PhotoQuilt: Training-Free Arbitrary-Resolution Photomosaics via Bootstrapped Tiled Denoising

arXiv:2606.30968 · cs.CV · Submitted 2026-06-29 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "PhotoQuilt: Training-Free Arbitrary-Resolution Photomosaics via Bootstrapped Tiled Denoising".

Jane: Here's the gist of the paper. Photomosaics are large images whose local regions are seen as independent tiles while their overall arrangement forms a coherent scene The single most informative sentence about the paper's subject and finding is PhotoQuilt, a training-free framework that generates photomosaics at arbitrary resolution by decoupling the global composition from the generation of local tiles.

Tom: Core Problem Addressed Photomosaics present challenges because they must resolve into two different images depending on the distance from which it is viewed The dual-scale behavior makes it difficult to produce since the target image must remain recognizable as a coherent scene from afar while every tile stays a convincing, meaningful image, and satisfying both at the same time is far from trivial.

Jane: Existing methods fall into two categories: retrieval-based methods construct a photomosaic by selecting the best-matching tile from a fixed image pool for each region of the target and applying local color corrections to improve the fit Generative methods replace retrieval with synthesis, editing, or conditioning each tile on a text prompt or reference image.

Tom: Diffusion models struggle to satisfy both scales at once because direct high-resolution generation is costly and most methods generate tiles sequentially which is time-consuming and still requires additional coordination procedures to assemble independently generated tiles into a coherent global image.

Title and authors: Tom: So we're looking at this paper called PhotoQuilt: Training-Free Arbitrary Resolution Photomosaics via Bootstrapped Tiled Denoising. Jane, what’s the big picture here?

Jane: Well, essentially, they’ve tackled the problem of making these large images that look good up close and from far away. They call them photomosaics, and the core difficulty is balancing that two-scale behavior.

Lu: What makes it interesting is how they solve this by decoupling the global composition from generating each local tile separately, which is a big shift because diffusion models usually struggle to handle both scales at once efficiently.

Meng: But I’m wondering about the cost, right? If you have to generate every single tile independently and then stitch them together later, doesn't that just lead to a slow process when the canvas gets really big?

Tom: That’s exactly where PhotoQuilt claims it’s faster and more scalable. They propose a bootstrapped tiled denoising procedure where they first create a global composition at low resolution to set the layout, then upscale it in latent space and renoise it to restore the detail.

Jane: So, how does this process manage that balance between the coarse global structure and the fine local detail?

Lu: They use a shared renoised latent that acts as a single control for both levels; smaller noise strength enforces the global structure more strictly, while larger noise strength gives each tile more freedom to diverge locally.

Meng: So, if I’m running this on real traffic or large-scale data, how does that tiling strategy actually work in practice? Does it introduce quadratic costs somewhere?

Tom: They claim they achieve linear cost scaling with the size of the canvas because confining the attention within each tile keeps the generation cost linear rather than quadratic. That makes high-resolution mosaics practical to produce on modern backbones like DiTs.

Jane: It sounds like a really clever way to keep things manageable when you’re dealing with massive images that need high fidelity in both areas. But what does this mean for how we actually use these models?

Title and authors: Lu: They show they can work without any extra training or architectural changes, meaning it applies to different diffusion backbones, whether it's a U-Net or a DiT model.

Meng: That’s good because I don't want to spend weeks retraining my production pipeline just for this new technique. But what about conditioning? Can we control what each tile actually looks like?

Tom: Yes, they have flexible conditioning modes. You can use a shared global prompt where the condition for every tile is the same as the original prompt, or you can use per-tile text prompts to give explicit control over individual parts.

Jane: And there’s another thing that really catches my attention: they address image gallery conditioning too. They propose assigning a gallery entry to each tile via a mapping function, so tiles can be conditioned on real images from a provided set of images, which brings it closer to the old tile-pool setting.

Lu: That mapping function essentially sets the condition for each tile as the reference image multiplied by some function, which is how they recover that classical approach where you pick a specific tile from a pool based on colors and structure.

Meng: That sounds powerful for specialized applications, but I have to ask about deployment. How does this affect inference speed when we are trying to generate these ultra-high resolution canvases in production?

