageNet FID while changing the model’s coverage of the data distribution. Original paper, p. 7 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Begin with the two LDM rows, which make the role of guidance easier to see than a comparison across architectures. Both use 250 sampling steps. The G row applies classifier-free guidance, abbreviated c.f.g., with scale s=1.5. FID is better when lower; Inception Score, precision and recall are better when higher, but the columns need not move together. Then compare with ADM-G and note its larger reported parameter count. Appendix E specifies 50K generated samples and training-set reference statistics for these generation metrics. Appendix D identifies the updated LDM training run as 178K iterations with batch size 1,200. e-imagenete-evaluatione-hyperparameters

What it supports. LDM-4 moves from FID 10.56 without guidance to 3.60 with guidance, while precision rises from 0.71 to 0.87 and recall falls from 0.62 to 0.48. Its guided FID is below ADM-G’s 4.59, but ADM-G retains higher recall, 0.52. A single best FID therefore hides a meaningful coverage tradeoff.

Where the evidence stops. Training, guidance and architecture differ across methods, and the table does not provide FID confidence intervals. Its LDM parameter count is 400M; Table 15 instead lists 395M, an unresolved configuration discrepancy.

Table 7. Resolution fine-tuning improves inpainting FID, while the reconstruction metric favors a different method. Original paper, p. 9 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Keep the two column groups separate: the left group contains difficult examples with 40–50% of the image masked, and the right group aggregates all test examples. Both FID and LPIPS are lower-is-better measures. Compare the first two rows to examine the large LDM before and after fine-tuning, denoted ft. The smaller-model attention labels concern the first-stage autoencoder; the large diffusion UNet itself uses attention. For the direct LaMa comparison, use the daggered row, recomputed on the authors’ test set. The retained note explains why: the original LaMa test set was unavailable. Evaluation uses 30K Places crops at 512×512. e-inpaintinge-inpainting-finetunee-inpainting-efficiency

What it supports. Fine-tuning the large LDM for half an epoch at 512 resolution lowers overall FID from 2.40 to 1.50, and hard-subset FID from 12.89 to 9.39. Recomputed LaMa has overall FID 2.23, but its LPIPS of 0.134 is slightly lower than LDM’s 0.137. Distributional realism and similarity to the particular target image remain different criteria.

Where the evidence stops. The fine-tuned model also receives extra training at the evaluation resolution. Other LDM rows change capacity or first-stage attention, so the table cannot attribute every improvement to latent diffusion alone. The reported ± values are not identified as run-to-run confidence intervals.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
Class-conditional ImageNet synthesis

256×256; 50K generated samples versus training-set statistics; updated LDM-4, 250 DDIM steps, classifier-free guidance scale 1.5.

FID 3.60; IS 247.67 ± 5.59; precision 0.87; recall 0.48.

FID ↓; Inception Score ↑; precision/recall ↑

Unguided LDM-4: FID 10.56, recall 0.62. ADM-G: FID 4.59, recall 0.52. Table 3 reports 400M versus 608M parameters.

Guidance improves FID while sacrificing coverage. Latent space, guidance, training and architecture differ across baselines. e-imagenete-evaluation

Text-to-image synthesis

LAION-400M training; 256×256 MS-COCO validation evaluation with 30K reference samples; 250 DDIM steps.

LDM-KL-8-G: 12.63 at guidance scale 1.5; 1.45B parameters.

FID ↓

Unguided LDM-KL-8: 23.31; GLIDE: 12.24 with 6B parameters; Make-A-Scene: 11.84 with 4B.

Competitive FID at fewer reported parameters, not the table’s best FID. Baseline training and sampling differ. e-text

Unconditional CelebA-HQ synthesis

256×256; 50K generated samples versus training-set statistics; LDM-4, 500 DDIM steps.

FID 5.11; precision 0.72; recall 0.49.

FID ↓; precision/recall ↑

LSGM: FID 7.22; PGGAN: 8.0.

Strong on this dataset; other Table 1 datasets contain GAN baselines with lower FID. e-unconditionale-evaluation

Layout-to-image synthesis

COCO 256×256; 2,048 Segmentation Challenge examples; LDM-4 pretrained on OpenImages and fine-tuned on COCO, 200 steps.

40.91.

