arxiv: 2604.03249 · v1 · submitted 2026-03-10 · 💻 cs.CY · cs.AI· cs.CV· cs.HC

Recognition: no theorem link

BLK-Assist: A Methodological Framework for Artist-Led Co-Creation with Generative AI Models

Daniel Grimes , Rachel M. Harrison

Authors on Pith no claims yet

Pith reviewed 2026-05-15 13:25 UTC · model grok-4.3

classification 💻 cs.CY cs.AIcs.CVcs.HC

keywords artist-led co-creationdiffusion modelsparameter-efficient fine-tuningprivacy-preserving AIstylistic fidelitygenerative frameworksconsent-based adaptation

0 comments

The pith

BLK-Assist shows how artists can fine-tune diffusion models on their own private corpus while preserving stylistic fidelity.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents BLK-Assist as a modular framework that lets a single professional artist adapt generative diffusion models to their specific style using efficient fine-tuning methods. It breaks the system into three parts for creating conceptual sketches, generating transparent assets, and producing high-resolution outputs, all while keeping the original artwork private. The authors document the full dataset setup, training steps, and inference process so others can follow the same approach with publicly available models. If the method holds, artists gain a concrete way to collaborate with AI without losing control over their visual identity or data. The work focuses on consent and reproducibility as core design choices for this kind of co-creation.

Core claim

BLK-Assist is a modular framework for artist-specific fine-tuning of diffusion models that uses parameter-efficient adaptation. It consists of BLK-Conceptor for LoRA-adapted conceptual sketch generation, BLK-Stencil for LayerDiffuse-based transparency-preserving asset generation, and BLK-Upscale for hybrid high-resolution output. The framework is demonstrated through a complete case study with one artist's proprietary corpus, including documented dataset composition, preprocessing, training configurations, and inference workflows. This setup illustrates a privacy-preserving, consent-based approach to human-AI co-creation that maintains stylistic fidelity to the source corpus and can be used,

What carries the argument

BLK-Assist, a three-component modular system for parameter-efficient fine-tuning that handles sketch generation, asset creation, and upscaling while keeping the artist's data private.

If this is right

Artists gain a documented path to generate new work in their own style without exposing their full body of work to external systems.
The same modular structure can be replicated by other artists using publicly available diffusion models under comparable consent conditions.
Reproducibility is supported by explicit records of dataset handling, training settings, and inference steps.
Stylistic fidelity is achieved through targeted adaptation rather than broad model retraining.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The modular split into sketch, stencil, and upscale stages could be applied to other generative tasks like animation or 3D asset creation by artists.
Similar consent-based fine-tuning setups might reduce disputes over style appropriation in commercial AI tools.
Extending the framework to multi-artist shared corpora under controlled access rules would test its scalability beyond single-user cases.

Load-bearing premise

The fine-tuning and inference workflows will reliably preserve stylistic fidelity when applied to new artists or different diffusion models, even though only a single-artist case study is shown.

What would settle it

Running the full BLK-Assist pipeline on a second artist with a markedly different visual style and measuring whether the generated outputs match the new artist's corpus at the same fidelity level as the original case.

Figures

Figures reproduced from arXiv: 2604.03249 by Daniel Grimes, Rachel M. Harrison.

**Figure 1.** Figure 1: The Throat Shakra (2025), produced with the BLK-Assist pipeline. The rapid advancement of large-scale text-to-image (TTI) models has ignited significant debate across creative industries. On one hand, critics argue that these systems are built on non-consensual data scraping that raises profound concerns around copyright, attribution, and the displacement of artistic labor [12, 29, 22]. On the other, adv… view at source ↗

**Figure 2.** Figure 2: Overview of the BLK-Conceptor inference pipeline. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Figure 3: Overview of the BLK-Stencil inference pipeline. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** Figure 4: Overview of the BLK-Upscale Diffusion+Texture LoRA inference pipeline. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

**Figure 5.** Figure 5: A before and after example of lowres artwork upscaled with BLK-Upscale. BLK-Upscale (see [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

**Figure 6.** Figure 6: A before and after example of a low-res image upscaled with BLK-Upscale Real-ESRGAN [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗

**Figure 7.** Figure 7: A before and after example of a low-res image upscaled with BLK-Upscale Diffu [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗

**Figure 8.** Figure 8: Input training data (images and captions) for BLK-Conceptor [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗

**Figure 9.** Figure 9: Examples of prompts and their corresponding outputs provided by BLK-Conceptor [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗

**Figure 10.** Figure 10: Input training data (PNGs and captions) for BLK-Stencil [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗

**Figure 11.** Figure 11: Examples of prompts and their corresponding outputs provided by BLK-Stencil [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗

read the original abstract

This paper presents BLK-Assist, a modular framework for artist-specific fine-tuning of diffusion models using parameter-efficient methods. The system is implemented as a case study with a single professional artist's proprietary corpus and consists of three components: BLK-Conceptor (LoRA-adapted conceptual sketch generation), BLK-Stencil (LayerDiffuse-based transparency-preserving asset generation), and BLK-Upscale (hybrid Real-ESRGAN and texture-conditioned diffusion for high-resolution outputs). We document dataset composition, preprocessing, training configurations, and inference workflows to enable reproducibility with publicly available models to illustrate a privacy-preserving, consent-based approach to human-AI co-creation that maintains stylistic fidelity to the source corpus and can be adapted for other artists under similar constraints.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

BLK-Assist gives a documented, modular recipe for consent-based artist fine-tuning of diffusion models but rests on one unquantified case study.

read the letter

The main thing to know is that this paper lays out a practical three-part workflow for adapting public diffusion models to match one professional artist's private style without releasing the data. BLK-Conceptor handles LoRA-based sketch generation, BLK-Stencil uses LayerDiffuse for transparent assets, and BLK-Upscale combines Real-ESRGAN with texture-conditioned diffusion. The write-up includes dataset makeup, preprocessing steps, training hyperparameters, and inference details, which makes the whole thing replicable for someone who wants to try the same setup on their own artist corpus. That level of documentation is the paper's real strength; it turns an otherwise common set of techniques into a named, step-by-step process focused on consent and privacy. In a field where most work either uses public scraped data or gives no implementation specifics, this is a useful reference point for anyone building artist-controlled tools. The limitation is straightforward and not hidden: everything comes from a single proprietary corpus. No FID scores, style-loss numbers, perceptual studies, or even side-by-side visual comparisons against a held-out set appear. There is also no second-artist replication or cross-backbone test, so the claim that the same configurations will work for other artists under similar constraints stays untested. The paper is aimed at practitioners and applied researchers who need concrete recipes for ethical fine-tuning rather than new theory. A reader who wants implementation guidance on privacy-preserving artist adaptation will find the details helpful. It deserves peer review because the documentation is clear enough to build on and the topic is relevant, though any referee would likely ask for quantitative validation and at least one additional case before publication.

Referee Report

2 major / 2 minor

Summary. This paper presents BLK-Assist, a modular framework for artist-specific fine-tuning of diffusion models via parameter-efficient methods. It describes three components—BLK-Conceptor (LoRA-adapted conceptual sketch generation), BLK-Stencil (LayerDiffuse-based transparency-preserving asset generation), and BLK-Upscale (hybrid Real-ESRGAN and texture-conditioned diffusion)—implemented as a single-artist case study with a proprietary corpus. The work documents dataset composition, preprocessing, training configurations, and inference workflows, claiming to illustrate a privacy-preserving, consent-based approach to human-AI co-creation that maintains stylistic fidelity to the source corpus and can be adapted for other artists under similar constraints.