Tom: They show comparable generation speeds on the same backbone, and they even beat Phomosaic’s SD2 point 1 implementation even when running on heavier models like FLUX point one and point two DiTs <ref:2606.30968#pg1>. Plus, they have a multi-GPU distributed generation feature for synthesizing ultra-high resolution canvases without losing quality or needing approximation.

Jane: So, to put it simply, PhotoQuilt gives us a training-free way to build high-quality photomosaics by setting up a global structure first and then letting each tile grow its own detail from that shared foundation.

Lu: The key is that this approach decouples the two scales effectively, addressing the trade-off between global fidelity and local tile quality across all tested backbones.

Title and authors: Meng: I think what’s important is that it works without needing specialized procedures or extra training, which simplifies deployment immensely for production teams.

Tom: It really does simplify things because it just relies on the internal capabilities of the underlying diffusion model, making it a minimal yet effective approach for high-resolution mosaics.

Jane: So, looking at what we’ve seen with PhotoQuilt: Training-Free Arbitrary Resolution Photomosaics via Bootstrapped Tiled Denoising, the main implication is that we can generate these complex images at any resolution without the quadratic cost problem plaguing previous methods.

Lu: It fundamentally breaks that trade-off between global structure and local tile quality across all tested backbones because of how they manage the denoising process within fixed tiles.

Meng: For me, as someone shipping models, the fact that it scales linearly with canvas size is huge because it means we can actually deploy this on larger tasks without worrying about exponential slowdowns.

Tom: It’s impressive that it achieves this balance with just a default configuration of noise strength s equals zero point six and a resolution of seven hundred sixty eight by seven hundred sixty eight, which they found to be the optimal setting compared to other variants.

Jane: So, to wrap up this discussion on PhotoQuilt: Training-Free Arbitrary Resolution Photomosaics via Bootstrapped Tiled Denoising, the paper shows a way to generate high-quality mosaics by using bootstrapped tiled denoising instead of generating tiles in isolation and combining them later.

Lu: The final point is that this framework applies to both U-Net and DiT backbones, outperforming existing baselines on both global structure and local realism without requiring any training or architectural changes.

Meng: I just want to reiterate that the multi-GPU distributed generation feature means we can synthesize ultra-high resolution canvases in parallel with equal load balancing.

Tom: It’s a solid result because it maintains both global layout fidelity and local tile autonomy while being incredibly fast compared to competitors like Phomosaic.

Jane: That's all for this discussion on PhotoQuilt, and we’ll be back next time when we look at some other interesting work from arXiv.

The paper's summary: Tom: So, we just talked about how PhotoQuilt uses this bootstrapped denoising thing to handle two different scales in a mosaic image without needing extra training or special procedures. Now, let's look at what they actually claim this method achieves in terms of results.

Jane: Okay, so the summary says that PhotoQuilt consistently improves both the big picture coherence and how sharp the fine details are. And it does this using only the internal capabilities of the underlying model, which is pretty minimal for a high-resolution mosaic generator.

Lu: Exactly. The authors say they break that trade-off between global fidelity and local tile quality across all tested backbones, meaning you don't have to choose one over the other anymore. They found that their default settings—noise strength of zero point six and a resolution of seven hundred sixty eight by seven hundred sixty eight—gave the best balance for everything they tested.

Meng: I’m still thinking about the cost here. They claim they achieve this without quadratic attention cost because confining the attention within each tile keeps the generation cost linear with respect to the size of the canvas. That’s actually pretty important for deployment on large images, right?

Tom: It is. And they also report that inference efficiency stays comparable to other models on the same backbone, even when running on heavier ones like FLUX point one and point two DiTs. Plus, they have a multi-GPU distributed generation feature that lets you synthesize ultra-high resolution canvases without losing quality or needing any approximation.

Jane: So, for someone just listening who isn't deep in the math, what does this mean for creating these big images? It means you can generate them at whatever size you need, and they do it reliably with a setting that seems pretty good by default.

Lu: The big idea is the decoupling. They don't generate everything at once; they fix the global layout first and then let each tile grow its own image from that shared structure. It’s a clever way to manage complexity without having to build an entire new architecture or train a whole new system.