FID ↓

SPADE: 41.11; COCO-only LDM-8 at 100 steps: 42.06.

Pretraining, compression and sampling budget all change; this is not a clean pretraining ablation. e-layout

ImageNet fourfold super-resolution

64→256; antialiased bicubic degradation; ImageNet validation outputs; LDM-4, 100 steps.

FID 2.8 against validation reference features, or 4.8 against training features; PSNR 24.4 ± 3.8.

FID ↓; PSNR ↑

SR3: FID 5.2, PSNR 26.4; image regression: PSNR 27.9.

Distributional quality and aligned pixel accuracy disagree. The two LDM FIDs use different reference sets. e-superresolution

Places image inpainting

30K test crops at 512×512 with synthetic masks; large LDM-4 after half an epoch of 512-resolution fine-tuning.

Overall: FID 1.50, LPIPS 0.137 ± 0.080. With 40–50% masked: FID 9.39.

FID ↓; LPIPS ↓

Recomputed LaMa: FID 2.23, LPIPS 0.134 ± 0.080 overall; hard-subset FID 12.31. Large LDM without fine-tuning: overall FID 2.40.

Fine-tuning helps FID; LaMa retains slightly better LPIPS. Its original test set was unavailable. e-inpaintinge-inpainting-finetune

Compression versus generation efficiency

ImageNet sweep with comparable model sizes on an A100; 2M iterations; 5K-sample quality estimates.

Section 4.1 reports a 38-point FID gap favoring LDM-8 over pixel-space LDM-1 after 2M iterations.

FID ↓ and sample throughput ↑

Figures 6–7 favor moderate compression; LDM-32 loses quality and LDM-1 is slow.

Batches and learning rates vary: equal iterations do not imply equal examples or compute. Figure 17 separately fixes compute. e-compressione-sweep-protocole-compute-sweep

6.3 Ablations and diagnostic examples

Read component removals and qualitative examples within their stated evaluation conditions.

Table 8. The first stage sets a measurable reconstruction constraint before diffusion begins. Original paper, p. 21 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read the first three columns before comparing scores. The downsampling factor f sets spatial compression; the codebook column identifies VQ settings or the KL alternative; c gives latent channels. R-FID evaluates reconstructed images, whereas the generation tables evaluate newly sampled images. Focus on the VQ rows for f=4 with codebook size 8192: the dagger marks the autoencoder without attention, so it should not be mistaken for a different diffusion sampler. The bottom block changes regularization to KL. The caption fixes provenance for this comparison: training uses OpenImages and reconstruction evaluation uses ImageNet validation. Arrows in each metric header show the preferred direction. e-reconstructione-autoencodere-compressione-limitations

What it supports. The VQ f=4, c=3, codebook-8192 model achieves reconstruction FID 0.58, while the VQ f=32, c=16, codebook-16384 row has 31.83. The KL f=4, c=3 row reaches 0.27. These results make the compression bottleneck concrete, but reconstruction quality alone does not determine the quality of the subsequently learned prior.

Where the evidence stops. Rows change channel counts, codebooks and sometimes attention or regularization. This table is not a controlled ablation of f alone. R-FID is also not interchangeable with generated-image FID or a guarantee of exact pixel recovery.

Figure 7. Compression changes both the cost per sample and the fidelity attainable at a given sampling budget. Original paper, p. 6 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. The left panel evaluates CelebA-HQ and the right panel ImageNet. Lower vertical positions mean better log FID; movement to the right means higher throughput in samples per second. The shared legend identifies the compression factor by color. Along each curve, the markers correspond to 10, 20, 50, 100 and 200 DDIM steps from right to left: adding denoising steps usually trades speed for quality. The dashed path connects the 200-step operating points across compression factors. These are 5,000-sample diagnostic estimates, after 500K CelebA training iterations or 2M ImageNet iterations, rather than the 50K-sample final generation benchmarks. e-compressione-sweep-protocole-evaluation

What it supports. Moderate compression offers useful operating points between the slow pixel-space model and the heavily compressed models. ImageNet particularly exposes the quality cost of f=32. The defensible conclusion is a quality–throughput tradeoff favoring f=4 or f=8 in the studied setting, not one universally optimal compression factor for every dataset and latency budget.