Significance. If the documented workflows reliably preserve stylistic fidelity, the paper would supply a concrete, reproducible template for ethical artist-AI collaboration that prioritizes consent and privacy while leveraging publicly available models. This could serve as a reference for similar frameworks in creative AI applications, particularly by detailing end-to-end configurations that others might replicate.

major comments (2)

[Case Study] Case Study section: The manuscript reports only a single-artist proprietary corpus with no quantitative metrics (FID, style loss, perceptual scores, or held-out validation) or baseline comparisons to demonstrate that stylistic fidelity is maintained; claims of fidelity therefore rest solely on qualitative documentation rather than measured evidence.
[Abstract] Abstract and concluding discussion: The assertion that the approach 'can be adapted for other artists under similar constraints' is unsupported, as no second-artist replication, cross-style ablation, or transfer experiments across diffusion backbones are presented; the single-instance feasibility demonstration does not establish generalizability.

minor comments (2)

[Methodology] Methodology descriptions of LoRA rank, LayerDiffuse integration, and Real-ESRGAN conditioning could include explicit hyperparameter tables (e.g., learning rate schedules, epoch counts) to strengthen reproducibility claims.
The paper would benefit from a dedicated limitations subsection addressing potential failure modes when the same configurations are applied to artists with different stylistic variance or to newer diffusion architectures.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. Our manuscript presents BLK-Assist as a methodological framework and single-artist case study focused on documenting a privacy-preserving workflow using public models. We address the major comments below and will revise the manuscript accordingly.

read point-by-point responses

Referee: [Case Study] Case Study section: The manuscript reports only a single-artist proprietary corpus with no quantitative metrics (FID, style loss, perceptual scores, or held-out validation) or baseline comparisons to demonstrate that stylistic fidelity is maintained; claims of fidelity therefore rest solely on qualitative documentation rather than measured evidence.

Authors: We agree that the evaluation is qualitative only. The proprietary corpus precludes sharing data required for quantitative metrics such as FID, style loss, or perceptual scores, and no baseline comparisons were performed. The paper's contribution is the end-to-end documented workflow for consent-based fine-tuning. We will revise the case study section to state that stylistic fidelity is illustrated through qualitative examples and artist feedback rather than measured evidence. revision: yes
Referee: [Abstract] Abstract and concluding discussion: The assertion that the approach 'can be adapted for other artists under similar constraints' is unsupported, as no second-artist replication, cross-style ablation, or transfer experiments across diffusion backbones are presented; the single-instance feasibility demonstration does not establish generalizability.

Authors: We acknowledge that no replication or ablation studies are included, so generalizability is not empirically demonstrated. The phrasing was intended to highlight the modular, public-model design as a potential template. We will revise the abstract and conclusion to describe the framework as a documented case study that others may adapt under similar constraints, without claiming established generalizability. revision: yes

Circularity Check

0 steps flagged

No significant circularity in framework description

full rationale

The paper is a methodological framework description built around a single-artist case study. It documents dataset composition, preprocessing steps, LoRA training configurations, LayerDiffuse asset generation, and Real-ESRGAN upscaling workflows without presenting any mathematical derivations, equations, fitted parameters, or predictions. No self-citations are invoked as load-bearing uniqueness theorems, and no ansatz or renaming of known results occurs. The central claim that the workflows maintain stylistic fidelity for the given corpus and can be adapted under similar constraints is supported by the explicit documentation of the implementation rather than by any reduction to quantities defined by the paper's own inputs. This is a standard descriptive case-study paper whose claims do not exhibit circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The framework rests on standard assumptions about diffusion model adaptability and parameter-efficient fine-tuning; no new entities are postulated and no free parameters are numerically specified in the abstract.

axioms (2)

domain assumption Parameter-efficient methods such as LoRA can adapt diffusion models to preserve an individual artist's stylistic fidelity from a proprietary corpus.
Invoked in the description of BLK-Conceptor and overall system design.
domain assumption LayerDiffuse and hybrid Real-ESRGAN pipelines can produce transparency-preserving and high-resolution outputs without degrading style consistency.
Assumed in the design of BLK-Stencil and BLK-Upscale components.

pith-pipeline@v0.9.0 · 5431 in / 1366 out tokens · 45673 ms · 2026-05-15T13:25:25.036953+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages · 3 internal anchors

[1]

Towards personalizing generative ai with small data for co-creation in the visual arts

Ahmed M Abuzuraiq and Philippe Pasquier. Towards personalizing generative ai with small data for co-creation in the visual arts. InIUI workshops, pages 1–14, 2024

work page 2024
[2]

Stable Diffusion Web UI, August 2022

AUTOMATIC1111. Stable Diffusion Web UI, August 2022

work page 2022
[3]