Meng: If I were deploying this, the fact that it's training-free and works on both U-Net and DiT backbones is what makes it interesting for my pipeline. It just plugs in, right?

Tom: Right. It’s a minimal but effective approach that performs well across the board. The main thing they show is that this bootstrapping method really works to solve the problem of how to keep things coherent globally while letting each part be its own convincing image locally.

Jane: It seems like they’ve managed to make a very powerful technique accessible, moving away from the clunky retrieval methods we used before. But there's still one thing they flag as a limitation: if you're using it for gallery conditioning, the reconstruction quality can depend on which diffusion backbone you use because the tile might diverge from its original reference image.

Lu: That’s a fair caveat. It means that while the core tiling mechanism is robust, applying it to specific tasks like image editing or retrieval might still require tuning based on your choice of underlying model architecture.

Tom: So, they proved this technique works really well across different model types and gives you good control over both scales without the usual bottlenecks. The next step for them seems to be making sure that gallery conditioning works more consistently regardless of the backbone choice.

The paper's improvements: Tom: So, we’ve seen how PhotoQuilt works right now, and now let's look at what the authors suggest they can do to make it even better going forward. They are proposing a few specific improvements to their method that really tighten up how it handles those different scales.

Jane: The paper suggests making the global structure preservation stronger by using a single shared latent that essentially fixes what the large scale looks like, which helps reconstruct the target structure specified by a prompt. That sounds like a big win for global coherence, right?

Lu: Yeah, they are also enhancing local tile realism by letting each tile follow its own separate trajectory starting from that shared initialization. This lets them develop independently while still being guided by the overall structure. It’s about giving tiles more autonomy within the system.

Meng: That sounds like it would definitely help with production because it means we have better control over how much freedom each tile gets before we start denosing them. I wonder if that leads to more predictable output when scaling up to really high resolutions?

Tom: They’re also introducing flexible conditioning modes, letting you choose between a shared global prompt where every tile gets the same condition, or per-tile text prompts for explicit control over individual parts. That gives users more fine-grained steering over the final result.

Jane: And they're addressing that gallery conditioning issue we talked about earlier by proposing a mapping function to assign a specific entry from a gallery to each tile, effectively recovering the classical tile-pool setting you might be familiar with.

Lu: That mapping function sets the condition for each tile as the reference image multiplied by some function, which lets them recover that old method where you just pick a specific tile based on its colors and structure. It’s a way to bridge their new method with older, more specialized control techniques.

Meng: So, from an engineering standpoint, it sounds like they’re focusing on making the input controls more explicit and robust across different use cases. But what about the computational side of these proposed changes? Are these improvements going to slow down our inference time significantly?

Tom: The core goal seems to be maintaining efficiency while increasing control. They are also using a two-dimensional, distributed block-tiling strategy for the final pixel space decoding phase in the VAE decode. This ensures peak memory usage is only limited by aggregate host memory, making it purely throughput limited and efficient.

Jane: So, to summarize these improvements: stronger global structure fixing, better independent tile development with per-tile conditioning options, a cleaner way to handle image gallery input, and an optimized decoding strategy for speed. It’s a lot of refinement on top of the initial framework.

Lu: These changes are aimed at making the system more versatile. The authors are pushing towards a framework that is not just good at mosaics but adaptable to more specific creative tasks through these new control interfaces.

Meng: I like that focus on throughput and memory bounding. If they keep that linear cost scaling while adding these controls, then it’s much easier to predict the resource usage when we move this into our production environment.

Tom: It seems like they're really building a complete system here, not just a patch. They are pushing towards a framework that is adaptable across different tasks without needing to rewrite the whole thing from scratch. But we still have that one sticking point: the dependency on the underlying diffusion backbone for gallery conditioning quality.

Jane: Right, so while they’re making big gains in control and efficiency, they still need to nail down that consistency when you use different diffusion models for those specific image gallery tasks. That's a challenge they have left to solve before it’s fully ready for every single application.

Conclusion: Tom: So we’ve covered how PhotoQuilt uses bootstrapped denoising to create training-free photomosaics at any resolution, balancing global structure and local detail without needing extra training or architecture changes. Now for the wrap-up of this paper.