Where the evidence stops. The vertical axis is log FID, so its tick values are not raw FID scores. Tables 13–14 also vary batches, learning rates and model details; equal training iterations on an A100 do not imply identical exposure or compute.

7. Analysis & limitations

7.1 What the evidence leaves open

Source description

Sequential denoising remains slower than GAN sampling. Autoencoder reconstruction can limit fine pixel accuracy; the authors identify super-resolution as a likely example. e-limitations

Source description

Bicubic-only super-resolution transfers poorly to other degradations. LDM-BSR broadens corruption types, but delegates its adapted parameters to code. e-blind-sr

Reader analysis

The paper discusses manipulated imagery, training-data disclosure and inherited biases. Distribution coverage does not establish privacy, fairness or factual image content. e-limitations

Reader analysis

Headline FIDs lack run-to-run confidence intervals. The user study specifies tasks and viewing duration but omits participant and judgment counts, limiting uncertainty assessment. e-imagenete-texte-human-study

7.2 Questions for discussion

  1. At equal training exposure and equal end-to-end compute, how stable is the preferred compression factor?
  2. Which guidance scale gives the best FID–recall compromise for one checkpoint?
  3. How much restoration error comes from the autoencoder versus the diffusion prior?

8. Reproducibility audit

8.1 Requirements and known gaps

Source description

Recover the matching autoencoder, datasets, conditioner, diffusion schedule and sampler. Separate reconstruction from generation metrics, and fix resizing and reference statistics: Appendix E documents evaluator-dependent differences. e-autoencodere-hyperparameterse-evaluatione-superresolution

Source description

Table 18 estimates 271 V100-days for updated ImageNet diffusion and lists 29 separately for its f=4 autoencoder. The conversion assumes A100 is 2.2× faster than V100, rather than measuring an end-to-end V100 reproduction. e-compute

Open question

Resolve source discrepancies: Table 15 lists class-model size 395M versus 400M in Tables 3/10, and semantic-map f=8 versus f=4 in Section 4.3.2. Preserve the printed Q/V shape inconsistency rather than inventing a configuration. e-hyperparameterse-imagenete-spatiale-conditioning

Reader analysis

Reader-proposed checks: compare compression under matched exposure and matched compute; separately sweep guidance on one fixed ImageNet checkpoint to test whether FID gains accompany recall loss. These tests were not run. e-compressione-sweep-protocole-imagenet

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: Separate compression from training exposure and compute

Reader-proposed experiment: repeat a small ImageNet comparison of f=1, 4, 8 and 32 with comparable denoiser capacity and one fixed evaluation pipeline. Run one comparison at matched numbers of training examples, and another at matched accelerator time; report autoencoder pretraining separately. Hold DDIM settings and evaluation labels fixed, measure reconstruction R-FID, generated FID and throughput, and repeat seeds. The paper motivates a middle-ground advantage. If that advantage disappears after exposure or compute is controlled, the original sweep would not by itself isolate compression as the cause. e-compressione-sweep-protocole-reconstructione-compute-sweepe-computee-evaluation

Check 2: Test the ImageNet guidance–coverage tradeoff on one checkpoint

Reader-proposed experiment: use one verified updated LDM-4 checkpoint and compare unguided sampling with classifier-free guidance scales 1.25 and 1.5. Keep 250 DDIM steps, class frequencies, latent seeds, preprocessing and the 50K-sample evaluation protocol fixed. Record FID, precision, recall and throughput across repeated sample sets. Table 10 predicts lower FID but lower recall at scale 1.5. Failure to observe that joint pattern would challenge the reported operating-point interpretation; lower FID alone would not establish improved coverage. No training or sampling has been run for this proposal. e-imagenete-evaluatione-computee-revision

8.3 Reading coverage

Visual audit: The title/author/version block, main method and result pages, revision changelog, image-guidance formulation, latent-scale diagnostic, reconstruction and generation tables, configuration tables, evaluation protocols and compute/autoencoder details were rendered and visually inspected. This includes Figures 1–11 and 14–18 and Tables 1–18. All five final crops were individually viewed; the inpainting crop retains the caption because it contains the necessary test-split and recomputed-baseline note. Figure 3 arrows and routing were checked against Section 3.3; the printed projection-shape inconsistency is disclosed. Appendix C has a separate Figure 14 caption/column-reference mismatch, recorded in the base evidence. All nine text chunks were read, including remaining appendix captions; the galleries on pp. 14–15 and 30–45 were not visually inspected. No external supplements or code were inspected.

PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 16, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29. 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 (pp. 1–3)
  • Sections 3.1–3.3: compression, latent diffusion and conditioning (pp. 3–5)
  • Sections 4.1–4.5: compression tradeoffs, unconditional and conditional synthesis, super-resolution and inpainting (pp. 5–9)
  • Sections 5–6: Limitations & Societal Impact; Conclusion (p. 9)
  • References (pp. 10–13)
  • Appendix A: Changelog (p. 16)
  • Appendix B: diffusion formulation (pp. 16–17)
  • Appendix C: image guiding (pp. 18–19)
  • Appendix D.1–D.6.1: additional results, including compute-normalized analysis and blind super-resolution (pp. 20–23)
  • Appendix E.1–E.3.6: hyperparameters, implementations, evaluation and user study (pp. 24–27)
  • Appendices F–G: computational requirements and autoencoder details (pp. 28–29)
  • Appendix H and all supplied qualitative figure captions, including pp. 14–15 and 30–45

Outside the original text pass

  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
  • Separate supplemental material availability has not been fully verified.
  • All nine supplied text chunks were read individually, covering the complete 45-page PDF. Original images were inspected for Figures 1–11 and 14–18 and Tables 1–18; the remaining appendix galleries were read through their supplied captions, not visually reviewed.
  • Identity note: the title page identifies arXiv:2112.10752v2, 13 April 2022, and the same five authors as the catalog. The observed title uses lower-case “with”; the catalog capitalizes “With”. Appendix A documents a larger text-to-image model, retrained ImageNet results, classifier-free guidance and a new user study relative to v1. This review concerns that v2 artifact; the separately cataloged CVPR edition and v1 were not supplied for comparison.
  • No code or external links were inspected, no separate supplements were acquired, and no experiments were reproduced.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e-identityPDF p. 1, title/author block and arXiv marginInspect

Observed title: High-Resolution Image Synthesis with Latent Diffusion Models. Authors: Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser and Björn Ommer. The first two contributed equally. Affiliations: Ludwig Maximilian University of Munich & IWR, Heidelberg University, Germany; Runway ML. The margin identifies arXiv:2112.10752v2, 13 Apr 2022.

Go to primary source ↓
e-revisionPDF p. 16, Appendix A, all four changelog bulletsInspect

Relative to v1, v2 updates text synthesis with a 1.45B model and new baselines; retrains ImageNet with a larger batch; uses classifier-free guidance for both; adds a super-resolution/inpainting user study and changes qualitative figures.

Go to primary source ↓
e-motivationPDF pp. 2–3, Departure to Latent Space; Section 3 openingInspect

Perceptual compression and generative learning are separated. A pretrained autoencoder can be reused across diffusion runs and downstream tasks, reducing repeated pixel-space computation.

Go to primary source ↓
e-autoencoderPDF pp. 3–4, Section 3.1; p. 29, Appendix G, Eq. (25) and DM Training in Latent SpaceInspect

The encoder/decoder uses perceptual reconstruction, patch adversarial training, and weak KL or VQ regularization. KL weight is approximately 10^-6. VQ diffusion uses latents before quantization; the decoder includes quantization.

Go to primary source ↓
e-latentPDF p. 4, Section 3.2, Eqs. (1)–(2) and paragraph below Eq. (2)Inspect

The time-conditioned UNet predicts Gaussian noise on encoded images. The forward process is fixed. Generated latents are decoded with one decoder pass; sampling still requires iterative denoising.

Go to primary source ↓
e-conditioningPDF pp. 4–5, Figure 3, Section 3.3, attention expression and Eq. (3)Inspect

Figure 3 depicts cross-attention and concatenation. Queries come from UNet features; keys and values from the condition encoder. Denoiser and conditioner are jointly optimized. Printed Q and V projection dimensions conflict with the input dimensions implied by their assignments.

Go to primary source ↓
e-conditioner-configPDF pp. 25–26, Section E.2.1, Eqs. (18)–(24), Tables 16–17 and class-conditioning paragraphInspect

Unmasked transformer conditioning feeds self-attention/MLP/cross-attention blocks. The conditioner is not time-conditioned. Classes use a single learned 512-dimensional embedding; layouts encode discrete corners and class information.