SimpleTuner.https://github.com/bghira/SimpleTuner, 2025

bghira. SimpleTuner.https://github.com/bghira/SimpleTuner, 2025

work page 2025
[4]

HiDream-I1: A high-efficient image generative foundation model with sparse diffusion transformer.arXiv preprint arXiv:2505.22705, 2025

Qi Cai, Jingwen Chen, Yang Chen, Yehao Li, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Yiheng Zhang, Fengbin Gao, Peihan Xu, et al. HiDream-I1: A high-efficient image generative foundation model with sparse diffusion transformer.arXiv preprint arXiv:2505.22705, 2025

work page arXiv 2025
[5]

ComfyUI: The most powerful and modular visual AI engine and application

comfyanonymous. ComfyUI: The most powerful and modular visual AI engine and application. https://github.com/comfyanonymous/ComfyUI, 2025

work page 2025
[6]

conda: A system-level, binary package and environment manager running on all major operating systems and platforms

conda contributors. conda: A system-level, binary package and environment manager running on all major operating systems and platforms

work page
[7]

Human-computer interaction

Alan Dix. Human-computer interaction. InEncyclopedia of database systems, pages 1327–1331. Springer, 2009

work page 2009
[8]

Optimum quanto

Hugging Face. Optimum quanto. https://github.com/huggingface/optimum-quanto, 2025

work page 2025
[9]

Parameter-Efficient Fine-Tuning for Large Models: A Comprehensive Survey

Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey.arXiv preprint arXiv:2403.14608, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[10]

Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

work page 2020
[11]

Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

work page 2022
[12]

Ai art and its impact on artists

Harry H Jiang, Lauren Brown, Jessica Cheng, Mehtab Khan, Abhishek Gupta, Deja Workman, Alex Hanna, Johnathan Flowers, and Timnit Gebru. Ai art and its impact on artists. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 363–374, 2023

work page 2023
[13]

Large- scale text-to-image generation models for visual artists’ creative works

Hyung-Kwon Ko, Gwanmo Park, Hyeon Jeon, Jaemin Jo, Juho Kim, and Jinwook Seo. Large- scale text-to-image generation models for visual artists’ creative works. InProceedings of the 28th international conference on intelligent user interfaces, pages 919–933, 2023

work page 2023
[14]

Gonzalez, Hao Zhang, and Ion Stoica

Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large lan- guage model serving with pagedattention. InProceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023

work page 2023
[15]

FLUX.https://github.com/black-forest-labs/flux, 2024

Black Forest Labs. FLUX.https://github.com/black-forest-labs/flux, 2024

work page 2024
[16]

FLUX.1 Kontext: Flow matching for in-context image generation and editing in latent space, 2025

Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, Sumith Kulal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas Müller, Dustin Podell, Robin Rombach, Harry Saini, Axel Sauer, and Luke Smith. FLUX.1 Kontext: Flow matching for in-context image ...

work page 2025
[17]

The societal evolution of the roles and functions of artists’ studios

Chao Li. The societal evolution of the roles and functions of artists’ studios

work page
[18]

sd-forge-layerdiffuse: Transparent image layer diffusion using latent transparency

lllyasviel. sd-forge-layerdiffuse: Transparent image layer diffusion using latent transparency. https://github.com/lllyasviel/sd-forge-layerdiffuse, 2024

work page 2024
[19]

Decoupled Weight Decay Regularization

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017. 7

work page internal anchor Pith review Pith/arXiv arXiv 2017
[20]

Human-computer interaction: An empirical research perspective

I Scott MacKenzie. Human-computer interaction: An empirical research perspective. 2024

work page 2024
[21]

The creativity of text-to-image generation

Jonas Oppenlaender. The creativity of text-to-image generation. InProceedings of the 25th international academic mindtrek conference, pages 192–202, 2022

work page 2022
[22]

Text-to-image generation: Perceptions and realities.arXiv preprint arXiv:2303.13530, 2023

Jonas Oppenlaender, Aku Visuri, Ville Paananen, Rhema Linder, and Johanna Silvennoinen. Text-to-image generation: Perceptions and realities.arXiv preprint arXiv:2303.13530, 2023

work page arXiv 2023
[23]