Jane: Basically, the main implication is that we can generate these complex images at any size reliably by decoupling the global composition from every single tile generation. It’s a training-free way to tackle a very hard problem in image synthesis.

Lu: I think it really changes how we approach these images because it eliminates the need for those complicated retrieval or sequential generation steps that used to be necessary. It just works by fixing the big picture first and letting the details follow.

Meng: It’s good that they showed it works across different diffusion backbones, whether you use a U-Net or a DiT model. That means it’s not tied to one specific architecture, which is huge for deployment flexibility.

Lalam: I'm confident in this result; the method seems solid because it doesn't rely on any heavy retraining or complex fine-tuning procedures to get that quality. It’s a very clean approach to solving the scale problem in image synthesis.

Tom: I agree, Lalam, it’s that clean execution that gets my attention. The paper shows how this bootstrapped tiled denoising procedure effectively scales without incurring those quadratic attention costs we usually see with high-resolution generation.

Jane: That efficiency is key for anyone trying to build a system that can handle massive canvases in a reasonable amount of time. It moves the goalposts on what’s feasible for high-resolution image synthesis.

Lu: The future work they hint at is mostly about refining those specific control mechanisms, like making the gallery conditioning more consistent across different model choices. That's where the real engineering challenge lies next.

Meng: I hope they spend time on making sure that scaling behavior stays linear even when you push the resolution to really extreme levels. If it breaks its linear scaling at some point, it won't be deployable for large-scale industrial applications like edge processing or massive video synthesis.

Tom: We’ll definitely be watching those follow-up papers to see if they can fully stabilize that gallery conditioning aspect across different model types. It’s a solid foundation, though.

Jane: So, PhotoQuilt is a training-free framework for generating photomosaics at arbitrary resolution by decoupling global composition from the generation of local tiles through bootstrapped denoising. That’s the core finding here.

Lu: It provides a way to achieve high fidelity in both the global structure and the individual tile details simultaneously without quadratic cost concerns.

Meng: It’s a minimal approach that keeps things efficient enough for production if you use their default settings, which is something I respect.

Lalam: I'm sure this method will be used a lot because it solves a fundamental trade-off in image generation right here.

Tom: That covers PhotoQuilt for now. We’ll keep an eye on those refinements to the gallery conditioning next time we look at new arXiv submissions.

Koorosh Roohi, Javad Rajabi, Andrew Fleet, Babak Taati

University of Toronto · Vector Institute · Samsung Research · KITE Research Institute · Queen's University

cs.CV

Submitted: 2026-06-29

Updated: 2026-11-16

Comments: WACV 2027. Project page: https://kooroshrh.github.io/photo-quilt/

Project page: https://kooroshrh.github.io/photo-quilt

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 93/100

The gist: Photomosaics are large images whose local regions are seen as independent tiles while their overall arrangement forms a coherent scene The single most informative sentence about the paper's subject…

Key concepts

Bootstrapped Tiled Denoising
This is the core process where the model first generates a coarse, low-resolution version of the entire image to set up the overall layout. This shared structure then guides subsequent denoising steps for each tile. It acts as a single control point that locks in the large-scale composition before allowing local details to be refined.
Tile Autonomy
Each individual tile is treated as an independent generation process, starting from its own region of the shared latent structure. This means every tile follows its own unique denoising trajectory based on its specific condition. This isolation allows tiles to develop their own convincing local details without being overly constrained by neighboring tiles.
Bootstrapped Procedure
The method uses a two-step process: first, creating a base latent (low resolution) that captures the global structure, and second, iteratively renoising this latent. The low-resolution step fixes the 'what' of the scene globally, while subsequent steps add high-frequency detail locally. This bootstrapping ensures both scale fidelity and fine detail are achieved.
Linear Attention Cost
By confining the attention mechanism within each tile during denoising, PhotoQuilt avoids quadratic computational costs associated with generating large images. This confinement ensures that the generation time scales linearly with the size of the canvas, making it practical to produce very high-resolution mosaics without prohibitive computational expense.

Terminology

Summary

Photomosaics are large images whose local regions are seen as independent tiles while their overall arrangement forms a coherent scene The single most informative sentence about the paper's subject and finding is PhotoQuilt, a training-free framework that generates photomosaics at arbitrary resolution by decoupling the global composition from the generation of local tiles.