Go to primary source ↓
e-spatialPDF p. 7, Section 4.3.2 and Figure 9Inspect

Aligned conditioning is concatenated. The semantic-synthesis description uses VQ f=4 trained on 256-resolution crops and demonstrates larger convolutional outputs; latent scaling affects this behavior.

Go to primary source ↓
e-scalePDF p. 20, Section D.1 and Figure 15; p. 29, Appendix G, latent-rescaling formulaInspect

Latent variance affects denoising signal-to-noise ratio and convolutional sampling. Appendix G rescales KL latents with a first-batch standard-deviation estimate. Although called component-wise, its displayed statistic sums across batch, channels and spatial positions.

Go to primary source ↓
e-reconstructionPDF p. 21, Table 8, VQ and KL rows and captionInspect

Autoencoders train on OpenImages and are evaluated on ImageNet validation. VQ f=4, codebook 8192, c=3 has R-FID 0.58 and PSNR 27.43 ± 4.26; its attention-free counterpart has R-FID 1.06. VQ f=32, codebook 16384, c=16 has R-FID 31.83. KL f=4, c=3 has R-FID 0.27.

Go to primary source ↓
e-compressionPDF p. 5, Section 4.1; p. 6, Figures 6–7 and captionsInspect

The sweep varies f over 1, 2, 4, 8, 16, 32. Section 4.1 reports a FID gap of 38 between LDM-1 and LDM-8 after 2M ImageNet iterations. Figure 7 plots log FID against samples/second; markers represent 10, 20, 50, 100, 200 DDIM steps from right to left, with dashed connections between 200-step points.

Go to primary source ↓
e-sweep-protocolPDF p. 24, Table 13; p. 25, Table 14 and footnote; p. 27, Section E.3.5Inspect

Compression experiments use a single A100, comparable parameter counts, differing batches and maximized stable learning rates. Efficiency FIDs use 5K samples. CelebA uses 500K iterations or the best earlier checkpoint if convergence occurs; ImageNet uses 2M.

Go to primary source ↓
e-compute-sweepPDF p. 22, Section D.5 and Figure 17 captionInspect

The supplementary comparison plots class-conditional ImageNet progress over a fixed 35 V100-day budget, using 100 DDIM steps and 5K samples for FID.

Go to primary source ↓
e-imagenetPDF p. 7, Table 3; p. 22, Table 10, updated LDM-4 rowsInspect

At 250 steps, unguided LDM-4 gives FID 10.56, precision 0.71 and recall 0.62. Scale-1.5 guidance gives FID 3.60, IS 247.67 ± 5.59, precision 0.87 and recall 0.48. Scale 1.25 gives FID 3.95 and recall 0.55. Runs use 178K iterations, batch 1200; tables report 400M parameters. ADM-G has FID 4.59, recall 0.52 and 608M parameters.

Go to primary source ↓
e-evaluationPDF pp. 26–27, Section E.3.1Inspect

Unconditional/class-conditional evaluation uses 50K generated samples and training-set references. FID uses torch-fidelity and is checked against the Dhariwal/Nichol evaluator; ImageNet and LSUN-Bedrooms show small pipeline-dependent differences. Precision and recall use the latter evaluator.

Go to primary source ↓
e-textPDF p. 6, Table 2; p. 7, Section 4.3.1; p. 27, Section E.3.2Inspect

The 1.45B KL f=8 model trains on LAION-400M with BERT tokenization and transformer conditioning. COCO validation evaluation uses 30K references at 256×256. At 250 steps, unguided/guided FIDs are 23.31/12.63, guidance scale 1.5. GLIDE reports 12.24 and Make-A-Scene 11.84 with 6B/4B parameters respectively.

Go to primary source ↓
e-unconditionalPDF p. 6, Table 1, all four dataset panelsInspect

CelebA-HQ LDM-4 at 500 steps has FID 5.11, precision 0.72 and recall 0.49, versus LSGM 7.22 and PGGAN 8.0 FID. FFHQ, LSUN-Churches and LSUN-Bedrooms panels include GANs with lower FID than LDM.