AI-Toolkit.https://github.com/ostris/ai-toolkit, 2025

Ostris. AI-Toolkit.https://github.com/ostris/ai-toolkit, 2025

work page 2025
[24]

Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

work page 2022
[25]

Human-computer interaction

Jenny Preece, Yvonne Rogers, Helen Sharp, David Benyon, Simon Holland, and Tom Carey. Human-computer interaction. Addison-Wesley Longman Ltd., 1994

work page 1994
[26]

pyenv: Simple Python Version Management

pyenv. pyenv: Simple Python Version Management. https://github.com/pyenv/pyenv, 2025

work page 2025
[27]

Flux-version-LayerDiffuse

RedAIGC. Flux-version-LayerDiffuse. https://github.com/RedAIGC/ Flux-version-LayerDiffuse, 2025

work page 2025
[28]

ComfyUI_UltimateSDUpscale

ssitu. ComfyUI_UltimateSDUpscale. https://github.com/ssitu/ComfyUI_ UltimateSDUpscale, 2025

work page 2025
[29]

Creative ownership and control for generative ai in art and design

Alexa Steinbrück and Aeneas Stankowski. Creative ownership and control for generative ai in art and design. InIn Generative AI in HCI Workshop, CHI, volume 23, 2023

work page 2023
[30]

Copyright Office

U.S. Copyright Office. What does copyright protect?

work page
[31]

Copyright Office

U.S. Copyright Office. Report on copyright and artificial intelligence: Part 2: Copyrightability. Technical report, U.S. Copyright Office, Washington, DC, January 2025. Pre-publication release

work page 2025
[32]

Copyright Office, Library of Congress

U.S. Copyright Office, Library of Congress. Copyright registration guidance: Works containing material generated by artificial intelligence. Federal Register, March 2023. Statement of policy, effective March 16, 2023

work page 2023
[33]

Supreme Court

U.S. Supreme Court. Burrow-giles lithographic co. v. sarony, 111 u.s. 53, 57–58 (1884). U.S. Reports, 1884. United States Supreme Court decision

work page
[34]

A small-data mindset for generative ai creative work

Gabriel Vigliensoni, Phoenix Perry, Rebecca Fiebrink, et al. A small-data mindset for generative ai creative work. 2022

work page 2022
[35]

Diffusers: State-of-the-art diffusion models

Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, Steven Liu, William Berman, Yiyi Xu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models

work page
[36]

Wan: Open and Advanced Large-Scale Video Generative Models

Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pandeng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fang, T...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[37]

Real-esrgan: Training real-world blind super-resolution with pure synthetic data

Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. InProceedings of the IEEE/CVF international conference on computer vision, pages 1905–1914, 2021. 8

work page 1905
[38]

Qwen-image technical report, 2025

Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, Yuxiang Chen, Zecheng Tang, Zekai Zhang, Zhengyi Wang, An Yang, Bowen Yu, Chen Cheng, Dayiheng Liu, Deqing Li, Hang Zhang, Hao Meng, Hu Wei, Jingyuan Ni, Kai Chen, Kuan Cao, Liang Peng, Lin Qu, Minggang Wu, Peng Wang, Shuting Yu, Tingkun...

work page 2025
[39]

Navigating text-to-image customization: From lyCORIS fine-tuning to model evaluation

SHIH-YING YEH, Yu-Guan Hsieh, Zhidong Gao, Bernard B W Yang, Giyeong Oh, and Yanmin Gong. Navigating text-to-image customization: From lyCORIS fine-tuning to model evaluation. InThe Twelfth International Conference on Learning Representations, 2024

work page 2024
[40]

Transparent image layer diffusion using latent trans- parency.arXiv preprint arXiv:2402.17113, 2024

Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent trans- parency.arXiv preprint arXiv:2402.17113, 2024. 9 A Technical Appendices and Supplementary Material Figure 6: A before and after example of a low-res image upscaled with BLK-Upscale Real-ESRGAN . Figure 7: A before and after example of a low-res image upscaled with BLK-...

work page arXiv 2024