Core Problem Addressed

Photomosaics present challenges because they must resolve into two different images depending on the distance from which it is viewed The dual-scale behavior makes it difficult to produce since the target image must remain recognizable as a coherent scene from afar while every tile stays a convincing, meaningful image, and satisfying both at the same time is far from trivial. Existing methods fall into two categories: retrieval-based methods construct a photomosaic by selecting the best-matching tile from a fixed image pool for each region of the target and applying local color corrections to improve the fit Generative methods replace retrieval with synthesis, editing, or conditioning each tile on a text prompt or reference image. p. 1

Where the paper says it

Existing approaches fall into two categories. Retrievalbased methods construct a photomosaic by selecting the best-matching tile from a fixed image pool for each region of the target and applying local color corrections to improve the fit [14].

Diffusion models struggle to satisfy both scales at once because direct high-resolution generation is costly and most methods generate tiles sequentially which is time-consuming and still requires additional coordination procedures to assemble independently generated tiles into a coherent global image. p. 1
Where the paper says it

Generative methods replace retrieval with synthesis, editing, or conditioning each tile on a text prompt or reference image [7, 35, 52]. Diffusion models make this direction promising, but applying them to photomosaics at high resolution is non-trivial: direct generation is computationally heavy, so most methods generate tiles sequentially [9, 11], which is time-consuming and still requires additional coordination procedures to assemble independently generated tiles into a coherent global image.

Methodology Overview

PhotoQuilt resolves this with a bootstrapped tiled denoising procedure where it first produces a global composition at low resolution to fix the layout, then upscales it in latent space and re-inject noise to restore generative capacity. The denoising proceeds within fixed tiles so that each forms its own image while the shared global structure holds them in one layout. p. 0

Where the paper says it

We first produce a global composition at low resolution to fix the layout, then upscale it in latent space and re-inject noise to restore generative capacity. Denoising proceeds within fixed tiles, so each forms its own image while the shared global structure holds them in one layout.

The process involves two main steps: establishing a coarse global structure and then using that to guide the generation of every tile. p. 1
Where the paper says it

In this work, we introduce PhotoQuilt, a training-free framework that produces photomosaics at arbitrary resolution by decoupling the global composition from the generation of local tiles. Instead of generating tiles in isolation and stitching them together afterward, PhotoQuilt first establishes a coarse global structure and then uses that to guide the generation of every tile.

Bootstrapped Tiled Denoising

The method begins by obtaining a base latent at low resolution, either by generating it from the base prompt or by encoding a provided image. This base latent is then upsampled to the target latent grid and renoised once with strength s to create z˜s, which retains the coarse content of zˆ while leaving its high-frequency detail to be regenerated. p. 3

Where the paper says it

Pixel space representation has been used instead of latent space for simplicity. Because s < 1, the renoised latent z˜s retains the coarse content of zˆ while leaving its high-frequency detail to be regenerated.

The shared renoised latent acts as a single global/local control, where smaller s enforces the global structure more strictly and larger s grants each tile more freedom to diverge. p. 3
Where the paper says it

As z˜s is shared by every tile, it fixes a common low-frequency layout across the image. The strength s thus acts as a single global/local control: smaller s enforces the global structure more strictly, while larger s grants each tile more freedom to diverge.

Tile Autonomy and Scaling

The method achieves tile autonomy by splitting the target latent into K equal, non-overlapping tiles and denoising all tiles together, each as a separate trajectory. Each tile starts from its own region of the shared renoised latent and follows its own condition ck. p. 3

Where the paper says it

We split the target latent into K equal, non-overlapping tiles {Ωk} and denoise all tiles together, each as a separate trajectory. Each tile starts from its own region of the shared renoised latent and follows its own condition ck, z (k) 0 = Φ z˜s Ωk ; s→0, ck , k = 1, .

This per-tile confinement keeps the cost of generation linear in the size of the canvas, which makes high-resolution mosaics practical to produce. p. 1
Where the paper says it

Confining attention within each tile keeps the cost of generation linear in the size of the canvas. This per-tile confinement is what makes high-resolution mosaics practical to produce, yielding outputs in which the target image is clearly recognizable as a whole while every tile remains a meaningful, high-quality image in its own right, a pairing that direct high-resolution generation struggles to achieve.