Go to primary source ↓
e-layoutPDF p. 20, Section D.3; p. 22, Table 9; p. 27, Section E.3.3Inspect

COCO evaluation uses 2048 unaugmented Segmentation Challenge examples. LDM-4 fine-tuned from OpenImages at 200 steps scores FID 40.91; COCO-only LDM-8 at 100 steps scores 42.06; SPADE scores 41.11.

Go to primary source ↓
e-superresolutionPDF p. 8, Section 4.4 and Table 5; p. 27, Section E.3.4Inspect

ImageNet 64→256 uses antialiased bicubic downsampling and excludes short sides below 256. At 100 steps LDM-4 gives FID 2.8 against validation features or 4.8 against training features, and PSNR 24.4 ± 3.8; SR3 reports FID 5.2 and PSNR 26.4, and regression PSNR is 27.9.

Go to primary source ↓
e-inpainting-efficiencyPDF p. 8, Table 6 and Section 4.5Inspect

After six epochs VQ LDM-4 with attention gives validation FID 14.99 versus pixel LDM-1 24.74, with epoch times 7.04 versus 20.66 hours. This uses 2K validation examples, not the final 30K-test benchmark. GPU/batch settings differ from Figure 7.

Go to primary source ↓
e-inpaintingPDF p. 9, Table 7 and caption; p. 26, Section E.2.2Inspect

Places uses synthetic masks, 2K validation and 30K test examples, 256 training crops and 512 test crops. Large fine-tuned LDM-4 has overall FID 1.50 and LPIPS 0.137 ± 0.080; recomputed LaMa has 2.23 and 0.134 ± 0.080. Hard-mask FIDs are 9.39 and 12.31. Original LaMa test images were unavailable.

Go to primary source ↓
e-inpainting-finetunePDF pp. 8–9, Section 4.5 continuation; p. 9, Table 7 first two rowsInspect

The larger inpainting UNet has 387M rather than 215M parameters and additional attention levels. Half an epoch of 512-resolution fine-tuning improves overall FID from 2.40 to 1.50 and hard-mask FID from 12.89 to 9.39. Authors hypothesize changed feature statistics explain the resolution gap.

Go to primary source ↓
e-image-guidancePDF pp. 18–19, Appendix C, Eqs. (16)–(17) and Figure 14Inspect

Post-hoc guidance evaluates an image target through the decoder and a differentiable transformation using L2 or perceptual similarity. Figure 14 headers label the middle column guided and the right column unguided; its caption points to column 2 when describing unguided incoherence, an internal reference mismatch.

Go to primary source ↓
e-blind-srPDF p. 23, Section D.6.1 and Figure 18Inspect

Bicubic-only LDM-SR transfers poorly to differently degraded images. LDM-BSR adds randomized JPEG, sensor noise, interpolation, blur and Gaussian noise. Authors weaken the referenced settings but point to code rather than tabulating adapted parameters.

Go to primary source ↓
e-human-studyPDF p. 8, Table 4; p. 27, Section E.3.6Inspect

The study compares generated images with ground truth and competing generated images using two-alternative forced choice after three-second viewing. The supplied description omits participant and judgment counts.

Go to primary source ↓
e-limitationsPDF p. 9, Section 5Inspect

Authors identify sequential sampling speed, fine-detail reconstruction limits, manipulation, possible training-data disclosure and inherited dataset bias as limitations or societal concerns.

Go to primary source ↓
e-computePDF p. 28, Appendix F, Table 18 ImageNet rows, caption and conversion paragraphInspect

Updated LDM-4 training is reported as 271 V100-days; its f=4 first stage is separately listed at 29. A100 time is converted using an assumed 2.2× speedup. Throughput is measured on one A100; the scale-1.5 guided ImageNet row reports 0.4 samples/second.

Go to primary source ↓
e-hyperparametersPDF p. 25, Table 15, conditional-task columns and captionInspect

Conditional models use 1000 diffusion steps and linear schedules. The class run lists batch 1200 and 178K iterations. Table 15 lists 395M parameters for class synthesis and f=8 for semantic maps, differing from Tables 3/10 and Section 4.3.2. The caption states one A100 except inpainting on eight V100s.

Go to primary source ↓

8.5 Primary sources

Scroll across the image to inspect details. Press Esc to close.