Experimental Findings

Experiments show that PhotoQuilt consistently improves both global coherence and fine-detail fidelity, outperforming existing baselines. It achieves this without any additional or specialized procedures, drawing only on the internal capabilities of the underlying model, which makes it a minimal yet effective approach to high-resolution photomosaic generation that remains stable and efficient across a wide range of output resolutions. p. 1

Where the paper says it

Extensive experiments show that PhotoQuilt consistently improves both global coherence and fine-detail fidelity, outperforming existing baselines. It achieves this without any additional or specialized procedures, drawing only on the internal capabilities of the underlying model, which makes it a minimal yet effective approach to high-resolution photomosaic generation that remains stable and efficient across a wide range of output resolutions.

PhotoQuilt breaks the trade-off between global fidelity and local tile quality across all tested backbones. p. 6
Where the paper says it

PhotoQuilt breaks this trade-off across all tested backbones. 1, it outperforms Phomosaic in global structure while matching or exceeding its tile quality.

The default configuration (s = 0. p. 7
Where the paper says it

) Figure 5. The default configuration (s = 0.

6, 768 × 768) provides the optimal balance between global layout fidelity and local tile realism compared to the ablated variants. p. 7
Where the paper says it

The default configuration (s = 0. 6, 768 × 768) provides the optimal balance between global layout fidelity and local tile realism compared to the ablated variants.

The full 768 × 768 bootstrap consistently provides the richest structural guidance and is the default for all reported results. p. 7
Where the paper says it

12, SSIM −0. The full 768 × 768 bootstrap consistently provides the richest structural guidance and is the default for all reported results.

Inference Efficiency

PhotoQuilt achieves comparable generation speeds on the same backbone while strictly preserving both global layout fidelity and local tile autonomy. The algorithmic efficiency of our bootstrapped tiled denoising approach becomes most apparent when compared to Phomosaic, our primary fully generative competitor. p. 0

Where the paper says it

PhotoQuilt resolves this with a bootstrapped tiled denoising procedure. We first produce a global composition at low resolution to fix the layout, then upscale it in latent space and re-inject noise to restore generative capacity.

PhotoQuilt remains faster than Phomosaic’s SD2. p. 6
Where the paper says it

05 Phomosaic 20. 35 PhotoQuilt (SD2.

1 implementation even when executing on the heavier FLUX. p. 2
Where the paper says it

reference block with a per-step low-frequency loss to balance global structure with local detail, though it remains slow even for modest tile counts. Doyle et al.

1 and FLUX. p. 2
Where the paper says it

Earlier text-to-image models were built on U-Net backbones, most notably Stable Diffusion [35, 39]. Modern text-to-image generation has since shifted to the diffusion transformer (DiT) [34], and a rapidly growing family of open DiT foundation models now occupies this space, including SD3 [13], PixArtα [8], Sana [48], FLUX [3, 5], Qwen-Image [46], Microsoft Lens [17], and the open-weight Ideogram 4 [23].

2 DiT backbones. p. 2
Where the paper says it

Earlier text-to-image models were built on U-Net backbones, most notably Stable Diffusion [35, 39]. Modern text-to-image generation has since shifted to the diffusion transformer (DiT) [34], and a rapidly growing family of open DiT foundation models now occupies this space, including SD3 [13], PixArtα [8], Sana [48], FLUX [3, 5], Qwen-Image [46], Microsoft Lens [17], and the open-weight Ideogram 4 [23].

The multi-GPU distributed generation feature allows for the synthesis of ultra-high-resolution canvases without compromising quality or requiring approximation. p. 9
Where the paper says it

3 [51] Jinjin Zhang, Qiuyu Huang, Junjie Liu, Xiefan Guo, and Di Huang. Diffusion-4k: Ultra-high-resolution image synthesis with latent diffusion models.

Conclusion

We presented PhotoQuilt, a training-free framework for generating photomosaics at arbitrary resolution. Instead of generating tiles in isolation and combining them afterward, it decouples global composition from local generation through bootstrapped tiled denoising. p. 7

Where the paper says it

Conclusion We presented PhotoQuilt, a training-free framework for generating photomosaics at arbitrary resolution. Instead of generating tiles in isolation and combining them afterward, it decouples global composition from local generation through bootstrapped tiled denoising, fixing a coarse layout at low resolution, upscaling it in latent space, and renoising it once so that a single shared latent enforces global structure while each tile denoises as its own trajectory.

Since tiles are generated separately, the method scales to large canvases without quadratic attention cost. p. 7
Where the paper says it

Instead of generating tiles in isolation and combining them afterward, it decouples global composition from local generation through bootstrapped tiled denoising, fixing a coarse layout at low resolution, upscaling it in latent space, and renoising it once so that a single shared latent enforces global structure while each tile denoises as its own trajectory. Since tiles are generated separately, the method scales to large canvases without quadratic attention cost.

Requiring no training or architectural change, PhotoQuilt applies to both UNet and DiT backbones, outperforming existing baselines on both global structure and local realism. p. 7
Where the paper says it

Since tiles are generated separately, the method scales to large canvases without quadratic attention cost. Requiring no training or architectural change, PhotoQuilt applies to both UNet and DiT backbones, outperforming existing baselines on both global structure and local realism.

One remaining challenge is that in image gallery conditioning mode, reconstruction quality depends on the diffusion backbone since a denoised tile may diverge from its reference image. p. 7
Where the paper says it

Requiring no training or architectural change, PhotoQuilt applies to both UNet and DiT backbones, outperforming existing baselines on both global structure and local realism. One remaining challenge is that in image gallery conditioning mode, reconstruction quality depends on the diffusion backbone, since a denoised tile may diverge from its reference image.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Global structure preservation is improved by using a single shared latent that fixes what it looks like at the large scale, which ensures global reconstruction by having the low-frequency content of I mosaic reconstructs the target structure specified by g.

  2. Local tile realism is enhanced because each tile follows its own trajectory from a shared initialization, allowing them to develop independently, as described by the mechanism where each tile denoises as its own trajectory from that shared latent.

  3. The system can generate photomosaics at arbitrary resolution and scale without quadratic attention cost because the confining attention within each tile keeps the cost of generation linear in the size of the canvas.

  4. The framework supports flexible conditioning modes, allowing for a Shared Global Prompt (where ck = c0) or Per-Tile Text Prompts, giving explicit control over individual tile content.

  5. The model can be adapted to use image gallery conditioning by assigning a gallery entry to each tile via a mapping function, enabling tiles to be conditioned on real images from a provided gallery G where ck = R g pi(k), which recovers the classical tile-pool setting.

  6. The system is compatible with any diffusion backbone, including UNet or DiT models, as it treats the generator as a black box requiring only text-to-image and image-to-image conditioning interfaces.

  7. Ultra-high resolution photomosaics can be synthesized using a multi-GPU distributed generation feature, where each band is assigned to a separate GPU holding a full model replica, allowing denoising to proceed in parallel with equal load balancing.

  8. The final pixel space decoding uses a two-dimensional, distributed block-tiling strategy for the VAE decode phase to ensure peak memory is bounded only by aggregate host memory, making the process purely throughputlimited.

Abstract

Photomosaics are large images whose local regions are seen as independent tiles while their overall arrangement forms a coherent scene. Generating them at high resolution, with every tile convincing in its own right, is computationally expensive, since the canvas must hold many detailed tiles at once. We present PhotoQuilt, a training-free framework that generates photomosaics at arbitrary resolution. Diffusion models struggle to satisfy both scales at once, as direct high-resolution generation is costly and tends toward one smooth image rather than a mosaic, while patch-based tiling keeps local detail but loses global structure. PhotoQuilt resolves this with a bootstrapped tiled denoising procedure. We first produce a global composition at low resolution to fix the layout, then upscale it in latent space and re-inject noise to restore generative capacity. Denoising proceeds within fixed tiles, so each forms its own image while the shared global structure holds them in one layout. Because tile generation is handled separately, PhotoQuilt scales to large canvases without quadratic attention cost. Experiments show that PhotoQuilt outperforms current baselines on both global structure and local realism.

Sources

Related papers